Keywords: SWE-Explore, coding agents, repository exploration, line budget, agentic exploration

Category: AI Systems and Tools

Research Objective:

– SWE-Explore introduces a benchmark to assess the repository exploration capabilities of coding agents, focusing on ranked lists of code within line budgets to surpass traditional retrieval methods.

Research Methods:

– The study evaluates 848 issues across 10 programming languages and 203 repositories, emphasizing metrics like coverage, ranking, and context-efficiency, derived from agent trajectories.

Research Conclusions:

– Agentic explorers demonstrate superior performance compared to classical retrieval methods, notably in line-level coverage and efficient ranking, which differentiate state-of-the-art explorers.

Paper link: https://huggingface.co/papers/2606.07297

2. Latent Spatial Memory for Video World Models

Keywords: Video world models, Latent spatial memory, Diffusion latent space, End-to-end video generation

Category: Computer Vision

Research Objective:

– Introduce latent spatial memory for video world models to eliminate pixel-space reconstruction overhead.

Research Methods:

– Implement a framework called Mirage that constructs 3D memory directly in diffusion latent space through depth-guided back-projection and novel view synthesis.

Research Conclusions:

– The proposed method achieves up to 10.57 times faster video generation and reduces memory footprint by 55 times compared to traditional methods, achieving state-of-the-art performance on benchmarks like WorldScore and RealEstate10K.

Paper link: https://huggingface.co/papers/2606.09828

3. Agents’ Last Exam

Keywords: AI agents, real-world tasks, industry clusters, task taxonomy, living benchmark

Category: AI Systems and Tools

Research Objective:

– The paper introduces Agents’ Last Exam (ALE), a benchmark tailored to evaluate AI agents on long-term, economically valuable real-world tasks that cover 13 industry clusters and over 1,000 tasks.

Research Methods:

– ALE was developed in collaboration with over 250 industry experts, structured around a task taxonomy with 55 subfields, and designed to continuously grow with new workflows and industries.

Research Conclusions:

– Current evaluation results show a significant gap between AI benchmark performance and real-world deployment, with an average full pass rate of only 2.6%. ALE is intended to bridge this gap by offering a more practical measure of AI impact on GDP-relevant tasks.

Paper link: https://huggingface.co/papers/2606.05405

4. FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention

Keywords: Lookahead Sparse Attention, Neural Memory Indexer, KV Cache, FlashMemory, Dual-Encoder Architecture

Category: Natural Language Processing

Research Objective:

– The study aims to address the GPU memory bottleneck caused by conventional LLMs during decoding by introducing Lookahead Sparse Attention (LSA) empowered by a Neural Memory Indexer.

Research Methods:

– The LSA technique involves proactive management of the KV cache by predicting future context demands.

– Utilizes a decoupled training strategy with a standard dual-encoder architecture, trained independently using standard retrieval training frameworks.

Research Conclusions:

– The proposed LSA approach significantly reduces GPU memory usage for long-context tasks, compressing the average physical KV cache footprint to 13.5% compared to the full-context baseline.

– Maintains or slightly improves downstream accuracy, achieving a +0.6% margin on average and reducing KV cache overhead by over 90% at extreme scales without impacting reasoning capabilities.

Paper link: https://huggingface.co/papers/2606.09079

5. Human Psychometric Questionnaires Mischaracterize LLM Behavior

Keywords: LLM behavior, real-world interactions, generation-based profiling, psychometric questionnaires, generation probabilities

Category: Natural Language Processing

Research Objective:

– To determine the reliability of human psychometric questionnaires in predicting LLM behavior during everyday user interactions.

Research Methods:

– Comparison of LLM value and personality profiles derived from Likert self-reports and generation probabilities over value-laden responses to user queries.

Research Conclusions:

– Psychometric questionnaires are insufficient for predicting LLM behavior as they fail to replicate realistic user query responses, while generation-based profiling provides a more accurate understanding.

Paper link: https://huggingface.co/papers/2509.10078

6. End-to-End Context Compression at Scale

Keywords: Latent Context Language Models, Compression Ratios, Encoder-decoder Compression, Architecture Search, Pre-training

Category: Natural Language Processing

Research Objective:

– The study aims to enhance encoder-decoder compression techniques through architectural search and extensive pre-training to develop Latent Context Language Models (LCLMs) that efficiently manage long contexts with improved performance and memory usage compared to traditional KV cache methods.

Research Methods:

– The authors conducted an architecture search and pre-trained various encoder-decoder model variants from scratch to determine optimal design and training strategies. They introduced LCLMs with pre-trained encoder-decoder models at different compression ratios of 1:4, 1:8, and 1:16 on over 350B tokens.

Research Conclusions:

– The LCLMs effectively enhance the Pareto frontier of general-task performance, compression speed, and memory usage. They serve as efficient backbones for long-horizon agents, facilitating skim through compressed contexts and adaptive expansion of relevant segments when needed.

Paper link: https://huggingface.co/papers/2606.09659

7. A Geometric Account of Activation Steering through Angle-Norm Decomposition

Keywords: hidden-state norm, angular structure, spherical steering, activation steering, language models

Category: Natural Language Processing

Research Objective:

– The study aims to challenge the notion that hidden-state norms carry concept-relevant information in language models and to understand how concepts are represented through angular structures and norms.

Research Methods:

– Conducted a controlled empirical study to explore the roles of angular and radial components by comparing different steering methods in language models.

Research Conclusions:

– Findings indicate concepts are mainly represented through angular structures, yet the hidden-state norm is crucial for stability and effectiveness in steering methods. The study suggests that activation steering should be parameterized by angular and radial components for better interpretability and performance.

Paper link: https://huggingface.co/papers/2606.06735

8. SwiftVR: Real-Time One-Step Generative Video Restoration

Keywords: real-time video restoration, consumer GPUs, efficient attention mechanisms, lightweight autoencoding, causal chunk-wise protocol

Category: Generative Models

Research Objective:

– To enable real-time video restoration on consumer GPUs achieving high frame rates at 4K resolution through efficient attention mechanisms and lightweight autoencoding.

Research Methods:

– Utilization of mask-free shifted-window self-attention for efficient spatial window processing and lightweight restoration-aware autoencoding for fast, quality-preserving chunk-wise decoding.

Research Conclusions:

– SwiftVR sustains significant frame rates on high-resolution settings and is the first generative VR model enabling real-time 1080p streaming on consumer-grade GPUs, ensuring strong no-reference perceptual quality with low inference cost.

Paper link: https://huggingface.co/papers/2606.09516

9. Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders

Keywords: Hallucinations, Whisper ASR, Sparse AutoEncoder, Activation-space steering, SAE latent-space steering

Category: Natural Language Processing

Research Objective:

– The study aims to detect and reduce hallucinations in Whisper ASR using internal representations from audio encoder activations and Sparse AutoEncoder latents.

Research Methods:

– The research involves extracting audio encoder activations and evaluating two representation spaces: raw Whisper activations and Sparse AutoEncoder latents. Two strategies are proposed for steering: activation-space steering and SAE latent-space steering.

Research Conclusions:

– The results show a remarkable reduction in hallucination rates, from 72.63% to 14.11% for Whisper small and from 86.88% to 27.33% for Whisper large-v3, with minimal transcription degradation, demonstrating the effectiveness of the proposed strategies.

Paper link: https://huggingface.co/papers/2606.07473

10. DEI: Diversity in Evolutionary Inference for Quality-Diversity Search

Keywords: Heterogeneous Large Language Models, Evolutionary Inference, Quality-Diversity, Mutation Operators, Cross-Model Adversarial Pressure

Category: Generative Models

Research Objective:

– To showcase how using heterogeneous large language models as mutation operators in a distributed Quality-Diversity search framework can enhance evolutionary inference.

Research Methods:

– Implemented DEI, a distributed Quality-Diversity framework that utilizes heterogeneous LLMs for mutation operations, extending the Digital Red Queen framework for cross-model competition and robustness.

Research Conclusions:

– DEI’s use of model diversity significantly improves performance over homogeneous setups, demonstrated by higher QD-Scores and coverage in the Core War domain. This highlights the importance of model diversity as opposed to mere parallelism in distributed LLM-based QD search.

Paper link: https://huggingface.co/papers/2605.27130

11. SlimSearcher: Training Efficiency-Aware Web Agents via Adaptive Reward Gating

Keywords: SlimSearcher, Pareto-efficient trajectory filtering, Adaptive Reward Shaping, Reinforcement Learning, computational efficiency

Category: Reinforcement Learning

Research Objective:

– The paper introduces SlimSearcher, a framework designed to improve the efficiency of deep research agents by balancing the trade-off between computational costs and accuracy.

Research Methods:

– The framework employs Pareto-efficient trajectory filtering during Supervised Fine-Tuning and Adaptive Reward Shaping during Reinforcement Learning to enhance efficiency and accuracy.

Research Conclusions:

– Experiments on benchmarks such as GAIA, BrowseComp, and XBenchDeepSearch show that SlimSearcher can reduce tool-call rounds by 17%-58% while maintaining or enhancing accuracy.

Paper link: https://huggingface.co/papers/2606.07074

12. Liberating LLM Capabilities in Full-Duplex Speech Models

Keywords: tri-channel speech interface, full-duplex interaction, Listen-Write-Speak, text-first paradigm, autoregressive LLM

Category: Human-AI Interaction

Research Objective:

– The paper proposes a text-first tri-channel speech interface that emphasizes the importance of visible text output alongside spoken responses for real-time and structured conversational tasks.

Research Methods:

– Introduces the Listen-Write-Speak (LWS) paradigm using an autoregressive LLM to handle audio, text, and speech concurrently with a shared causal attention context, leveraging a Token Schema without architectural changes.

Research Conclusions:

– The study demonstrates that visible writing can effectively serve as a first-class output channel for speech interaction, maintaining high responsiveness and performance across multiple benchmarks.

Paper link: https://huggingface.co/papers/2606.07547

13. Light-WAM: Efficient World Action Models with State-Fusion Action Decoding

Keywords: Light-WAM, World Action Models, robot manipulation, video backbone, StateFusionActionExpert

Category: Robotics and Autonomous Systems

Research Objective:

– The main goal is to develop a lightweight World Action Model called Light-WAM for efficient robot manipulation that incorporates future-video supervision to enhance temporal structure representation.

Research Methods:

– Use a compact video backbone and downsampled latent space to reduce video co-training costs.

– Implement the StateFusionActionExpert to directly predict action chunks with learned-query pooling from multiple backbone layers.

Research Conclusions:

– Light-WAM demonstrates strong performance on LIBERO and achieves functional multi-task performance on RoboTwin 2.0 with only 0.44B trainable parameters.

– Achieves low inference latency of 72.03ms with 4.1GiB peak GPU memory usage and improved training throughput.

Paper link: https://huggingface.co/papers/2606.08242

14. SDR: Set-Distance Rewards for Radiology Report Generation

Keywords: set-based rewards, embedding distances, chest X-ray report generation, vision–language models, reinforcement learning

Category: AI in Healthcare

Research Objective:

– To improve chest X-ray report generation by employing set-based rewards using embedding distances that facilitate effective post-training and test-time selection without the need for causal reasoning structures.

Research Methods:

– Used a set-based approach where reports are split into sentences and transformed into unordered embedding sets.

– Proposed set-to-set distances between generated and reference embeddings as continuous, permutation-invariant rewards.

– Conducted experiments across two datasets and three vision-language models, comparing post-training set-to-set distance based rewards against supervised fine-tuning.

Research Conclusions:

– Set-to-set distance based rewards consistently outperform supervised fine-tuning on all headline metrics with notable improvements in BERTScore, RadGraph F1, and CheXbert F1 scores.

– The approach facilitates test-time best-of-N selection, providing a significant performance improvement over random selection.

– Set-distance rewards enable more efficient test-time scaling, reducing generated tokens while maintaining quality, thus establishing them as a unified signal for both post-training and test-time scaling in chest X-ray report generation.

Paper link: https://huggingface.co/papers/2606.00440

15. Trajectory-Refined Distillation

Keywords: On-policy distillation, Prefix failure, Trajectory-refined distillation, Large language models, Teacher guidance

Category: Natural Language Processing

Research Objective:

– The research aims to address prefix failure in on-policy distillation (OPD) for large language models by proposing trajectory-level corrections.

Research Methods:

– The paper introduces Trajectory-Refined Distillation (TRD), a method that corrects student’s rollouts at the trajectory level under teacher guidance to mitigate prefix failure.

Research Conclusions:

– TRD successfully improves exploration by exposing students to alternative valid derivations, enhances single-attempt accuracy, and broadens reasoning coverage, outperforming prior baselines across various benchmarks and scales.

Paper link: https://huggingface.co/papers/2606.08432

16. DuMate-DeepResearch: An Auditable Multi-Agent System with Recursive Search and Rubric-Grounded Reasoning

Keywords: Deep Research, Multi-agent Framework, Long-form Synthesis, Planning, Tool Ecosystem

Category: AI Systems and Tools

Research Objective:

– The aim is to create a multi-agent framework called DuMate-DeepResearch to address the challenges of deep research tasks, improving planning, evidence acquisition, and report synthesis.

Research Methods:

– The framework decouples task understanding, planning, and scheduling from evidence acquisition and report rendering, making decisions traceable using a dynamic optimization approach.

– Introduces dynamic graph-based planning, recursive two-level execution, and rubric-based test-time optimization mechanisms.

Research Conclusions:

– DuMate-DeepResearch achieved state-of-the-art results on two benchmarks, marking top scores in both overall performance and specific metrics such as information recall and analysis.

Paper link: https://huggingface.co/papers/2606.07299

17. Evaluation Cards: An Interpretive Layer for AI Evaluation Reporting

Keywords: EvalCards, AI evaluation, interpretive signals, benchmark metadata, score comparability

Category: AI Systems and Tools

Research Objective:

– To develop EvalCards, a framework that standardizes and unifies AI evaluation reporting across various platforms to overcome inconsistencies and facilitate reliable comparisons.

Research Methods:

– A structured review of 52 papers and 10 stakeholder interviews to derive a reporting schema.

– Implementation of four key interpretive signals: reproducibility, documentation completeness, provenance and risk, and score comparability.

Research Conclusions:

– A new operational reporting layer was created, deploying a monitoring tool that applied to thousands of models, benchmarks, and results, exposing systematic gaps in current AI evaluation reporting practices.

Paper link: https://huggingface.co/papers/2606.09809

18. WorldCraft: From Camera Navigation to Object Manipulation in Interactive Video World Models

Keywords: WorldCraft, video-based world models, object-level trajectory actions, camera navigation, trajectory-centric control

Category: Computer Vision

Research Objective:

– Introduce WorldCraft, a framework that extends video-based world models to include object-level trajectory control while maintaining camera navigation functionalities.

Research Methods:

– Utilizes a trajectory-centric control pipeline with components like Normalized World Trajectory, Spatial-Pathway LoRA, and Trajectory-Anchored State Persistence to achieve simultaneous object manipulation and camera navigation.

Research Conclusions:

– WorldCraft successfully enables precise object control, preserves camera navigation fidelity, and maintains object state across extended scenarios, even during off-camera excursions.

Paper link: https://huggingface.co/papers/2605.25077

19. AsyncWebRL: Efficient Multi-Step RL for Visual Web Agents

Keywords: AsyncWebRL, asynchronous reinforcement learning, trajectory normalization

Category: Reinforcement Learning

Research Objective:

– This research aims to enhance vision-language web agent training by employing asynchronous reinforcement learning and modifying trajectory normalization to achieve faster throughput and better performance on challenging tasks.

Research Methods:

– The study introduces AsyncWebRL, which combines an asynchronous design with specific adaptations like an everlasting rollout pool and lightweight screenshot handling, resulting in a significant speedup in training throughput.

– Implementing a modification in trajectory normalization by replacing 1/|τ_i| with a constant 1/k, improving trajectory shortening while maintaining success rates.

Research Conclusions:

– The AsyncWebRL approach sets a new open-source state-of-the-art performance on the WebGym out-of-distribution test split, with notable performance improvements on more difficult tasks, achieving up to +48% relative gain on the hardest slice.

Paper link: https://huggingface.co/papers/2606.05597

20. Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loops

Keywords: LLM agents, exploit-resistant verifiers, reward hacking, hacker-fixer loop, Terminal Wrench

Category: Reinforcement Learning

Research Objective:

– The study aims to identify vulnerabilities in agent benchmark verification systems and develop an automated iterative process using LLM agents to create robust verifiers that resist exploitation while maintaining legitimate task performance.

Research Methods:

– An audit was conducted on 1,968 tasks across five terminal-agent benchmarks to assess hackability by frontier models.

– The introduction of the hacker-fixer loop, which uses three LLM agents iteratively: a hacker, a fixer, and a solver to build exploit-resistant verifiers without per-task manual patching.

Research Conclusions:

– The hacker-fixer loop significantly reduced attack success rates; for example, it brought the attack success rate from 62% to 0% on KernelBench.

– Weaker agents in the loop were effective against more powerful attackers, underscoring the loop’s robustness in identifying and mitigating exploits.

– Terminal Wrench was released as a snapshot of the current attack surface and a basis for future research, including patched verifiers and discovered exploits.

Paper link: https://huggingface.co/papers/2606.08960

21. Where Rectified Flows Leak: Characterising Membership Signals Along the Interpolation Path

Keywords: Rectified Flows, Membership Inference Attack, training data traces, interpolation path, bell-shaped curve

Category: Generative Models

Research Objective:

– To understand what generative models retain from training data and its implications for privacy and copyright.

Research Methods:

– Analysis of the interpolation path in Rectified Flows to identify differences in the reconstruction of training and test data, and derivation of a maximum point under Gaussian assumptions.

Research Conclusions:

– The study suggests that Rectified Flows encode subtle traces of training data exploitable for membership inference attacks, with a universal bell-shaped structure identified in the data reconstruction curve.

Paper link: https://huggingface.co/papers/2606.07271

22. Chiaroscuro Attention: Spending Compute in the Dark

Keywords: CHIAR-Former, spectral entropy, self-attention, DCT, attention FLOPs

Category: Natural Language Processing

Research Objective:

– The research aims to enhance transformer efficiency on large text datasets by dynamically routing tokens using spectral entropy to select optimal operators.

Research Methods:

– A 4-layer hybrid transformer, CHIAR-Former, is proposed, which routes tokens among DCT spectral mixing, RBF kernel mixing, and full self-attention, with evaluations conducted on datasets like WikiText-103 and IMDB sentiment classification.

Research Conclusions:

– The CHIAR-Former demonstrated a 45% improvement in performance over traditional full-attention models on WikiText-103 with reduced computational resources, indicating the advantages of spectral routing in large-scale text processing.

Paper link: https://huggingface.co/papers/2606.08327

23. OASIS: From Simulation Data Collection to Real-World Humanoid Loco-Manipulation

Keywords: simulation-data-driven framework, humanoid loco-manipulation, 3D generative model, hierarchical visuomotor policy, domain randomization

Category: Robotics and Autonomous Systems

Research Objective:

– The paper aims to enhance humanoid robot loco-manipulation by utilizing a simulation-data-driven framework named OASIS, leveraging simulation to overcome limitations in traditional robot manipulation task demonstrations.

Research Methods:

– The authors employ a 3D generative model to create realistic assets and collect trajectories through teleoperation in simulation, which are further augmented with domain randomization. They design a hierarchical visuomotor policy based on this augmented simulation data.

Research Conclusions:

– The proposed framework, OASIS, demonstrates that policies trained on simulated data achieve better zero-shot performance compared to those trained on real-robot teleoperation data, ensuring higher success rates on various tasks by capturing broader environmental variations.

Paper link: https://huggingface.co/papers/2606.08548

24. Precision Is Not Faithfulness: Coverage-Aware Evaluation of Grounded Generation with a Complete Oracle

Keywords: Reference-free faithfulness, precision, recall, grounded generation

Category: Natural Language Processing

Research Objective:

– The study aims to address the limitation of reference-free faithfulness metrics that only measure precision, proposing a new metric that combines precision and recall to offer a more comprehensive evaluation of generated content’s faithfulness.

Research Methods:

– The researchers utilize Formula 1 telemetry as a deterministic domain to measure both precision and recall by having access to complete ground truth. They conduct experiments on a multilingual benchmark and a second complete-oracle domain, NOAA weather forecasts, to validate their metric.

Research Conclusions:

– The study finds that high-precision models often have poor fact coverage, thus ranking lower when evaluated with both precision and recall. A new verifier-guided generation method is proposed, improving precision and recall without needing references, demonstrating the effectiveness of their proposed metric.

Paper link: https://huggingface.co/papers/2606.09376

25. Lean4Agent: Formal Modeling and Verification for Agent Workflow and Trajectory

Keywords: Large Language Models, Lean4, Formal Verification, Multi-step Workflows, Agent Behavior

Category: AI Systems and Tools

Research Objective:

– The primary aim is to enhance the reliability and performance of multi-step workflows in Large Language Models (LLMs) using a formal verification framework with Lean4, a dependent-type formal language.

Research Methods:

– Development of Lean4 Agent, which utilizes Lean4 to model and verify agent behavior, with the FormalAgentLib library to ensure semantic consistency and debug workflow execution, and LeanEvolve to enhance workflows utilizing results from FormalAgentLib.

Research Conclusions:

– Achieved an 11.94% improvement on verification-passing workflows over failing ones and enhanced SWE performance by 7.47% using LeanEvolve. Stablished a foundational framework for formal modeling and verification of agent behavior with dependent-type formal languages.

Paper link: https://huggingface.co/papers/2606.06523

26. Cosine Misleads: Auxiliary Losses Reshape Vision Language Models, Not Their Latents

Keywords: Latent Visual Reasoning, Supervised Latent Tokens, Cosine Similarity, Information Bottleneck, Vision-Language Models

Category: Multi-Modal Learning

Research Objective:

– Challenge conventional views on the relationship between latent mechanisms and accuracy in vision-language models (VLMs), specifically focusing on the correlation between cosine alignment of supervised latents and model accuracy.

Research Methods:

– Experimentation with a designed matrix of five different Latent Visual Reasoning (LVR) variants to evaluate the correlation between cosine alignment and accuracy.

– Introduction of PRISM diagnostics: a linear probe to determine answer decodability and a corruption test to assess the dependency on latent states.

Research Conclusions:

– The study reveals an inverse correlation between the cosine alignment of supervised latents and model accuracy (r=-0.94).

– Answers in VLMs are decoded from downstream latents rather than directly within them, indicating limited dependency on these latent states.

– An Information Bottleneck approach demonstrates that auxiliary objectives reshape models through shared parameters, rather than exclusively through the targeted latent variables.

Paper link: https://huggingface.co/papers/2606.05753

27. Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?

Keywords: Large language models, machine translation, low-resource languages, linguistic reasoning traces, in-context learning

Category: Natural Language Processing

Research Objective:

– To explore the potential of large language models to improve machine translation for extremely low-resource languages by using structured linguistic reasoning traces.

Research Methods:

– Proposing a pipeline for generating linguistic reasoning traces from resources like Universal Dependencies treebanks, dictionaries, and grammar-rule banks.

– Evaluating linguistic reasoning traces in the contexts of in-context learning, supervised fine-tuning, and reinforcement fine-tuning, specifically for Xibe and Chintang languages.

Research Conclusions:

– Linguistic reasoning traces significantly enhance translation performance during inference as reliable sentence-specific traces improve performance across models and languages.

– Using traces as training data results in less consistent improvements, indicating that effective inference-time guidance can better leverage grammatical information for low-resource machine translation, while generating reliable analyses remains challenging.

Paper link: https://huggingface.co/papers/2606.03782

28. Set-Based Transformer for Atmospheric Compensation in Standoff LWIR Hyperspectral Imaging

Keywords: Lightweight deep learning, LWIR hyperspectral imaging, atmospheric compensation, transmittance estimation, sparse autoencoder

Category: Machine Learning

Research Objective:

– Develop a lightweight deep learning framework for atmospheric compensation in passive long-wave infrared hyperspectral imaging.

Research Methods:

– Utilize a set-based deep learning framework to jointly estimate transmittance, atmospheric path radiance, and downwelling spectrum from multi-range radiance measurements. Analyze learned representation using a sparse autoencoder.

Research Conclusions:

– The framework demonstrates low spectral distortion in atmospheric compensation tasks, with geographically coherent latent features emerging without location supervision. Publicly available dataset and code.

Paper link: https://huggingface.co/papers/2606.08324

29. Pruning and Distilling Mixture-of-Experts into Dense Language Models

Keywords: Mixture-of-Experts, Knowledge Distillation, Memory-Constrained Deployment, Dense Architectures, Scoring Method

Category: Natural Language Processing

Research Objective:

– To develop a systematic framework for converting Mixture-of-Experts (MoE) models into fully dense architectures, addressing memory constraints during deployment.

Research Methods:

– Experts within MoE are scored, selected, and grouped through a variety of methods, then concatenated into a dense feedforward network. Knowledge distillation is applied from the original MoE to refine the dense architecture.

Research Conclusions:

– The novel diversity-aware scoring method outperforms previous methods in various testing configurations, demonstrating a significant improvement in downstream accuracy (+6.3 pp) and training speed (1.6x faster) compared to traditional pruning techniques.

Paper link: https://huggingface.co/papers/2605.28207

30.

Paper link: 

31. EMMA: Extracting Multiple physical parameters from Multimodal Data

Keywords: EMMA, physics-informed, multimodal, Liquid Time-Constant, dynamical parameters

Category: Multi-Modal Learning

Research Objective:

– To introduce and utilize EMMA, a novel physics-informed multimodal framework, for recovering dynamical parameters directly from raw video, audio, and image data.

Research Methods:

– Utilization of a Liquid Time-Constant network and physics-constrained loss for learning latent dynamics and enforcing consistency with differential equations.

– A unified feature pipeline enables the alignment of data across various modalities without the need for additional segmentations or sensors.

Research Conclusions:

– EMMA achieves robust multi-parameter recovery and outperforms existing single-modality baselines. It is established as a scalable solution for extracting physics-consistent models from multimodal data.

Paper link: https://huggingface.co/papers/2605.24047

32. Honest Lying: Understanding Memory Confabulation in Reflexive Agents

Keywords: Reflexion-style agents, self-generated reflections, memory confabulation, Reflection Repetition Rate, trajectory-level failure signals

Category: Reinforcement Learning

Research Objective:

– The study aims to identify and address the issue of persistent errors in Reflexion-style agents due to incorrect self-generated reflections, specifically measured by a new metric called the Reflection Repetition Rate (RRR).

Research Methods:

– The research utilizes the Reflection Repetition Rate (RRR) metric to detect repeated reliance on incorrect reflections across environments such as ALFWorld and HumanEval and employs programmatic extraction of trajectory-level failure signals to mitigate these errors.

Research Conclusions:

– The study finds that memory confabulation leads to persistent errors and incorrect task interpretations. Mitigation strategies significantly increase correct object mention and reduce RRR, demonstrating that improvements in reflective memory processes can reduce errors and support more accurate task execution.

Paper link: https://huggingface.co/papers/2605.29463

33. CIPER: A Unified Framework for Cross-view Image-retrieval and Pose-estimation

Keywords: CIPER, Cross-view geo-localization, 3-DoF pose estimation, transformer encoder, multi-task objective

Category: Computer Vision

Research Objective:

– The study presents CIPER, aiming to solve cross-view geo-localization by improving simultaneous city-scale retrieval and precise 3-DoF pose estimation.

Research Methods:

– Utilizes a shared transformer encoder with task-specific tokens and a two-way transformer pose decoder for disentangling retrieval features and improving localization accuracy.

Research Conclusions:

– CIPER demonstrates competitive performance on VIGOR, KITTI, and Ford Multi-AV datasets, particularly in scenarios with limited field-of-view and variable orientations.

Paper link: https://huggingface.co/papers/2606.05011

34. PIPE-Cypher: Automatic Enterprise Benchmark Generation for Text-to-Cypher Systems

Keywords: Local Benchmark-Generation Pipeline, Property Graphs, Text2Cypher, Execution Validation, Diversity Controls

Category: Knowledge Representation and Reasoning

Research Objective:

– To develop PIPE-Cypher, a local benchmark-generation pipeline that transforms live property graphs and seed queries into balanced NL-to-Cypher datasets for enterprise knowledge graphs.

Research Methods:

– Utilizes processes such as schema profiling, reverse-query grounding, constrained generation, execution validation, and the employment of a calibrated local LLM judge, utilizing local Qwen3.5-9B generation and judging.

Research Conclusions:

– PIPE-Cypher consistently creates repeatable and adaptable Text2Cypher benchmarks. It highlights that zero-shot transfer is limited whereas schema-specific example banks enhance compatible model performances.

Paper link: https://huggingface.co/papers/2606.08481

35. Send a SCOUT First: Pre-hoc Reasoning for Adaptive Detector Allocation in Prompt-Injection Defense

Keywords: SCOUT framework, prompt-injection detection, detector allocation, safety-utility threshold, SCOUT-450

Category: AI Systems and Tools

Research Objective:

– The SCOUT framework aims to dynamically allocate prompt-injection detection by predicting the reliability and latency of detectors to improve safety and efficiency over single-detector approaches.

Research Methods:

– The framework reframes defense as detector allocation, deciding which detectors to run per request and whether to escalate to an LLM judge, using predictions based on past detector behavior.

Research Conclusions:

– SCOUT reduces the attack-success rate by 46% and total wall-clock time by 40%, with only a 5.1-point drop in benign utility, when compared to an always-on GPT-4o judge. It also shows improved performance on external benchmarks, enhancing the safety-utility frontier.

Paper link: https://huggingface.co/papers/2605.30837

36. Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data

Keywords: Self-Evaluation Elicitation, Reinforcement Learning, Calibration, Masked Distillation, Transferable Quality Evaluation

Category: Reinforcement Learning

Research Objective:

– The objective is to improve model calibration for quality assessment through a novel method called Self-Evaluation Elicitation (SEE).

Research Methods:

– The SEE method employs calibration-coupled reinforcement learning and masked distillation in a short cycle to enhance prediction accuracy whilst maintaining answer quality.

Research Conclusions:

– The SEE method successfully surfaces a model’s latent ability to predict judge scores beyond specific preferences, demonstrating a transferable quality evaluation on various benchmarks.

Paper link: https://huggingface.co/papers/2606.05122

37. SigmaScale: LLM Compression with SVD-based Low-Rank Decomposition and Learned Scaling Matrices

Keywords: Large Language Model, compression, scaling matrices, activation-aware compression, effective-rank entropy

Category: Natural Language Processing

Research Objective:

– To present SigmaScale, a method for learning auxiliary scaling matrices to improve the compression of Large Language Models using truncated Singular Value Decomposition.

Research Methods:

– SigmaScale optimizes vectors for diagonal row and column scaling transformations based on activation-aware compression loss, lowering the effective intrinsic rank of weight matrices.

Research Conclusions:

– SigmaScale demonstrates competitive performance with state-of-the-art SVD-based compression methods across benchmarks, offering a flexible route for low-rank LLM compression, which is beneficial in reducing LLM-inference computing costs.

Paper link: https://huggingface.co/papers/2606.07098

38. Skill-3D: Evolving Scene-Aware Skills for Agentic 3D Spatial Reasoning

Keywords: Skill-3D, 3D spatial reasoning, scene-aware skills, tool utilization, self-evolving

Category: Knowledge Representation and Reasoning

Research Objective:

– The paper introduces the Skill-3D framework which aims to improve agent performance in 3D spatial reasoning tasks by developing scene-aware skills through a self-evolving system.

Research Methods:

– Skill-3D utilizes a self-evolving memory and skill library to develop and refine scene-aware skills. This system tracks tool-use trajectories across different scenes, distilling successful ones into reusable skills and using failed attempts as learning lessons.

Research Conclusions:

– Experiments demonstrate significant improvements in tool utilization in 3D reasoning tasks, such as a 67% enhancement for Gemini-3-Flash on MMSI-Bench and a 43% improvement for Qwen3-VL-8B on VSI-Bench, highlighting the effectiveness of skill-guided tool use strategies.

Paper link: https://huggingface.co/papers/2606.07436

39. EmpiriGraph-Psy: A Dataset and LLM Pipeline for Extracting Empirical Relation Graphs from Psychology Abstracts

Keywords: Empirical Graph Extraction, Variable-Centered, Psychology, Staged Pipeline, Typed Graphs

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to map psychology abstracts to typed graphs using normalized variables and empirical relations, specifically targeting variable-oriented empirical fields like psychology to bridge existing gaps in scientific relation extraction benchmarks.

Research Methods:

– A staged pipeline approach is employed for graph extraction, which involves separate steps for variable extraction, normalization, hierarchy construction, evidence selection, relation extraction, and edge validation. This method is compared against direct extraction methods.

Research Conclusions:

– The staged pipeline approach significantly improves performance, achieving a macro-F1 of 0.74, though challenges in moderating relations and concept hierarchies remain, particularly in extracting higher-order empirical claims from abstracts.

Paper link: https://huggingface.co/papers/2606.08362

40. PBSD: Privileged Bayesian Self-Distillation for Long-Horizon Credit Assignment

Keywords: Privileged Bayesian Self-Distillation, Credit Assignment, Reinforcement Learning, Bayesian Evidence Scoring, Autoregressive Decomposition

Category: Reinforcement Learning

Research Objective:

– Introduce Privileged Bayesian Self-Distillation to enable fine-grained credit assignment in long-horizon tasks by converting sparse rewards into calibrated turn-level signals.

Research Methods:

– Utilization of Bayes’ rule to transform the posterior-to-prior probability ratio into a tractable likelihood ratio between a student model and a privileged teacher model.

– Implementation of autoregressive decomposition to derive turn-level signals from Bayesian evidence scoring.

Research Conclusions:

– PBSD enhances performance across various settings and facilitates effective policy learning and improved generalization by transforming sparse outcome supervision into Bayes-calibrated credit signals.

Paper link: https://huggingface.co/papers/2606.09348

41. Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory

Keywords: SkeMex, Medical Agents, Skill Memory, Clinical Decision Making, Contextual Utility

Category: AI in Healthcare

Research Objective:

– To develop SkeMex, a self-evolving framework that enhances medical agent systems through structured skill memory to improve long-term clinical reasoning.

Research Methods:

– Utilization of a post-deployment self-evolution framework distilling informative interaction trajectories into structured skills, organized in a multi-branch repository and governed by context-dependent utility.

Research Conclusions:

– SkeMex outperforms existing memory-based agents in clinical tasks, generalizes across model backbones, and supports the adaptation of transferable skill memory.

Paper link: https://huggingface.co/papers/2606.09365

42. Trust Functions: Near-Lossless Weak-to-Strong Generalization by Learning When to Trust the Weak Teacher

Keywords: Trust Functions, Weak-to-Strong Generalization, Reliable Labels, Data Selection

Category: Machine Learning

Research Objective:

– To enhance weak-to-strong generalization by leveraging trust functions to identify reliable weak labels for training across various domains.

Research Methods:

– Introduce trust functions that assign trust scores to weak labels, using these scores to filter weak supervision and enable iterative training chains.

Research Conclusions:

– Trust functions enable students to match or exceed performance with ground-truth supervision, facilitating an effective weak-to-strong generalization process.

Paper link: https://huggingface.co/papers/2606.01000

43. Phase Marginalization for Patch-Grid Instability in Vision Transformers

Keywords: Phase Marginalization, Vision Transformers, patch-grid phase, dense prediction, Uniform Phase Marginalization

Category: Computer Vision

Research Objective:

– The objective is to address phase-dependent instability in Vision Transformers by proposing the novel method of Phase Marginalization for evaluating structured patch-grid phases and aggregating outputs in the original image coordinate system.

Research Methods:

– A post-hoc marginalization approach called Phase Marginalization is formalized to handle patch-grid phases for dense predictions without additional training. This method includes evaluating structured patch-grid phases and inverse-aligning dense outputs.

Research Conclusions:

– This research demonstrates that Uniform Phase Marginalization with K = 4 surpasses the traditional K = 1 baseline in segmentation, depth, and local matching experiments. It provides better performance with a modest compute-matched advantage over generic test-time augmentation in Cityscapes experiments. The study also highlights that using K = 8 or K = 16 offers minimal accuracy gain at higher costs.

Paper link: https://huggingface.co/papers/2606.08132

44. Optical Reasoning: Rethinking Images as an Expressive Reasoning Medium Beyond Text

Keywords: Optical reasoning, Chain-of-Thought, Large Language Models, Multimodal Large Language Models, Token efficiency

Category: Multi-Modal Learning

Research Objective:

– This study proposes the concept of optical reasoning, exploring the use of images as a standalone reasoning medium for language and multimodal tasks to achieve higher token efficiency compared to traditional text-based approaches.

Research Methods:

– The study introduces two variants of optical reasoning: typographic-based, which optimizes visual layouts, and graphical-based, which composes text and graphical elements into visual rationales.

Research Conclusions:

– Optical reasoning can match or surpass traditional text reasoning, reducing reasoning tokens by 28.57% on language tasks and 16% on multimodal tasks, thus enhancing token efficiency by 1.96 times. This indicates that images can effectively encode rationales while providing a unified visual platform for reasoning.

Paper link: https://huggingface.co/papers/2606.09585

45. Robotic Policy Adaptation via Weight-Space Meta-Learning

Keywords: WIZARD, Vision-Language-Action models, LoRA parameters, meta-learning, task adaptation

Category: Robotics and Autonomous Systems

Research Objective:

– The objective is to develop WIZARD, a framework providing task-specific adaptation for Vision-Language-Action models without requiring fine-tuning, utilizing language instructions and demonstration videos.

Research Methods:

– WIZARD operates by predicting task-specific LoRA parameters in a single forward pass, using language instructions and demonstration videos during the meta-training phase to generate expert LoRA updates without target-task action labels.

Research Conclusions:

– Experiments on LIBERO demonstrate that WIZARD significantly improves performance, enhancing results by up to ~2x on unseen datasets and up to ~14x on unseen tasks, especially on a Franka Emika Panda, where WIZARD surpasses a real-domain adapted baseline.

Paper link: https://huggingface.co/papers/2606.07217

46. Text-to-Image Models Need Less from Text Encoders Than You Think

Keywords: Text-to-image models, text embeddings, diffusion transformer-based models, visual quality, text fidelity

Category: Generative Models

Research Objective:

– To explore which aspects of text representation are essential for image generation in text-to-image models.

Research Methods:

– Developed a new text embedding that captures only individual word meanings and order, lacking complete contextual information, to evaluate its impact on image generation.

Research Conclusions:

– Text-to-image models primarily depend on simple text representation aspects like word merging and order, rather than exploiting richer contextual information. The study finds that such simplified text embeddings can still guide image generation successfully, maintaining high visual quality and text fidelity.

Paper link: https://huggingface.co/papers/2606.03715

47. Answer Presence Drives RAG Rewriting Gains

Keywords: QA performance, gold answer, intervention audit, LLM rewriter, sentinel changes

Category: Natural Language Processing

Research Objective:

– The study investigates the causal factors behind the performance boost in multi-hop QA systems, specifically determining whether the presence of the gold answer in rewritten contexts is the main driver.

Research Methods:

– The researchers conducted controlled interventions where they manipulated rewritten contexts by either removing or injecting the gold answer and assessed the impact on QA performance across multiple reader configurations and datasets.

Research Conclusions:

– The presence of the gold answer in rewritten contexts significantly enhances QA performance, with its removal causing notable F1 decrease, and injection causing improvement. Conventional probing methods demonstrated fragility to sentinel changes.

Paper link: https://huggingface.co/papers/2606.05633

48. Reasoning Arena: Trace Tournaments When Verifiable Rewards Fall Short

Keywords: Reinforcement Learning, Verifiable Rewards, Reasoning Arena, Trace Tournaments, Bradley-Terry Models

Category: Reinforcement Learning

Research Objective:

– To enhance reinforcement learning for large language models by introducing a more informative reward system through the Reasoning Arena framework.

Research Methods:

– Implementing trace tournaments to differentiate reasoning quality among non-diverse reward groups.

– Utilizing a judge system and dynamically updated trace pools for efficient relative ranking.

– Applying Bradley-Terry models on incomplete comparison graphs to facilitate scalable RL integration.

Research Conclusions:

– Reasoning Arena outperforms the baseline RLVR by 7.6% in reasoning tasks, accelerates training speed by 27% to 41%, and reduces computation by nearly 50%.

Paper link: https://huggingface.co/papers/2606.09380

49. Why Muon Outperforms Adam: A Curvature Perspective

Keywords: Muon, Adam, large language model training, curvature penalty, Normalized Directional Sharpness (NDS)

Category: Natural Language Processing

Research Objective:

– The study aims to uncover the reasons behind Muon’s superior performance over Adam in large language model training, focusing on curvature perspectives.

Research Methods:

– Application of second-order Taylor approximation to the training landscape.

– Analysis of curvature penalties through decomposition into components like squared update norm and NDS.

– Investigation of training data imbalance using Zipf-Probabilistic Context-Free Grammar (PCFG).

Research Conclusions:

– Muon exhibits a larger one-step loss decrease and incurs a smaller second-order curvature penalty than Adam, attributed to lower NDS.

– Data imbalance and heterogeneous curvature conditions amplify Muon’s advantages.

Paper link: https://huggingface.co/papers/2606.04662

50. OmniCap-IF: Benchmarking and Improving Instruction Following Abilities for Omni-Video Captioning

Keywords: OmniCap-IF, omni-modal captioning, instruction-following, format-content tradeoff, Temporal Grounding

Category: Multi-Modal Learning

Research Objective:

– The paper introduces OmniCap-IF, the first comprehensive benchmark to evaluate instruction-following capabilities in omni-modal captioning, addressing the gap in assessing multi-modal reasoning under complex user instructions.

Research Methods:

– A systematic framework is employed to evaluate captions on format correctness and content correctness across 50 distinct constraint types in pure visual, pure audio, and audio-visual modalities, including Temporal Grounding for spatio-temporal precision.

Research Conclusions:

– The study reveals significant performance disparities among models and identifies a critical “format-content tradeoff,” explaining that increased formatting complexity degrades omni-modal reasoning. OmniCaptioner-IF, a new model, demonstrates notable improvements through a curated 54K instruction-tuning dataset.

Paper link: https://huggingface.co/papers/2606.08572

51. Skill-RM: Unifying Heterogeneous Evaluation Criteria via Agent Skill

Keywords: Reward Models, Reinforced Fine-Tuning, Reinforcement Learning, Structured Agentic Task, Evidence Aggregation

Category: Reinforcement Learning

Research Objective:

– Introduce a unified reward modeling framework called Skill-RM that treats reward computation as a structured agentic task.

Research Methods:

– Utilizes a consistent interface for orchestrating heterogeneous resources, dynamically selecting and aggregating evidence based on specific input requirements.

Research Conclusions:

– Skill-RM outperforms traditional judge baselines in reward benchmarks and downstream applications by providing a unified solution for reward modeling with superior performance.

Paper link: https://huggingface.co/papers/2606.03980

52. Bayesian-Agent: Posterior-Guided Skill Evolution for LLM Agent Harnesses

Keywords: Bayesian-Agent, SOPs, hypotheses, task performance, model success

Category: AI Systems and Tools

Research Objective:

– The study aims to introduce Bayesian-Agent, a framework using Bayesian inference for enhancing agent behavior and task performance by treating reusable skills and SOPs as hypotheses for success.

Research Methods:

– Utilizes Bayesian inference to guide agent behavior and optimize task performance through posterior-guided harness optimization. The framework records trajectory evidence and maintains a categorical posterior over each skill.

Research Conclusions:

– Bayesian-Agent framework improves the performance of different benchmarks significantly, suggesting that agent skill evolution is more effective when viewed as posterior-guided harness optimization rather than uncalibrated prompt accumulation.

Paper link: https://huggingface.co/papers/2606.08348

53. AHA-WAM:Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing

Keywords: AHA-WAM, dual Diffusion Transformers, asynchronous world-action model, horizon-adaptive offset training, Observation-Guided Video-Context Routing

Category: Robotics and Autonomous Systems

Research Objective:

– To develop an Asynchronous Horizon-Adaptive World-Action Model (AHA-WAM) for efficient long-horizon planning and real-time action execution in robotic manipulation tasks.

Research Methods:

– Utilization of dual Diffusion Transformers architecture to decouple temporal resolutions for world prediction and action execution.

– Implementation of horizon-adaptive offset training and Observation-Guided Video-Context Routing for asynchronous execution.

Research Conclusions:

– AHA-WAM achieved state-of-the-art performance, with 92.80% success on RoboTwin and 78.3% success in real-world tasks without robot-data pretraining.

– The model demonstrated a 24.17 Hz closed-loop control with a 4.59x speedup over Fast-WAM.

Paper link: https://huggingface.co/papers/2606.09811

54. OmniGameArena: A Unified UE5 Benchmark for VLM Game Agents with Improvement Dynamics

Keywords: Vision-language model, OmniGameArena, reflection-based improvement, Unreal Engine 5, Improvement Dynamics Curve

Category: Multi-Modal Learning

Research Objective:

– To establish a unified benchmark, OmniGameArena, for evaluating Vision-language model (VLM) agents across diverse game settings to track their performance evolution and skill generalization.

Research Methods:

– Introduced twelve games built with Unreal Engine 5 encompassing Solo, PvP, and Coop modes.

– Developed a unified action interface and Improvement Dynamics Curve (IDC) which uses a tool-using reflector LLM to refine bounded skill prompts over multiple rounds.

Research Conclusions:

– Demonstrated the effectiveness of IDC by reporting performance metrics on a cold-start leaderboard and additional observables, showcasing how agent scores evolve and how skills generalize across tasks.

Paper link: https://huggingface.co/papers/2606.09826

55. Echo-Memory: A Controlled Study of Memory in Action World Models

Keywords: Echo-Memory, memory mechanisms, action-conditioned world models, replay quality, state-space recurrence

Category: Generative Models

Research Objective:

– To investigate the impact of memory structure and capacity on the performance of action-conditioned world models.

Research Methods:

– Conducted a controlled study using Echo-Memory, varying only the memory storage and retrieval mechanisms while keeping other factors like the video diffusion backbone constant.

Research Conclusions:

– Raw context provides a robust capacity baseline, significantly enhancing open-domain return performance.

– Compact memory designs, while efficient, can lead to the loss of essential evidence for accurate memory recall.

– State-space recurrence stands out as the most effective mechanism for open-domain returns, demonstrating the critical role of implicit memory structure.

Paper link: https://huggingface.co/papers/2606.09803

56. SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks

Keywords: SpatialWorld, multimodal agents, spatial reasoning, partial observability, text-based actions

Category: Multi-Modal Learning

Research Objective:

– Introducing SpatialWorld as a unified benchmark to evaluate interactive spatial understanding in multimodal agents through diverse real-world tasks.

Research Methods:

– Integration of eight heterogeneous simulation backends under a unified protocol, enabling tasks with vision-only partial observability and decision-making via a text-based action interface.

Research Conclusions:

– Highlighting challenges in robust spatial task solving, with the most advanced model achieving a low task success rate, revealing inefficiencies and performance variations across domain-specific tasks.

Paper link: https://huggingface.co/papers/2606.09669

57. CoVEBench: Can Video Editing Models Handle Complex Instructions?

Keywords: CoVEBench, compositional video editing, multi-point editing instructions, video fidelity, video quality

Category: Computer Vision

Research Objective:

– The paper introduces CoVEBench, a benchmark designed to assess the capabilities of current models in compositional video editing, specifically focusing on handling complex and multi-step editing tasks while preserving spatiotemporal content.

Research Methods:

– CoVEBench consists of 416 curated source videos, 626 multi-point editing instructions, and 9,990 fine-grained checklist items to evaluate models on instruction compliance and video fidelity using automated metrics for assessing video quality.

Research Conclusions:

– The study finds that compositional video editing remains challenging; models often fail to implement all edits correctly, breach preservation constraints, or generate artifacts when executing multiple operations at the same time.

Paper link: https://huggingface.co/papers/2606.08415

58. LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents

Keywords: LatentSkill, LoRA adapters, weight space, semantic geometry, parameter-space arithmetic

Category: AI Systems and Tools

Research Objective:

– The objective of the research is to develop LatentSkill, a framework designed to efficiently convert textual skills into LoRA adapters for agent systems, reducing context overhead while maintaining modularity and composability.

Research Methods:

– The researchers utilized LatentSkill to transform textual skills into plug-and-play LoRA adapters using a pretrained hypernetwork, allowing these skills to be stored in weight space instead of context space.

Research Conclusions:

– LatentSkill was shown to outperform in-context skill baselines in specific benchmarks, achieving significant improvements in task success and efficiency in ALFWorld and Search-QA. It demonstrated that weight-space skills are efficient, modular, and offer less exposure compared to context-space skills.

Paper link: https://huggingface.co/papers/2606.06087

59. On the Geometry of On-Policy Distillation

Keywords: On-policy distillation, Parameter space, Subspace locking, Reinforcement learning, Supervised fine-tuning

Category: Natural Language Processing

Research Objective:

– The research investigates the unique geometric patterns in parameter space dynamics of On-policy distillation (OPD) and compares them with supervised fine-tuning and reinforcement learning with verifiable rewards.

Research Methods:

– The study uses parameter-space diagnostics to compare the trajectory of OPD updates in parameter space with other methods, highlighting subspace locking and relaxed off-principal updates.

Research Conclusions:

– OPD forms a distinct update geometry, characterized by fewer weight updates and subspace locking, which is functionally sufficient for OPD but not for supervised fine-tuning. The study highlights how OPD’s dynamics are unique and not merely intermediate between other methods.

Paper link: https://huggingface.co/papers/2606.07082

Keywords: Large language models, EmbedFilter, text embeddings, high-frequency tokens, dimensionality reduction

Category: Natural Language Processing

Research Objective:

– This study aims to address the deficiency in large language models’ embedding capabilities by introducing EmbedFilter, a linear transformation that enhances semantic representations and enables dimensionality reduction.

Research Methods:

– The authors identified that text embeddings tend to align with frequent, uninformative tokens, and thus, applied EmbedFilter to suppress the influence of these high-frequency tokens, refining the semantic quality of the embeddings.

Research Conclusions:

– Experiments show that models integrated with EmbedFilter achieve better zero-shot performance on downstream tasks, even with reduced embedding dimensions, suggesting enhanced efficiency and quality of semantic representations.

Paper link: https://huggingface.co/papers/2606.07502

2. AnchorWorld: Embodied Egocentric World Simulation with View-based Evolution Customization

Keywords: egocentric simulation, 3D human motion, spatial grounding, self-evolving worlds, anchor views

Category: Computer Vision

Research Objective:

– To enhance egocentric simulation through improved interaction integrity and world customization using 3D human motion and anchor view definitions.

Research Methods:

– Utilization of 3D human motion as the primary interaction modality, and incorporation of auxiliary training supervision with exogenous viewpoints to improve spatial grounding of human-world interactions.

– Introduction of a mechanism for customizing self-evolving worlds by defining anchor views within a unified world coordinate system and using textual descriptions for dynamic evolution of scenes.

Research Conclusions:

– AnchorWorld outperforms state-of-the-art baselines, with ablation studies supporting the effectiveness of its designs. The proposed customization scheme shows promising spatio-temporal geometric consistency and adheres strictly to the prescribed evolutionary dynamics.

Paper link: https://huggingface.co/papers/2606.07326

3. GENEB: Why Genomic Models Are Hard to Compare

Keywords: GENEB, genomic foundation models, diagnostic benchmark, probing-based protocol, model rankings

Category: Foundations of AI

Research Objective:

– To introduce GENEB, a comprehensive benchmark for evaluating genomic foundation models across diverse tasks and architectures under a unified protocol.

Research Methods:

– Implementation of a large-scale diagnostic benchmark called GENEB, which evaluates frozen representations from 40 genomic foundation models across 100 tasks spanning 13 functional categories using a unified probing-based protocol.

Research Conclusions:

– Current evaluation practices show limitations, as model rankings vary sharply across task categories, with scale providing modest and inconsistent gains. GENEB is positioned as a reference framework for principled comparison and category-aware model selection in genomic machine learning.

Paper link: https://huggingface.co/papers/2606.04525

4. Robots Need More than VLA and World Models

Keywords: Generalist robot intelligence, unstructured behavioral data, embodiment mapping, world modeling, reward inference

Category: Robotics and Autonomous Systems

Research Objective:

– The paper argues for a shift from focusing solely on policy scaling to incorporating unstructured behavioral data through specialized interfaces to enhance robot intelligence.

Research Methods:

– The study highlights the need for interfaces for autolabelling unstructured behavior, retargeting human motion, 3D reasoning for world modeling, and inferring rewards.

Research Conclusions:

– The authors propose a research agenda for building robotic systems capable of learning from the broader physical world, not just robot demonstrations.

Paper link: https://huggingface.co/papers/2606.06556

5. OpenSkill: Open-World Self-Evolution for LLM Agents

Keywords: OpenSkill, self-evolving agents, open-world deployment, verification signals, transferable skills

Category: Reinforcement Learning

Research Objective:

– This study introduces OpenSkill, aiming to enable agents to develop skills and verification signals independently using open-world resources, without relying on target-task supervision.

Research Methods:

– OpenSkill employs a framework to bootstrap the learning loop by acquiring grounded knowledge and verification anchors from various sources and synthesizing them into transferable skills.

Research Conclusions:

– OpenSkill showcases high automated performance across benchmarks without breaching the no-supervision constraint, effectively transferring skills across models and aligning self-built verifiers with ground-truth outcomes.

Paper link: https://huggingface.co/papers/2606.06741

6. UniSHARP: Universal Sharp Monocular View Synthesis

Keywords: AI Native, universal monocular rendering, omnidirectional latent space, Gaussian primitives, photorealistic view synthesis

Category: Computer Vision

Research Objective:

– The study aims to extend SHARP for universal monocular rendering across various camera systems by aligning images in an omnidirectional latent space.

Research Methods:

– UniSHARP is proposed, which performs implicit alignment in both feature and Gaussian spaces using Gaussian primitives arranged in a ray-based universal representation. A benchmark stratified by field of view is constructed for evaluation.

Research Conclusions:

– UniSHARP demonstrates superior performance in universal monocular rendering across diverse imaging systems, outperforming alternative methods by a large margin.

Paper link: https://huggingface.co/papers/2606.07514

7. LIMMT: Less is More for Motion Tracking

Keywords: Motion Tracking, High-Quality Data, Data-Centric Study, Physics-Based Humanoid Motion Tracking, Data Cleaning

Category: Robotics and Autonomous Systems

Research Objective:

– The objective is to improve tracking policy optimization for physics-based humanoid motion tracking using high-quality motion data, specifically by utilizing minimal data subsets to outperform full datasets.

Research Methods:

– Introduction of LIMMT (Less Is More for Motion Tracking) as a framework, focusing on data quality defined by physics feasibility, diversity, and complexity. Data cleaning on web-sourced mocap data was also conducted.

Research Conclusions:

– Demonstrated that less than 3% of the AMASS dataset yields better tracking performance than the full dataset, and extensive experiments validate the effectiveness of the LIMMT framework.

Paper link: https://huggingface.co/papers/2606.06953

8. dots.tts Technical Report

Keywords: Continuous Autoregressive, AudioVAE, Flow-Matching Head, Low-Latency Speech Generation

Category: Generative Models

Research Objective:

– The goal is to develop a state-of-the-art continuous autoregressive text-to-speech model, dots.tts, capable of efficient low-latency speech generation across multiple languages.

Research Methods:

– Utilizes a novel training approach with AudioVAE for a semantically structured continuous speech space.

– Incorporates full-history conditioning and reward-free self-corrective post-training to enhance robustness and acoustic quality.

– Applies CFG-aware MeanFlow distillation to minimize latency in speech generation.

Research Conclusions:

– The model, trained on a large multilingual corpus, shows superior performance on Seed-TTS-Eval benchmark with impressive WERs and SIM scores.

– Achieves open-source state-of-the-art results on multiple benchmarks, showcasing strong stability, voice cloning, and emotional expressiveness.

– Efficient inference is possible with dual-streaming modes, facilitating practical deployment and reproducible research.

Paper link: https://huggingface.co/papers/2606.07080

9. PaperFlow: Profiling, Recommending, and Adapting Across Daily Paper Streams

Keywords: PaperFlow, scientific paper recommendation, profiling, interest drift, multi-signal aggregation

Category: AI Systems and Tools

Research Objective:

– To develop a framework called PaperFlow for recommending scientific papers by processing user profiles, daily paper streams, and addressing interest drift through a three-stage process.

Research Methods:

– Implemented a longitudinal benchmark with 24 users, 50 daily streams, and 1,200 episodes to evaluate PaperFlow.

– Organized the framework into three stages: Profiling, Recommending, and Adapting.

Research Conclusions:

– PaperFlow demonstrates superior oracle-based ranking, high behavioral alignment with simulated reading selections, and outperforms scientific recommendation baselines in blind human-evaluation scores.

Paper link: https://huggingface.co/papers/2606.07454

10. Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators

Keywords: Astra, Vision-Language Models, action-conditioned visual imagination, world simulator, spatial reasoning

Category: Multi-Modal Learning

Research Objective:

– To enhance Vision-Language Models with action-conditioned visual imagination through a spatial reasoning framework called Astra.

Research Methods:

– Astra employs a reinforcement learning-trained policy coupled with a Bagel-based world simulator to generate novel-view observations, utilizing view consistency tuning and a world-simulator-in-the-loop two-phase RL curriculum.

Research Conclusions:

– The Astra framework significantly improves spatial reasoning by providing useful imagined observations, demonstrating improvements on benchmarks such as MMSI-Bench and MindCube; effective reasoning requires learning the optimal use of imagined evidence.

Paper link: https://huggingface.co/papers/2606.06476

11. Stream3D-VLM: Online 3D Spatial Understanding with Incremental Geometry Priors

Keywords: 3D vision-language model, autoregressive control modeling, Visual-Spatial Feature Integration, Geometry-Adaptive Voxel Compression

Category: Computer Vision

Research Objective:

– The paper presents an online 3D vision-language model aimed at achieving real-time spatial understanding from streaming video.

Research Methods:

– Utilizes autoregressive streaming control modeling to determine response timing.

– Employs a Visual-Spatial Feature Integration (VSFI) module to incrementally inject geometry priors.

– Proposes a Geometry-Adaptive Voxel Compression (GAVC) module for efficient visual token compression.

Research Conclusions:

– Extensive experiments demonstrate the model’s superior performance over existing proprietary and open-source models in tasks related to 3D spatial understanding, reasoning, and grounding.

Paper link: https://huggingface.co/papers/2606.06891

12. SIA: Self Improving AI with Harness & Weight Updates

Keywords: Self-Improving AI, Language-Model Agent, Task-Specific Agent, GPU Optimization, Biological Data Denoising

Category: AI Systems and Tools

Research Objective:

– To develop a self-improving AI framework that can update both the model weights and task-specific agent architecture using a language-model feedback agent across diverse tasks.

Research Methods:

– Introduced SIA, a self-improving loop that simultaneously updates the harness and weights of a task-specific agent across different domains such as legal classification, GPU optimization, and biological data denoising.

Research Conclusions:

– The proposed method, combining both harness and weight updates, outperformed traditional scaffold-only iterations across various tasks, achieving significant improvements in benchmarks like LawBench, GPU kernel runtime, and RNA denoising.

Paper link: https://huggingface.co/papers/2605.27276

13. Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation

Keywords: post-hoc compression, knowledge distillation, accuracy-efficiency trade-off, reasoning traces

Category: Knowledge Representation and Reasoning

Research Objective:

– To explore the benefits of post-hoc compression of reasoning traces for more efficient and cost-effective knowledge distillation.

Research Methods:

– Two instruction-tuned models were used to compress reasoning traces from large teacher models, reducing them to 8.6-21.0% of their original length. Experiments conducted included main grid runs and truncation ablations to compare efficiency and accuracy.

Research Conclusions:

– Compressed traces significantly reduce training time and token usage while maintaining a high level of accuracy. Raw traces retain the highest accuracy, but compressed models provide substantial efficiency improvements, up to 18x per token efficiency, especially beneficial for smaller models.

Paper link: https://huggingface.co/papers/2606.05988

14. ECI_{sem}: Semantic Residual Effective Contrastive Information for Evaluating Hard Negatives

Keywords: ECI_sem, dense retrieval, semantic residual, BEIR, MS MARCO

Category: Natural Language Processing

Research Objective:

– Introduce ECI_sem, a semantic residual variant of Effective Contrastive Information, to rank negative sources for dense retrieval without training, using frozen embeddings.

Research Methods:

– ECI_sem constructs a weighted residual information matrix based on target consistency, semantic locality, lexical residuality, and log-determinant diversity.

Research Conclusions:

– ECI_sem achieves strong performance on MS MARCO and BEIR benchmarks, with high alignment depending on the target encoder and stability under various perturbations.

Paper link: https://huggingface.co/papers/2603.20990

15. Towards Retrieving Interaction Spaces for Agentic Search

Keywords: RISE framework, BM25, agentic search, corpus exploration, interaction space

Category: AI Systems and Tools

Research Objective:

– The research aims to develop the RISE framework for efficient corpus exploration by constructing bounded interaction spaces that maintain high accuracy at scale.

Research Methods:

– The study combines BM25 retrieval with preprocessed document indexing to create an interaction space for agentic search, optimizing for shell-style navigation.

Research Conclusions:

– The RISE framework, when evaluated on BrowseComp-Plus, demonstrated comparable accuracy to the pure-shell DCI baseline at 78% accuracy with lesser costs and outperformed it significantly in larger corpus settings, achieving 81% accuracy on a 1M document set.

Paper link: https://huggingface.co/papers/2606.06880

16. A Cookbook of 3D Vision: Data, Learning Paradigms, and Application

Keywords: 3D vision, geometric representations, learning frameworks, datasets, multimodal geometric grounding

Category: Computer Vision

Research Objective:

– This study aims to create a data-centric taxonomy for 3D vision, integrating key elements such as geometric representations, datasets, learning frameworks, and applications into a unified conceptual map.

Research Methods:

– The methods include analyzing various structural representations of 3D data like point clouds, meshes, voxels, and 3D Gaussians, and exploring dataset design, benchmark construction, and supervision regimes.

Research Conclusions:

– The research provides a clarified view of the intricate interactions between representations, learning paradigms, and tasks, highlighting the trends toward balancing efficiency and fidelity and emphasizing multimodal geometric grounding.

Paper link: https://huggingface.co/papers/2606.04291

17. When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges

Keywords: Multi-objective LLM judge customization, textual gradients, Gradient specificity, instruction interference, Spearman’s rho

Category: Natural Language Processing

Research Objective:

– The study aims to customize a large language model (LLM) judge for specific tasks or domains by optimizing its prompts across multiple evaluation criteria using textual gradients.

Research Methods:

– The authors tested five decomposition modes of textual gradient optimizers by altering the shared cross-task information between loss, gradient, and optimizer LLMs.

Research Conclusions:

– The research identified two key failure modes: gradient dilution during optimization and instruction interference during inference, which limit the effectiveness of multi-objective judge customization with textual feedback.

Paper link: https://huggingface.co/papers/2605.26046

18. HarnessForge: Joint Harness and Policy Evolution for Adaptive Agent Systems

Keywords: LLM agents, meta-adaptation, HarnessForge, co-evolution, harness-conditioned policy alignment

Category: AI Systems and Tools

Research Objective:

– To address the challenges faced by LLM agents in heterogeneous task regimes by proposing a meta-adaptive framework, HarnessForge, which facilitates the co-evolution of agent systems.

Research Methods:

– Introduction of a stable adaptation space through harness–policy pairs separating execution structure from reasoning behavior, and employing fault-guided harness tailoring and harness-conditioned policy alignment for co-evolution.

Research Conclusions:

– HarnessForge enhances the performance of LLM agents like Qwen3-4B and Qwen3-8B, achieving up to 12.0% improvement over baselines, and emphasizes the importance of harmonizing harness and policy to optimize agent-system adaptability.

Paper link: https://huggingface.co/papers/2606.01779

19. Empirical Study on the Characteristics and Evolution of AI-usage in GitHub Repositories: Evidence from Code Comments

Keywords: AI tools, LLMs, code implementation, bug fixing, human oversight

Category: AI Systems and Tools

Research Objective:

– The study aims to analyze how AI tools, particularly LLMs, are utilized by developers in real-world software development workflows and the evolution of AI-assisted code over time.

Research Methods:

– Analysis of 35,361 GitHub code comments referencing AI use, deriving a taxonomy of AI-assisted development activities, and annotating the dataset using LLM-based classifiers. Additionally, the study examines 12,996 subsequent commit messages to understand the evolution of AI-assisted code.

Research Conclusions:

– The findings indicate that developers primarily use LLMs for tasks like code implementation, enhancement, debugging, and documentation, with sustained human oversight through refactoring and bug fixing. AI tools are increasingly seen as collaborative support mechanisms, shifting from direct code generation to enhancing conceptual support over time.

Paper link: https://huggingface.co/papers/2606.06843

20. Towards Human-Like Interactive Speech Recognition With Agentic Correction and Semantic Evaluation

Keywords: Interactive ASR, semantic correction, multi-turn refinement, Sentence-level Semantic Error Rate, reasoning-based editing

Category: Human-AI Interaction

Research Objective:

– The primary objective is to reduce semantic errors in Automatic Speech Recognition (ASR) through the integration of semantic correction and reasoning-based editing in a multi-turn refinement process.

Research Methods:

– The study introduces the Agentic ASR framework which combines a single-pass ASR front-end with semantic correction, intent routing, and reasoning-based editing, validated through a new Sentence-level Semantic Error Rate (S^2ER) metric and an Interactive Simulation System.

Research Conclusions:

– Iterative interaction in multilingual, named-entity-intensive, and code-switching benchmarks significantly reduces semantic errors more effectively in S^2ER than conventional token-level metrics, demonstrating enhanced alignment and robustness of the proposed framework.

Paper link: https://huggingface.co/papers/2605.29430

21. Measuring Model Robustness via Fisher Information: Spectral Bounds, Theoretical Guarantees, and Practical Algorithms

Keywords: Fisher Information Matrix, spectral norm, robustness metric, deep neural networks, adversarial vulnerability

Category: Foundations of AI

Research Objective:

– The paper aims to introduce a novel attack-agnostic robustness metric for deep neural networks using the spectral norm of the Fisher Information Matrix.

Research Methods:

– The study develops scalable evaluation methods such as power iteration and Hutchinson-based estimation for robustness assessment across different architectures, including VGG, ResNet, DenseNet, and Transformer in both white-box and black-box settings.

Research Conclusions:

– The research demonstrates a strong correlation between the proposed metric and adversarial vulnerability, suggesting that the framework serves as an interpretable diagnostic tool for complementing attack-based evaluations and guiding robust model design.

Paper link: https://huggingface.co/papers/2606.04767

22. WorldBench: A Challenging and Visually Diverse Multimodal Reasoning Benchmark

Keywords: WorldBench, Multimodal Large Language Models, Visual Diversity, Reasoning Benchmark

Category: Multi-Modal Learning

Research Objective:

– The objective is to introduce WorldBench, a visually diverse reasoning benchmark for evaluating Multimodal Large Language Models (MLLMs) and to reveal limitations in current models’ visual understanding capabilities.

Research Methods:

– The study involves constructing a taxonomy of thousands of visual concepts across multiple domains, and curating a broad collection of images from search engines and datasets to represent the visual world comprehensively. It uses structured trial-and-error to design challenging questions for MLLMs.

Research Conclusions:

– WorldBench demonstrates higher visual diversity compared to existing benchmarks, revealing weaknesses in visual understanding where even the strongest MLLMs only reach a 64.0% accuracy, emphasizing the importance of visual diversity in building multimodal benchmarks.

Paper link: https://huggingface.co/papers/2606.06538

23. The Distillation Game: Adaptive Attacks & Efficient Defenses

Keywords: Distillation attacks, Minimax game, Adaptive evaluation, Defense strategy, Product-of-Experts

Category: Machine Learning

Research Objective:

– To study the trade-off between model utility and vulnerability to imitation attacks through a minimax game framework.

Research Methods:

– Developed a minimax game between a utility-constrained teacher and an adaptive student to explore defensive strategies, including adaptive evaluation and a forward-pass-only defense called Product-of-Experts.

Research Conclusions:

– The adaptive student recovers more capabilities than passive evaluation reveals, narrowing the robustness gap between costly defenses and the cheaper Product-of-Experts.

– The study suggests that strong distillation prevention requires evaluation against adaptive students for progress in antidistillation efforts.

Paper link: https://huggingface.co/papers/2605.22737

24.

Paper link: 

25. Augmenting Attention with Exponentially Decaying Memory Improves Query-Aware KV Sparsity

Keywords: RAT+, Memory Module, Sparse Inference, Long-Context Language Models, Query-Aware

Category: Natural Language Processing

Research Objective:

– The study investigates whether the RAT+ memory module can enhance accuracy in query-aware sparse inference methods for long-context language models.

Research Methods:

– The paper utilizes RAT+ for various representative methods including Quest, MoBA, and SnapKV, validating the improvements in accuracy across different sparse budgets and tasks.

Research Conclusions:

– RAT+ consistently improves accuracy over standard attention in eight needle-in-a-haystack tasks and is verified on both RAT+ released checkpoints and continued pretraining on OLMo2-7B with a new memory module.

– Two hypotheses are proposed and supported by targeted experiments to explain the benefits of this memory module for query-aware sparse inference.

Paper link: https://huggingface.co/papers/2605.28640

26. Critic-R: Improving Agentic Search using Instruction-tuned Retrievers with Natural Language Introspective Feedback

Keywords: Agentic search, Retrieval models, Critic model, Feedback loop, Query refinement

Category: Knowledge Representation and Reasoning

Research Objective:

– The main aim is to enhance agentic search by improving the interaction between reasoning agents and retrieval models through a feedback loop mechanism.

Research Methods:

– Introduces Critic-R, a framework utilizing a critic model to evaluate reasoning and retrieval outcomes via dual optimization mechanisms: Critic-R-Zero for inference-time query refinement and Critic-Embed for optimizing retrieval models using automatic supervision.

Research Conclusions:

– Critic-R significantly improves retrieval quality and answer accuracy, as demonstrated by evaluations on HotpotQA, 2WikiMultihopQA, MuSiQue, and Bamboogle.

Paper link: https://huggingface.co/papers/2606.00590

27. How Far Can Chord-Symbol Time-Series Adaptation Carry Genre Identity? Capabilities and Boundaries in Multi-Genre Chord-Symbol Modeling

Keywords: Music Transformer, harmonic prediction, LoRA, IA3, genre adaptation

Category: Generative Models

Research Objective:

– To evaluate the effectiveness of small adaptation interfaces in extending a frozen Music Transformer model to handle multiple genres in harmonic prediction.

Research Methods:

– The study compared five methods, including LoRA and IA3, across a complete 165-cell grid for 11 genres and three seeds, analyzing improvements in chord prediction.

Research Conclusions:

– All methods improved the model over the base, with LoRA and IA3 scoring highest. Chord-symbol adaptation is shown to reliably enhance genre-local harmonic prediction, but it does not fully represent genre identity.

Paper link: https://huggingface.co/papers/2606.07334

28. CORE: Contrastive Reflection Enables Rapid Improvements in Reasoning

Keywords: Contrastive Reflection, reasoning tasks, verifiable rewards, natural-language insights

Category: Knowledge Representation and Reasoning

Research Objective:

– The paper aims to enhance AI Native reasoning capabilities in language models by using Contrastive Reflection to generate concise and interpretable insights for model self-improvement.

Research Methods:

– Introduced Contrastive Reflection (CORE), a non-parametric algorithm analyzing differences in reasoning traces to derive insights, allowing efficient and faster reasoning task improvements.

Research Conclusions:

– CORE demonstrated rapid and cost-effective performance improvements across various reasoning tasks compared to traditional parametric and non-parametric methods.

– It achieves comparable or superior outcomes with limited training samples and is more context-efficient, offering a more interpretable path to model self-improvement than existing approaches.

Paper link: https://huggingface.co/papers/2605.28742

29. Parametric Social Identity Injection and Diversification in Public Opinion Simulation

Keywords: Large language models, public opinion simulation, Diversity Collapse, Parametric Social Identity Injection, representation-level control

Category: Natural Language Processing

Research Objective:

– Address the issue of reduced social diversity in public opinion simulations with large language models by introducing a parametric framework to enhance demographic representation fidelity and diversity.

Research Methods:

– Implemented a Parametric Social Identity Injection framework to inject explicit demographic and value-oriented representations into LLMs.

– Conducted extensive experiments using the World Values Survey and multiple open-source LLMs.

Research Conclusions:

– The proposed method significantly improves the distributional fidelity and diversity of simulated public opinion data, reducing KL divergence and enhancing overall diversity, offering new insights into scalable, diversity-aware simulations.

Paper link: https://huggingface.co/papers/2603.16142

30. Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models

Keywords: Imaginative Perception Tokens, spatial reasoning, Vision language models, Perspective Taking, Path Tracing

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance vision-language models’ spatial reasoning capabilities by utilizing Imaginative Perception Tokens (IPT), which provide intermediate perceptual representations for improved interpretation of unseen viewpoints.

Research Methods:

– Three tasks, Perspective Taking, Path Tracing, and Multiview Counting, were formulated and tested using datasets with approximately 20K examples and the unified vision-language model BAGEL as the backbone.

Research Conclusions:

– IPT supervision enhances spatial reasoning and often outperforms traditional text-based methods, improving accuracy by 3.4% on Multiview Counting and showing competitive performance on Path Tracing. Combining IPT with label-only supervision further improves results, whereas textual chain of thought training could hinder performance.

Paper link: https://huggingface.co/papers/2606.03988

31. Data-Efficient Autoregressive-to-Diffusion Language Models via On-Policy Distillation

Keywords: Autoregressive Models, Diffusion Language Models, On-Policy Distillation, Train-Inference Mismatch, Bidirectional Attention

Category: Generative Models

Research Objective:

– To transform autoregressive language models (ARLMs) into diffusion language models (DLMs) using on-policy distillation to address train-inference mismatch and reduce training token requirements.

Research Methods:

– Implementing an On-Policy Diffusion Language Model (OPDLM) where on-policy distillation (OPD) is used for transforming ARLMs to DLMs, incorporating bidirectional attention for generating trajectories and using original ARLMs for knowledge distillation.

Research Conclusions:

– OPDLM significantly reduces the need for training tokens (15x to 7,000x fewer) while maintaining strong performance across various tasks, thus eliminating the high cost of DLM pretraining and improving knowledge retention from ARLMs.

Paper link: https://huggingface.co/papers/2606.06712

32. Streaming Video Generation with Streaming Force Control

Keywords: StreamForce, causal model, video generation, distillation pipeline, autoregressive efficiency

Category: Computer Vision

Research Objective:

– Introduce StreamForce, a streaming video generation framework that provides real-time, physically grounded responses to time-varying forces.

Research Methods:

– Design a unified force representation as a control signal and develop a distillation pipeline for force-controllable video generation.

– Combine autoregressive efficiency with force responsiveness to achieve stable photometric and dynamic realism.

Research Conclusions:

– StreamForce achieves state-of-the-art performance in force adherence and motion realism, running at up to 16.6 FPS on a single GPU.

Paper link: https://huggingface.co/papers/2606.07508

33. LayerRoute: Input-Conditioned Adaptive Layer Skipping via LoRA Fine-Tuning for Agentic Language Models

Keywords: LayerRoute, transformer blocks, LoRA adapters, inference, compute savings

Category: AI Systems and Tools

Research Objective:

– To develop a lightweight adapter, LayerRoute, that selectively skips transformer blocks during inference to save computational resources while maintaining or improving model quality.

Research Methods:

– Utilized LayerRoute with a per-layer router and LoRA adapters for gated routing on Qwen2.5-0.5B-Instruct, alongside a single end-to-end training pass on agentic data with gate regularization.

Research Conclusions:

– Achieved a 12.91% skip differential in FLOPs: tool calls skip 15.25% of FLOPs, while planning steps skip 2.34%. Quality improved over the base model due to LoRA adaptation, with reduced perplexity for both tool calls and planning steps.

Paper link: https://huggingface.co/papers/2606.01838

34. Entropy as a Structural Prior: How a Log-Barrier on DiT Belief Space Drives Musical Diversity and Development

Keywords: Confidence-based loss weighting, generative models, entropy, diffusion training, Stable Audio 3

Category: Generative Models

Research Objective:

– The study aims to improve audio generation through adaptive gradient scaling using confidence-based loss weighting in supervised diffusion training.

Research Methods:

– Introduces the Eisbach log-barrier, a parameter-free weight derived from the entropy of the DiT output’s spatial energy distribution, influencing gradient dampening or preservation.

– Applies this method to LoRA fine-tuning of Stable Audio 3 Medium on MusicCaps.

Research Conclusions:

– Achieves stronger thematic development, clearer acoustic differentiation, and higher textural diversity in audio generation.

– Demonstrates the emergence of a self-referential data curriculum purely from the forward pass with testable predictions.

Paper link: https://huggingface.co/papers/2606.07207

35. Reinforcement Learning from Rich Feedback with Distributional DAgger

Keywords: Forward Cross-Entropy, Distributional Imitation Learning, Monotonic Policy Improvement, Reasoning Tasks

Category: Reinforcement Learning

Research Objective:

– Enable monotonic policy improvement and enhance performance in reasoning tasks through forward cross-entropy objective with distributional imitation learning compared to traditional reinforcement learning methods.

Research Methods:

– Utilization of a distributional variant of the classic imitation learning algorithm DAgger, where learners have local access to expert distribution on visited states for the current policy.

Research Conclusions:

– Forward cross-entropy provides monotonic policy improvement and guarantees on regret, illustrating improvements over traditional RL and RL with self-distillation baselines in areas like scientific reasoning, coding, and solving complex mathematical problems.

Paper link: https://huggingface.co/papers/2606.05152

36. Almieyar-Oryx-BloomBench: A Bilingual Multimodal Benchmark for Cognitively Informed Evaluation of Vision-Language Models

Keywords: Vision-Language Models, multimodal benchmark, cognitive asymmetry, cross-lingual multimodal reasoning

Category: Multi-Modal Learning

Research Objective:

– Introduce BloomBench, a cognitively grounded bilingual multimodal benchmark to reveal cognitive asymmetries and cross-lingual performance gaps in Vision-Language Models.

Research Methods:

– Utilize Bloom’s Taxonomy to systematically evaluate six cognitive levels through image-question-answer tasks, employing a semi-automated pipeline and hybrid quality assurance protocol.

Research Conclusions:

– Identify strong performance in semantic understanding but weaknesses in factual recall and creative synthesis, highlighting cognitive asymmetries and performance gaps between languages in current models.

Paper link: https://huggingface.co/papers/2606.05531

37. SPACENUM: Revisiting Spatial Numerical Understanding in VLMs

Keywords: Vision-Language Models, Spatial Numerical Understanding, Coordinate-Aware Representations

Category: Multi-Modal Learning

Research Objective:

– The study revisits spatial numerical understanding of Vision-Language Models (VLMs) using the framework SpaceNum for evaluating map capabilities between spatial structures and numerical representations.

Research Methods:

– Formulated bidirectional tasks (Num2Space and Space2Num) to systematically study if VLMs understand numerical values in spatial settings.

Research Conclusions:

– Current VLMs fail to ground numerical values in spatial meaning, perform near random guesses, and continue to rely on shallow spatial cues instead of developing stable coordinate-aware representations. Explicit reasoning provides marginal improvements, but tuning can partially enhance spatial numerical understanding.

Paper link: https://huggingface.co/papers/2605.23898

38. Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills

Keywords: Socratic-SWE, self-evolving software engineering agents, historical solving traces, closed-loop self-evolution framework, repair patterns

Category: AI Systems and Tools

Research Objective:

– To enhance LLM-driven software engineering agents using Socratic-SWE, which generates targeted repair tasks by leveraging historical solving traces.

Research Methods:

– Implement a closed-loop self-evolution framework that distills historical solving traces into structured agent skills, guiding the generation of repair tasks.

– Validate tasks through execution-based validation and solver-gradient alignment rewards.

Research Conclusions:

– Socratic-SWE improves self-evolution in software engineering agents across various benchmarks, achieving significant performance gains on constrained budgets.

– The approach demonstrates that solving traces can be a scalable substrate for enhancing the capabilities of self-evolving SWE agents.

Paper link: https://huggingface.co/papers/2606.07412

39. Physics in 2-Steps: Locking Motion Priors Before Visual Refinement Erases Them

Keywords: PhaseLock, Image-to-Video diffusion models, physical consistency, motion priors, Latent Delta Guidance

Category: Generative Models

Research Objective:

– To improve physical consistency in image-to-video diffusion models by preserving motion priors during the denoising process.

Research Methods:

– Utilized a training-free approach called PhaseLock to maintain motion priors from early-step inference throughout the denoising trajectory, using spectral analysis and Latent Delta Guidance.

Research Conclusions:

– PhaseLock effectively mitigates phase degradation, improving physical consistency by an average of 6.2 points across diverse models, while preserving visual fidelity with minimal computational overhead.

Paper link: https://huggingface.co/papers/2606.06361

40. LLM Explainability with Counterfactual Chains and Causal Graphs

Keywords: Causal graphs, Large Language Models, concept discovery, counterfactual augmentation, concept-level explainability

Category: Knowledge Representation and Reasoning

Research Objective:

– The paper aims to model Large Language Model inference processes using causal graphs to enhance transparency and explainability.

Research Methods:

– A four-phase method involving concept discovery, mapping, and MCMC-inspired counterfactual augmentation is proposed to construct interpretable graphs, applied across various tasks including disease diagnosis and sentiment analysis.

Research Conclusions:

– The discovered causal graphs reflect meaningful dependencies aligned with LLMs’ reasoning, supporting concept-level explainability of language models.

Paper link: https://huggingface.co/papers/2606.05972

41. Watch, Remember, Reason: Human-View Video Understanding with MLLMs

Keywords: Multimodal large language models, video understanding, perceptual representations, memory modeling, reasoning traces

Category: Multi-Modal Learning

Research Objective:

– The paper aims to transform video understanding through Multimodal Large Language Models (MLLMs) that handle complex video scenarios by focusing on watching, remembering, and reasoning capabilities.

Research Methods:

– The study introduces a human-view perspective, organizing LLMs by their roles in video tasks: perceptual representation, memory states, and reasoning. Challenges are identified in areas such as spatio-temporal perception and memory modeling.

Research Conclusions:

– It offers insights into the future of scalable, memory-aware video intelligence, emphasizing the development of unified models for comprehensive video analysis and the exploration of application domains such as sports and medical videos.

Paper link: https://huggingface.co/papers/2606.07433

42. UnpredictaBench: A Benchmark for Evaluating Distributional Randomness in LLMs

Keywords: UnpredictaBench, Large Language Models, Simulation, Output Diversity, Distributional Sampling

Category: Natural Language Processing

Research Objective:

– To evaluate the capacity of large language models (LLMs) in simulating target distributions and assessing the unpredictability of systems.

Research Methods:

– Introduction of UnpredictaBench, which includes 448 problems aimed at testing LLMs’ ability to sample outcomes from individual target distributions, using the KS@N evaluation metric to quantify performance.

– Utilization of the Kolmogorov-Smirnov test to measure how well LLMs’ samples approximate target distributions.

Research Conclusions:

– Significant variations exist among models in their distributional capabilities, with scores indicating room for improvement in distributional sampling.

– Current advancements in reasoning and output diversity have yet to provide a complete solution for accurate distributional simulations, highlighting ongoing challenges.

Paper link: https://huggingface.co/papers/2606.06622

43. SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents

Keywords: Persistent AI assistants, memory relations, long-term memory, relational memory, SubtleMemory

Category: Knowledge Representation and Reasoning

Research Objective:

– To evaluate AI agents’ capacity to manage complex relational memory structures using the SubtleMemory benchmark, focusing on long-term memory relations.

Research Methods:

– Introduced SubtleMemory, a benchmark specifically designed to assess fine-grained relational memory discrimination in prolonged AI interactions, consisting of 1,522 evaluation instances and grounded in 1,090 memory-variant sets.

Research Conclusions:

– Existing memory systems exhibit limitations in discriminating fine-grained relational memory, and distinct capability profiles emerge across preservation, retrieval, and reasoning stages.

Paper link: https://huggingface.co/papers/2606.05761

44. When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents

Keywords: ToolMaze, Tool-Integrated Reasoning (TIR), implicit semantic failures, dynamic replanning, agentic fault-tolerance

Category: AI Systems and Tools

Research Objective:

– Introduce ToolMaze, a benchmark designed for dynamic path discovery and error recovery in Tool-Integrated Reasoning (TIR) agents, addressing real-world tool failures.

Research Methods:

– Employ a two-dimensional design incorporating DAG-based topological complexity and a 2×2 taxonomy of tool perturbations (explicit/implicit, transient/permanent) to evaluate performance under various conditions.

Research Conclusions:

– Real-world tool failures, especially implicit semantic ones, significantly degrade TIR performance, with an approximate 37% drop in Perturbation Recovery Rate; dynamic replanning emerges as a crucial bottleneck inadequately addressed by model scaling or prompting.

Paper link: https://huggingface.co/papers/2606.05806

45. Direct 3D-Aware Object Insertion via Decomposed Visual Proxies

Keywords: Object Insertion, Diffusion-based Methods, Pose Control, High-fidelity 2D Image Synthesis

Category: Generative Models

Research Objective:

– The paper introduces DIRECT, a novel framework designed to enable pose-controllable object insertion with high-fidelity 2D image synthesis.

Research Methods:

– The method utilizes decomposed guidance components comprising appearance guidance, geometry guidance, and context guidance to ensure accurate pose manipulation and visual detail integration.

Research Conclusions:

– The proposed method, DIRECT, demonstrates superior performance in geometric controllability and visual quality compared to previous approaches, with the help of an automated data construction pipeline enhancing data diversity and quality.

Paper link: https://huggingface.co/papers/2606.06601

46. MMAE: A Massive Multitask Audio Editing Benchmark

Keywords: Instruction-based Audio Editing, MMAE, Multitask Audio Editing, Audio Modalities

Category: Multi-Modal Learning

Research Objective:

– Introduction of MMAE as a comprehensive benchmark for instruction-based audio editing across various modalities and complexity levels.

Research Methods:

– The benchmark includes a taxonomy covering 7 audio modalities, 6 task complexity levels, and 8 operation types, based on a rubric-based evaluation framework.

Research Conclusions:

– Current models show significant gaps in capabilities, with an Exact Match Rate below 5% and 0% in complex tasks, indicating challenges in execution precision and robustness.

Paper link: https://huggingface.co/papers/2606.07229

47. SoCRATES: Towards Reliable Automated Evaluation of Proactive LLM Mediation across Domains and Socio-cognitive Variations

Keywords: SoCRATES, LLM mediators, socio-cognitive adaptation, consensus gap, multi-domain testbeds

Category: Natural Language Processing

Research Objective:

– To present SoCRATES, a realistic benchmark for evaluating proactive LLM mediators across various socio-cognitive adaptation axes.

Research Methods:

– Construct scenarios from real conflicts using an agentic pipeline in eight domains; probe five socio-cognitive adaptation axes; evaluate using a topic-localized evaluator aligned with human experts.

Research Conclusions:

– Even top-performing Large Language Models (LLMs) resolve only about a third of the consensus gap; performance varies sharply by socio-cognitive axis, indicating need for better social adaptation.

Paper link: https://huggingface.co/papers/2606.05563

Keywords: Code2LoRA, LoRA adapters, hypernetwork framework, GRU hidden state, repository-specific

Category: AI Systems and Tools

Research Objective:

– Introduce Code2LoRA, a hypernetwork framework designed for generating repository-specific LoRA adapters to enhance code language models.

Research Methods:

– Created RepoPeftBench, a benchmark with static and evolution tracks, to evaluate performance on Python repositories against parameter-efficient fine-tuning baselines.

Research Conclusions:

– Code2LoRA-Static and Code2LoRA-Evo achieved significant exact match rates in cross-repo and in-repo scenarios, demonstrating their effectiveness over existing LoRA fine-tuning methods.

Paper link: https://huggingface.co/papers/2606.06492

2. TIDE: Proactive Multi-Problem Discovery via Template-Guided Iteration

Keywords: TIDE, iterative discovery, thought templates

Category: AI Systems and Tools

Research Objective:

– The paper aims to introduce TIDE, a framework for discovering hidden problems in context using templates and iterative methods.

Research Methods:

– TIDE employs two mechanisms: iterative discovery for extending problem coverage and thought templates to anchor predictions within recognizable problem classes.

Research Conclusions:

– TIDE demonstrates significant improvements in task coverage, problem identification, and resolution in personal workspaces and software repositories, surpassing traditional single-shot and parallel multi-agent approaches.

Paper link: https://huggingface.co/papers/2606.04743

3. VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding

Keywords: VideoKR, Knowledge-Intensive, Human-in-the-loop, Video Reasoning, Expert-Domain

Category: Knowledge Representation and Reasoning

Research Objective:

– Introduce VideoKR, a pioneering large-scale training dataset focused on enhancing knowledge- and reasoning-intensive video understanding.

Research Methods:

– Developed a human-in-the-loop, skill-oriented example generation pipeline to cultivate progressively deeper video reasoning capabilities.

Research Conclusions:

– Models post-trained on VideoKR showcase superior performance on knowledge-intensive video reasoning tasks while maintaining competitiveness in general video reasoning, emphasizing data design’s pivotal role.

Paper link: https://huggingface.co/papers/2606.05259

4. RobotValues: Evaluating Household Robots When Human Values Conflict

Keywords: RobotValues, value-conflict scenarios, household robots, default value preferences, VLMs

Category: Robotics and Autonomous Systems

Research Objective:

– Evaluate household robot planners in value-conflict scenarios using the RobotValues benchmark to test their ability to prioritize human values over task completion.

Research Methods:

– Utilization of LLM-assisted scenario generation, stakeholder-grounded value extraction, image generation, and automatic quality control to construct a benchmark with 10,000 scenarios.

Research Conclusions:

– Vision-language models (VLMs) used in robotics display default value preferences and often fail to prioritize specific conflicting values when instructed, making incorrect decisions 80% of the time.

Paper link: https://huggingface.co/papers/2606.03312

5. LoomVideo: Unifying Multimodal Inputs into Video Generation and Editing

Keywords: LoomVideo, video generation, video editing, Multimodal Large Language Model, Scale-and-Add conditioning

Category: Multi-Modal Learning

Research Objective:

– Introduce LoomVideo, a 5B-parameter efficient architecture for unified video generation and editing.

Research Methods:

– Utilizes a Multimodal Large Language Model and Deepstack injection for feature alignment.

– Implements Scale-and-Add conditioning to significantly reduce computational cost.

Research Conclusions:

– LoomVideo achieves state-of-the-art performance with superior efficiency and speed, particularly excelling in e-commerce and fashion generation scenarios.

– The model provides a 5.41x acceleration in inference speed over similarly capable models.

Paper link: https://huggingface.co/papers/2606.06042

6. Rethinking Continual Experience Internalization for Self-Evolving LLM Agents

Keywords: Experience internalization, Continual learning, Large language models, Capability collapse, Internalization regime

Category: Natural Language Processing

Research Objective:

– To investigate the mechanisms of experience internalization to enable continual learning in large language models.

Research Methods:

– Systematic examination of three dimensions: experience granularity, experience injection pattern, and internalization regime.

Research Conclusions:

– Principle-level experience is more durable than instance-level experience for transferability.

– Step-wise injection is superior to global injection for aligning experiences with decision states.

– Off-policy context-distillation provides a more stable training signal than on-policy context-distillation for improving stability in experience internalization.

Paper link: https://huggingface.co/papers/2606.04703

7. The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset

Keywords: KITScenes Multimodal dataset, high-fidelity sensors, HD maps, embodied AI, geographic diversity

Category: Robotics and Autonomous Systems

Research Objective:

– The research aims to provide a comprehensive European driving dataset with high-fidelity sensors, including rich 3D maps and diverse urban environments, to advance embodied AI research.

Research Methods:

– Utilizes a fully synchronized sensor suite comprising high-resolution cameras, long-range lidar, 4D imaging radar, and GNSS/INS for precise localization. The dataset features complete HD maps validated through autonomous driving trials.

Research Conclusions:

– The KITScenes Multimodal dataset enriches existing datasets by offering unprecedented map completeness and geographic diversity, setting benchmarks in HD map construction, long-range depth estimation, novel view synthesis, and end-to-end driving.

Paper link: https://huggingface.co/papers/2606.02956

8. LLMs Can Leak Training Data But Do They Want To? A Propensity-Aware Evaluation of Memorization in LLMs

Keywords: PropMe, memorization evaluation, SimpleTrace, propensity-aware framework, prefix-based capability attacks

Category: Natural Language Processing

Research Objective:

– The study aims to evaluate language model memorization by differentiating between forced reproduction capabilities and natural propensity using the PropMe framework.

Research Methods:

– Introduces PropMe, a propensity-aware framework, and SimpleTrace, a lightweight tracing tool. Utilizes propensity-transformed metrics across open models and datasets, focusing on prefix-based capability attacks versus non-adversarial evaluations.

Research Conclusions:

– The study concludes that large language models can reveal training data when directly elicited, but do so less frequently under non-adversarial circumstances. It also highlights the importance of assessing both worst-case extractability and ordinary leakage propensity for a comprehensive view of memorization capabilities.

Paper link: https://huggingface.co/papers/2606.06286

9. MAOAM: Unified Object and Material Selection with Vision-Language Models

Keywords: unified vision-language model, object selection, material selection, interactive image editing, MAOAM

Category: Computer Vision

Research Objective:

– Introduce MAOAM, a unified selection framework to enhance object and material selection via text and click interactions, supporting diverse editing workflows with improved robustness.

Research Methods:

– Utilizes a vision-language model (VLM) with a segmentation head to create pixel-accurate masks from user-defined prompts, aiming at both object and material selection.

Research Conclusions:

– Demonstrated accurate and coherent selection capabilities in diverse scenarios, improving image editing workflows by integrating text and click-based interactions effectively.

Paper link: https://huggingface.co/papers/2606.04880

10. Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents

Keywords: Memory-augmented language models, Belief Entropy, Metacognitive Memory Policy Optimization, long-horizon tasks, epistemic uncertainty

Category: Reinforcement Learning

Research Objective:

– The research aims to improve performance in memory-augmented language models tackling long-horizon tasks by focusing on memory quality instead of solely outcome success.

Research Methods:

– The study introduces Belief Entropy as a self-supervised proxy to assess uncertainty about latent task states.

– Metacognitive Memory Policy Optimization (MMPO) is proposed to provide memory-specific supervision and penalize summaries increasing epistemic uncertainty, departing from traditional outcome-based signals.

Research Conclusions:

– Experiments demonstrate that MMPO consistently outperforms existing methods in diverse long-horizon tasks, maintaining high performance even in large contexts.

Paper link: https://huggingface.co/papers/2605.30159

11. World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis

Keywords: World-language-action models, Autoregressive Transformer, Long-horizon task execution, Cross-embodiment learning, State-of-the-art

Category: Robotics and Autonomous Systems

Research Objective:

– The objective is to develop world-language-action (WLA) models, integrating textual instructions, images, and robot state predictions to efficiently execute long-horizon tasks and enhance cross-embodiment learning.

Research Methods:

– An autoregressive (AR) Transformer backbone is used to predict future states by combining semantic-level textual intentions with fine-grained physical dynamics, enabled by a World Expert for supervising physical dynamics and meta-queries for world prediction impacting action generation.

Research Conclusions:

– The WLA-0 prototype, with 2B active parameters, demonstrates state-of-the-art capabilities in multi-task and long-horizon learning in simulated and real-world environments, achieving notable success rates on RoboTwin2.0 Clean and RMBench tasks, and shows potential for learning novel tasks from cross-embodiment robot videos without action annotations.

Paper link: https://huggingface.co/papers/2606.05979

12. AffordanceVLA: A Vision-Language-Action Model Empowering Action Generation through Affordance-Aware Understanding

Keywords: AffordanceVLA, Vision-Language-Action, perception-action mapping, affordance forecasting, robotic manipulation

Category: Robotics and Autonomous Systems

Research Objective:

– AffordanceVLA aims to establish a precise perception-action mapping by integrating structured affordance forecasting with Vision-Language models to enhance robotic manipulation.

Research Methods:

– The framework utilizes a Mixture-of-Transformer architecture with specialized experts, combined with a three-stage training strategy and automated data augmentation to tackle data scarcity issues in robotic datasets.

Research Conclusions:

– AffordanceVLA demonstrates strong performance in diverse manipulation scenarios through its spatially grounded and semantically conditioned affordance cues, bridging the gap between vision, language, and action.

Paper link: https://huggingface.co/papers/2606.06155

13. MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery

Keywords: MLEvolve, LLM-based, multi-agent framework, machine learning algorithm discovery, self-evolving

Category: Machine Learning

Research Objective:

– Introduce MLEvolve, an LLM-based self-evolving multi-agent framework designed for machine learning algorithm discovery to overcome existing limitations in long-horizon tasks.

Research Methods:

– Utilizes Progressive MCGS to enhance search mechanisms with graph-based reference edges and an entropy-inspired progressive schedule.

– Implements Retrospective Memory for dynamic knowledge retrieval and reuse to facilitate agent evolution.

Research Conclusions:

– MLEvolve exhibits state-of-the-art performance on MLE-Bench across various metrics and outperforms methods like AlphaEvolve in mathematical algorithm optimization tasks, showcasing strong cross-domain generalization.

Paper link: https://huggingface.co/papers/2606.06473

14. SePO: Self-Evolving Prompt Agent for System Prompt Optimization

Keywords: Self-Evolving Prompt Optimization, evolutionary search, task agents, prompt optimization, fine-tuning

Category: Natural Language Processing

Research Objective:

– The objective is to enhance AI agent performance by jointly optimizing both task and prompt agent system prompts using a novel method called Self-Evolving Prompt Optimization (SePO).

Research Methods:

– The study employs an evolutionary search strategy with a self-referential design, using a single prompt agent to improve both task agents’ system prompts and its own system prompt.

Research Conclusions:

– Self-Evolving Prompt Optimization significantly outperforms existing methods across five different benchmarks, demonstrating an average accuracy improvement of 4.49 points over the Manual-CoT technique. The optimization skill gained from pre-training efficiently generalizes across various tasks beyond the pre-training mixture.

Paper link: https://huggingface.co/papers/2606.04465

15. SEAOTTER: Sensor Embedded Autoencoding with One-Time Transcode for Efficient Reconstruction

Keywords: Cloud Robotics, Learned Latent, JPEG Compatibility, Asymmetric Autoencoders, SEAOTTER

Category: Robotics and Autonomous Systems

Research Objective:

– The study presents a compression framework for cloud robotics that merges learned latent representations with standard JPEG compatibility to enhance encoding and decoding speed while maintaining high perceptual quality.

Research Methods:

– Introduction of the SEAOTTER framework, which pairs a Sensor Embedded Autoencoder with a One-Time Transcode for Efficient Reconstruction, utilizing a learnable JPEG color and quantization transform to improve accuracy in various perception tasks.

Research Conclusions:

– The proposed framework achieves significant improvements over AVIF, with 7 times faster encoding, 3.5 times faster decoding, and +8% ImageNet top-1 accuracy, all while preserving compatibility with existing JPEG infrastructure.

Paper link: https://huggingface.co/papers/2606.03940

16. Regret Minimization with Adaptive Opponents in Repeated Games

Keywords: Repeated Policy Regret, adaptive opponents, game-theoretic, non-convex optimization

Category: Reinforcement Learning

Research Objective:

– The study focuses on minimizing regret in repeated games with adaptive opponents by introducing a new metric: Repeated Policy Regret.

Research Methods:

– Identification of necessary conditions to achieve sublinear RP-Regret.

– Proposal of three algorithms designed for minimizing non-convex RP-Regret.

Research Conclusions:

– Minimizing RP-Regret leads to finding better equilibria and more cooperative solutions in repeated games like Stag-Hunt.

Paper link: https://huggingface.co/papers/2606.06486

17. Latent Reasoning with Normalizing Flows

Keywords: Latent reasoning, normalizing flows, chain-of-thought, probabilistic sampling, KV-cache decoding

Category: Natural Language Processing

Research Objective:

– The study aims to enhance large language models’ reasoning capabilities by integrating latent reasoning through normalizing flows without losing the benefits of autoregressive generation.

Research Methods:

– The framework, called NF-CoT, uses normalizing flows to conduct intermediate computations in continuous states, maintaining compatibility with left-to-right generation and probabilistic sampling.

Research Conclusions:

– NF-CoT has been shown to improve code-generation pass rates and reduce reasoning costs compared to explicit chain-of-thought methods and previous latent reasoning frameworks.

Paper link: https://huggingface.co/papers/2606.06447

18. MechVQA: Benchmarking and Enhancing Multimodal LLMs on Comprehensive Mechanical Drawing Understanding

Keywords: Mechanical engineering drawing, Multimodal Large Language Models, MechVQA, High-density visual question answering, Domain knowledge

Category: Computer Vision

Research Objective:

– To enhance understanding of mechanical engineering drawings using a specialized dataset and domain-specific model.

Research Methods:

– Introduction of MechVQA dataset containing 3.3k images and 21k question-answer pairs.

– Development of MechVL model through a multi-stage training paradigm.

Research Conclusions:

– MechVL model outperforms existing baselines by 7.57 percentage points, providing a reusable foundation for deploying MLLMs in mechanical design and inspection.

Paper link: https://huggingface.co/papers/2605.30794

19. LLM Anonymization Against Agentic Re-Identification

Keywords: LLM-powered, anonymization, re-identification, contextual utility, adaptive privacy scope

Category: AI Ethics and Fairness

Research Objective:

– Develop AURA, a framework to balance privacy protection and utility retention in text anonymization using LLM-powered methods.

Research Methods:

– Utilization of adaptive privacy scopes and mask-reconstruct methods evaluated against re-identification attacks from web-search agents and utility based on real-user interviews.

Research Conclusions:

– AURA enhances the privacy-utility frontier by improving resistance to agentic re-identification while preserving contextual utility through adaptive privacy and mask-reconstruct techniques.

Paper link: https://huggingface.co/papers/2605.30848

20. Video2LoRA: Parametric Video Internalization for Vision-Language Models

Keywords: Video2LoRA, Low-Rank Adaptation, video processing, vision-language models, inference cost

Category: Multi-Modal Learning

Research Objective:

– The main objective is to enhance video processing efficiency in vision-language models by predicting Low-Rank Adaptation weights, thereby reducing computational costs while retaining video-faithful outputs.

Research Methods:

– The introduction of Video2LoRA method involves a perceiver hypernetwork that reads intermediate representations from a frozen Vision-Language Model (VLM) to generate Low-Rank Adaptation adapters in a single forward pass.

Research Conclusions:

– Video2LoRA achieves equivalent performance to direct video-in-context inference, reducing answer-time visual-token load and query TTFT significantly while being stable for extensive frame and pixel ranges.

Paper link: https://huggingface.co/papers/2606.04351

21. Discrete-WAM: Unified Discrete Vision-Action Token Editing for World-Policy Learning

Keywords: Discrete-WAM, Autonomous Driving, Causal Reasoning, Discrete Tokens, Discrete Diffusion Framework

Category: Robotics and Autonomous Systems

Research Objective:

– Introduce Discrete-WAM, a unified discrete latent vision-action world policy for autonomous driving, enabling compositional causal reasoning and counterfactual reasoning.

Research Methods:

– Utilize aligned discrete tokens and a shared discrete diffusion framework for compositional generalization across diverse driving scenarios.

Research Conclusions:

– Discrete-WAM achieves competitive performance on large-scale autonomous-driving benchmarks, supporting controllable generation and offering a principled path to more reliable decision-making.

Paper link: https://huggingface.co/papers/2606.05645

22. Trust Region Q Adjoint Matching

Keywords: Off-policy reinforcement learning, Trust Region Q-Adjoint Matching, pretrained flow policies, projected dual descent, model collapse

Category: Reinforcement Learning

Research Objective:

– The objective is to address the instability in off-policy reinforcement learning by introducing Trust Region Q-Adjoint Matching (TRQAM) to ensure stable fine-tuning of pretrained flow policies.

Research Methods:

– TRQAM uses projected dual descent to adaptively control the path-space KL divergence, and optimizes the trust-region parameter (λ) in stochastic optimal control dynamics to stabilize the learning process.

Research Conclusions:

– Experiments on 50 OGBench tasks demonstrate that TRQAM consistently outperforms existing approaches in offline RL and offline-to-online RL, achieving a 68% success rate compared to a strong baseline of 46%.

Paper link: https://huggingface.co/papers/2605.27079

23. Combinatorial Synthesis: Scaling Code RLVR via Atomic Decomposition and Recombination

Keywords: Atomic Decomposition, Recombination, Verifiable Code Tasks, Reinforcement Learning, Large Language Models

Category: Reinforcement Learning

Research Objective:

– To propose the Atomic Decomposition and Recombination (ADR) framework for generating novel and challenging verifiable code tasks to enhance the scalability of Reinforcement Learning with Verifiable Rewards (RLVR) in Large Language Models.

Research Methods:

– The ADR framework decomposes tasks into atomic elements and performs controlled recombination to produce new tasks, surpassing the limitations of previous heuristic approaches for data synthesis.

Research Conclusions:

– ADR demonstrates excellent originality, difficulty, diversity, and test quality over existing methods and improves coding performance across RLVR in various domains such as algorithmic programming, tool usage, and data science.

Paper link: https://huggingface.co/papers/2605.31058

24. Quality-Guided Semi-Supervised Learning for Medical Image Segmentation

Keywords: Semi-supervised learning, pseudolabels, segmentation quality, quality predictor, medical image segmentation

Category: AI in Healthcare

Research Objective:

– To develop a quality-guided semi-supervised learning framework that enhances medical image segmentation by improving pseudolabel reliability and segmentation performance.

Research Methods:

– Introduction of a dedicated quality predictor trained on variable-quality masks from synthetic corruptions and partially trained segmentation models.

– Integration of the quality predictor into SSL through quality-aware regularization loss and quality-based pseudolabel sample reweighting.

Research Conclusions:

– Proposed method consistently improves over existing semi-supervised learning methods in medical image segmentation, validated through extensive experiments across five datasets and multiple architectures.

Paper link: https://huggingface.co/papers/2606.01753

25. ForeSci: Evaluating LLM Agents for Forward-Looking AI Research Judgment

Keywords: LLM agents, forward-looking research, ForeSci, decision-making systems, AI domains

Category: AI Systems and Tools

Research Objective:

– Introduction of ForeSci, a benchmark aimed at evaluating the ability of LLM agents to make forward-looking research decisions based on historical evidence in AI domains.

Research Methods:

– Creation of tasks derived from pre-cutoff taxonomy branches and use of specific answer-generation backbones preceding task cutoffs to enhance accuracy and traceability.

– Evaluation of native LLMs, Hybrid RAG, and research-agent adaptations across various backbones to test evidence organization and decision-making capabilities.

Research Conclusions:

– Evidence organization improves traceability and factual support.

– The effectiveness of evidence organization depends significantly on the decision family, with a noted challenge of evidence-decision decoupling affecting research judgements.

Paper link: https://huggingface.co/papers/2606.00644

26.

Paper link: 

27. SABER: Benchmarking Operational Safety of LLM Coding Agents in Stateful Project Workspaces

Keywords: Large Language Models, Coding Agents, Environment-Aware Operational Safety, Safety Profiles, Harmful Safety-Violation Rate

Category: AI Systems and Tools

Research Objective:

– Introduce SABER, a benchmark for evaluating the safety of large language models as coding agents in realistic project environments.

Research Methods:

– SABER places models in realistic agent-style projects to assess safety based on the action sequence’s final environment state rather than binary prompt refusals.

Research Conclusions:

– Even the best large language models have over a 54% harmful safety-violation rate, revealing current alignment deficiencies and distinct safety profiles across models.

Paper link: https://huggingface.co/papers/2606.01317

28. BRepCLIP: Contrastive Multimodal Pretraining on BRep Primitives for CAD Understanding

Keywords: BRepCLIP, Multimodal Representation Learning, Contrastive Pretraining, CAD Models, Boundary Representations

Category: Multi-Modal Learning

Research Objective:

– Introduce BRepCLIP, a framework for aligning boundary representation (BRep) geometry of CAD models with language and image embeddings using contrastive pretraining.

Research Methods:

– Model CAD objects as sequences of face and edge tokens with discrete vocabularies for surface and curve geometry, augmented with spatial and semantic descriptors. Use a transformer encoder to aggregate these into a global BRep embedding aligned with CLIP’s text and image encoders.

Research Conclusions:

– BRepCLIP achieves superior retrieval and classification performance over point-based methods, showing significant improvements in retrieval and classification scores across various datasets and proving effective as a CAD-aware similarity metric.

Paper link: https://huggingface.co/papers/2606.05515

29. Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing

Keywords: RE-Edit, image editing systems, reasoning dimensions, logical consistency, Diffusion-based image editing

Category: Knowledge Representation and Reasoning

Research Objective:

– To introduce RE-Edit, a benchmark for reasoning-aware image editing that evaluates systems across five reasoning dimensions: physical, environmental, cultural, causal, and referential.

Research Methods:

– Development of a benchmark comprising 1,000 curated samples, each designed to test the logical consistency of image editing systems beyond visual plausibility.

– Evaluation of ten open-source and two commercial image editing models using dimension-aligned criteria for fine-grained analysis.

Research Conclusions:

– Finding that even advanced image editing systems struggle with multi-dimensional reasoning despite high-quality visuals.

– Introduction of a lightweight reasoning-guided post-edit baseline, demonstrating the potential of explicit reasoning to improve model performance in a model-agnostic manner.

Paper link: https://huggingface.co/papers/2606.05172

30. Multimodal Music Recommendation System using LLMs

Keywords: multimodal framework, session-based music recommendation, LLM-based sequential reasoning, audio and lyric embeddings, cross-modal integration

Category: Multi-Modal Learning

Research Objective:

– The paper aims to enhance music recommendation accuracy by integrating audio, lyric, and semantic signals using a multimodal framework that employs LLM-based sequential reasoning.

Research Methods:

– The study adopts a multimodal framework to enrich the LastFM-1K dataset with audio and lyric embeddings, LLM-generated semantic metadata, and listening completion ratios.

– The research leverages E4SRec, extending it with various item ID encoder backbones and LLM backbones, including SASRec, BERT4Rec, GRU4Rec, LLaMa-2-13B, Qwen2.5-7B-Instruct, and LLaMa-3-70B.

Research Conclusions:

– The integration of content-based features significantly improves recommendation accuracy, demonstrating up to 95% improvement in Recall and 79% in NDCG.

– The study highlights challenges in cross-modal integration, noting that naive multimodal fusion does not always yield additive improvements.

– A large-scale multimodal benchmark for music recommendation is released.

Paper link: https://huggingface.co/papers/2606.00125

31. The Shape of Addition: Geometric Structures of Arithmetic in Large Language Models

Keywords: Large Language Models, Arithmetic Fragility, Geometric Structures, Noisy Quantization Model, Geometric Slippages

Category: Natural Language Processing

Research Objective:

– Analyze the geometric structures causing arithmetic fragility in Large Language Models (LLMs) and propose a new model to address these issues.

Research Methods:

– Introduce and utilize the Noisy Quantization Model to explain arithmetic errors in LLMs and employ geometric frameworks to detect and correct quantization failures during inference.

Research Conclusions:

– Identified and explained the Iso-Raw-Sum Trajectory (IRST) as a key structure in arithmetic fragility, validating the insights through geometric consistency checks that successfully detect and correct arithmetic errors.

Paper link: https://huggingface.co/papers/2606.03645

32. Absorbing Complexity: An Interaction-Native Knowledge Harness for Financial LLM Agents

Keywords: Financial AI agents, InKH, knowledge management, temporal memory, AI Native

Category: AI in Finance

Research Objective:

– The study introduces the Interaction-native Knowledge Harness (InKH) architecture to embed complexity within financial AI agents, reducing the need for users to manage this complexity.

Research Methods:

– Utilized a controlled synthetic benchmark with 24 random seeds, 4 rounds, and 80 episodes per round, comparing InKH against 6 baselines in 46,080 evaluations.

Research Conclusions:

– InKH significantly reduces latency, token cost, and stale-knowledge usage, while improving task quality and traceability, demonstrating that system-absorbed complexity enhances financial AI agent efficacy.

Paper link: https://huggingface.co/papers/2606.01886

33. Benchmark Everything Everywhere All at Once

Keywords: Automated benchmark creation, LLMs, scalability, domain-specific reasoning, Benchmark Agent

Category: AI Systems and Tools

Research Objective:

– Develop an automated system, Benchmark Agent, for creating diverse evaluation datasets to facilitate continuous model assessment.

Research Methods:

– The framework encompasses user query analysis, subtask design, data annotation, and quality control. Evaluations include human assessments and consistency checks.

Research Conclusions:

– Benchmark Agent produces high-quality benchmarks with minimal human intervention and highlights weaknesses in current models, particularly in domain-specific reasoning tasks.

Paper link: https://huggingface.co/papers/2606.06462

34. AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents

Keywords: Intent Inference, Implicit Needs, Gap Scoring, Tool Usage, Probe Consumption

Category: Knowledge Representation and Reasoning

Research Objective:

– Enhance query answering by estimating implicit needs and optimizing tool usage through an intent inference step.

Research Methods:

– Incorporate an inference step that produces an IntentFrame to estimate implicit needs.

– Utilize gap scoring to control per-query probe budget and tool selection.

– Benchmark with a 100-query four-scene implicit-intent dataset.

Research Conclusions:

– AURA achieves improved implicit-need coverage compared to standard approaches, with significant gains across multiple scenes.

– The approach reduces probe consumption and maintains privacy compliance on factual lookups.

– The improvement is attributed to gap calibration rather than answer memorization.

Paper link: https://huggingface.co/papers/2606.05557

35. EvoDS: Self-Evolving Autonomous Data Science Agent with Skill Learning and Context Management

Keywords: EvoDS, Autonomous Skill Acquisition, Adaptive Context Compression, Reinforcement Learning

Category: AI Systems and Tools

Research Objective:

– The study aims to enhance automated data science capabilities by introducing a self-evolving autonomous data science agent, EvoDS, which utilizes skill acquisition and adaptive context management.

Research Methods:

– Implemented two strategies: Autonomous Skill Acquisition (ASA) and Adaptive Context Compression (ACC), within a two-stage multi-agent training scheme, leveraging reinforcement learning principles to improve context management and skill synthesis.

Research Conclusions:

– EvoDS demonstrates a significant performance improvement, outperforming state-of-the-art data science agents by an average of 28.9% across various benchmarks and eliminating out-of-token failures.

Paper link: https://huggingface.co/papers/2606.03841

36. Revising Context, Shifting Simulated Stance: Auditing LLM-Based Stance Simulation in Online Discussions

Keywords: LLM-based stance simulation, context sensitivity, counterfactual context revision, multimodal approaches

Category: Natural Language Processing

Research Objective:

– Investigate how Large Language Models (LLMs) simulate social media user stances, focusing on context sensitivity in counterfactual scenarios.

Research Methods:

– Applied controlled revision strategies to both text-only and multimodal conversational contexts to simulate stance changes.

– Evaluated the effectiveness of these strategies using metrics such as average directional stance shift and stance transition rate.

Research Conclusions:

– Both text-only and multimodal approaches exhibit robust stance transitions, highlighting the importance and complexity of context sensitivity in LLM-based stance simulations.

– The study provides a framework for understanding these simulations, bringing attention to both the potential and risks of using LLMs for simulating opinion dynamics online.

Paper link: https://huggingface.co/papers/2606.06443

37. AdaCodec: A Predictive Visual Code for Video MLLMs

Keywords: AdaCodec, Video MLLMs, Visual Tokens, Inter-frame Changes, Predictive Visual Code

Category: Multi-Modal Learning

Research Objective:

– The paper introduces AdaCodec, a system designed to reduce redundancy in video encoding by selectively transmitting full visual tokens only when scene prediction fails.

Research Methods:

– AdaCodec operates by first sending a full reference frame when the scene cannot be predicted well from prior context, otherwise it encodes compact inter-frame changes.

Research Conclusions:

– AdaCodec significantly outperforms a baseline model by improving efficiency in visual token usage and reducing time-to-first-token from 9.26s to 1.62s across multiple benchmarks, even at reduced visual-token budgets.

Paper link: https://huggingface.co/papers/2606.02569

38. Flash-WAM: Modality-Aware Distillation for World Action Models

Keywords: Flash-WAM, Modality-aware, World-action models, Real-time inference, Consistency function

Category: Robotics and Autonomous Systems

Research Objective:

– The study introduces Flash-WAM, a modality-aware step-distillation framework to enhance World-action models (WAMs) for real-time inference by addressing inconsistencies across noise regimes in video and action streams.

Research Methods:

– The paper employs a step-distillation process inspired by consistency distillation, utilizing different parametrization techniques (linear-gradient-scaling and variance-preserving) to optimize video and action streams’ noise conditions.

Research Conclusions:

– Flash-WAM significantly reduces latency and maintains high task success rates in simulation benchmarks, enabling real-time inference on platforms like RoboTwin 2.0 and improving performance on real-world tasks compared to naive consistency distillation.

Paper link: https://huggingface.co/papers/2606.05254

39. Towards Truly Multilingual ASR: Generalizing Code-Switching ASR to Unseen Language Pairs

Keywords: Automatic Speech Recognition, code-switching ASR, synthetic CS speech generation, model merging, domain generalization

Category: Natural Language Processing

Research Objective:

– To investigate the generalization capabilities of code-switching ASR models across unseen language pairs using model merging and domain generalization methods.

Research Methods:

– The study employed model merging and domain generalization techniques to test the transferability of bilingual code-switching capabilities to new language pairs.

Research Conclusions:

– Merged bilingual CS-ASR models showed modest generalization to unseen language pairs, indicating limited transferability of bilingual CS capabilities.

Paper link: https://huggingface.co/papers/2606.05846

40. Learning Geometric Representations from Videos for Spatial Intelligent Multimodal Large Language Models

Keywords: GeoVR, 3D awareness, geometric knowledge distillation, semantic latent space, spatial intelligence

Category: Multi-Modal Learning

Research Objective:

– GeoVR aims to enhance multimodal large language models (MLLMs) by introducing 3D awareness through a novel framework that restructures their semantic latent space using geometric knowledge from 3D foundation models.

Research Methods:

– The framework utilizes a multi-objective learning strategy employing four geometric targets, including camera pose estimation, dense depth map regression, metric scale factor prediction, and multi-scale 3D feature distillation, to develop strong 3D understanding.

Research Conclusions:

– GeoVR, through extensive experiments on spatial reasoning benchmarks, demonstrates state-of-the-art performance and establishes a new paradigm in endowing foundation models with spatial intelligence.

Paper link: https://huggingface.co/papers/2606.05833

41. Towards One-to-Many Temporal Grounding

Keywords: One-to-Many Temporal Grounding, Temporal Grounding, Count Accuracy, Effective Temporal F1, Chain-of-Thought reasoning

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the challenge of One-to-Many Temporal Grounding by introducing a comprehensive benchmark, novel reward functions, and improved policy optimization.

Research Methods:

– Establishes the first comprehensive benchmark for One-to-Many Temporal Grounding (OMTG) with new evaluation metrics like Count Accuracy and Effective Temporal F1.

– Develops a high-quality OMTG dataset with 56k samples and novel temporal and caption reward functions utilizing Chain-of-Thought reasoning for policy optimization.

Research Conclusions:

– The proposed model achieves a new state-of-the-art Effective Temporal F1 of 43.65% on the OMTG benchmark, outperforming previous models by significant margins.

Paper link: https://huggingface.co/papers/2606.06294

42. Imagine Before You Predict: Interleaved Latent Visual Reasoning for Video Event Prediction

Keywords: Future-L1, Video event prediction, Latent visual reasoning, Autoregressive decoding, State-of-the-art

Category: Multi-Modal Learning

Research Objective:

– Future-L1 is designed to improve video event prediction by maintaining visual semantics in latent space during autoregressive decoding, aiming for enhanced prediction accuracy.

Research Methods:

– The Future-L1 framework alternates between language tokens and continuous latent visual spans.

– Constructs Future-L1-50K dataset and employs LA-DAPO, a latent-aware RL objective with outcome-contrastive and temporal-diversity rewards, to optimize latent trajectories.

Research Conclusions:

– Future-L1 achieves state-of-the-art results, significantly improving benchmark scores on FutureBench and TwiFF-Bench, demonstrating the benefits of preserving visual semantics in latent space rather than converting all reasoning into text.

Paper link: https://huggingface.co/papers/2606.05769

43. The Shadow Price of Reasoning: Economic Perspective on Optimal Budget Allocation for LLMs

Keywords: Inference-time scaling, Large Language Models, constrained optimization, economic principles, global shadow price

Category: Natural Language Processing

Research Objective:

– The study aims to enhance performance in resource-constrained environments by improving inference-time scaling for Large Language Models through a novel economic principle-based optimization strategy.

Research Methods:

– The paper formulates inference budget allocation as a global constrained optimization problem, employing economic principles and modeling per-query reasoning utility with a shifted-surge function to derive an optimal allocation policy.

Research Conclusions:

– The proposed Constrained Latent-utility Equilibrium Allocation for Reasoning (CLEAR) strategy reallocates resources to solvable queries, significantly improving the Pareto frontier of token cost versus mean accuracy. In resource-scarce regimes, CLEAR enhances global accuracy up to 3 times compared to uniform allocation.

Paper link: https://huggingface.co/papers/2606.03092

44. OPRD: On-Policy Representation Distillation

Keywords: On-Policy Representation Distillation, hidden-state space, sampling variance, training efficiency

Category: Machine Learning

Research Objective:

– Improve the traditional on-policy distillation by aligning student and teacher representations in the hidden-state space to reduce variance and improve training efficiency.

Research Methods:

– Introduce On-Policy Representation Distillation (OPRD) that aligns student and teacher representations in hidden-state space, bypassing the LM head and providing richer per-layer structural information.

Research Conclusions:

– OPRD closes the student-teacher gap effectively, trains 1.44x faster, and uses 54% less memory than existing top-k OPD methods.

Paper link: https://huggingface.co/papers/2606.06021

45. Unsupervised Skill Discovery for Agentic Data Analysis

Keywords: DataCOPE, data-analytic agent, skill discovery, Adaptive Checklist Verifier, Answer Agreement Verifier

Category: Machine Learning

Research Objective:

– The main objective of the research is to develop DataCOPE, an unsupervised framework that discovers reusable data-analysis skills to improve the performance in report-style and reasoning-style tasks.

Research Methods:

– DataCOPE utilizes verifier-guided exploration, deriving verifier signals from exploration trajectories to evaluate quality and agreement. It involves a Data-Analytic Agent for trajectory generation, an Unsupervised Verifier for signal extraction, and a Skill Manager for contrastive skill distillation. The framework is instantiated with an Adaptive Checklist Verifier for report-style analysis and an Answer Agreement Verifier for reasoning-style analysis.

Research Conclusions:

– DataCOPE consistently enhances performance over baseline models, achieving an average improvement in mean score by 9.71% for report-style tasks and 32.30% for reasoning-style tasks across various model settings.

Paper link: https://huggingface.co/papers/2606.06416

46. Dream.exe: Can Video Generation Models Dream Executable Robot Manipulation?

Keywords: Video generation models, robotic manipulation, physics simulator, trajectory fidelity, execution success

Category: Robotics and Autonomous Systems

Research Objective:

– To evaluate video generation models on their ability to reflect physical reality through robotic manipulation tasks, and assess if visual quality predicts executable motion accuracy.

Research Methods:

– Introduce Dream.exe, an evaluation framework with a video-to-execution pipeline to assess the execution ability of videos generated for robotic tasks in a physics simulator.

Research Conclusions:

– Several models achieved measurable execution success, indicating that generative priors learned from large datasets may contain meaningful physical knowledge, although visual quality does not reliably predict executability.

Paper link: https://huggingface.co/papers/2606.04811

47. Complexity-Balanced Diffusion Splitting

Keywords: Complexity-Balanced Splitting, generative capacity, temporal capacity allocation, diffusion timeline, synthesis quality

Category: Generative Models

Research Objective:

– The objective is to improve synthesis quality in generative models without increasing inference costs by introducing Complexity-Balanced Splitting (CBS), which allocates generative capacity across specialized sub-networks based on local complexity measures.

Research Methods:

– CBS divides the diffusion timeline into segments with equal approximation burden using two monitor functions: a spatial measure based on Dirichlet energy and a geometric measure based on sampling trajectories’ acceleration. A lightweight auxiliary model estimates local complexity profiles to optimize temporal partitioning.

Research Conclusions:

– CBS consistently enhances synthesis quality across various architectures and datasets, achieving a ~35% improvement in FID on SiT-XL with CFG compared to naive temporal partitioning, without additional inference cost.

Paper link: https://huggingface.co/papers/2606.06477

48. Personal AI Agent for Camera Roll VQA

Keywords: Conversational AI, Visual Question Answering, Hierarchical Memory, Personal Camera Roll

Category: Human-AI Interaction

Research Objective:

– To develop a Conversational AI agent for answering visual questions using a personal camera roll with hierarchical memory and specialized tools.

Research Methods:

– Creation and manual annotation of a dataset named camroll containing 50 users, 31,476 images, and 2,500 QA pairs to support real-world usage.

Research Conclusions:

– The camroll-agent demonstrates superior performance in long-context understanding compared to existing baselines, indicating the need for different approaches in AI agents for personalized visual memory versus textual memory.

Paper link: https://huggingface.co/papers/2606.05275

49. Reinforcement Learning Elicits Contextual Learning of Unseen Language Translation

Keywords: Reinforcement learning, Large language models, Zero-shot transfer, Meta-skill, Linguistic context

Category: Natural Language Processing

Research Objective:

– The research aims to enhance the ability of large language models to translate unseen languages by utilizing in-context linguistic knowledge instead of memorizing specific languages.

Research Methods:

– The study employs a reinforcement learning approach using a surface-level translation metric, chrF, as the reward to train models on unseen language translation.

Research Conclusions:

– The reinforcement learning-trained models successfully leverage linguistic context, outperforming in-context learning and supervised fine-tuning, and potentially apply RL to tasks beyond traditional reasoning such as language learning from context.

Paper link: https://huggingface.co/papers/2606.06428

50. AdaPlanBench: Evaluating Adaptive Planning in Large Language Model Agents under World and User Constraints

Keywords: AdaPlanBench, adaptive planning, Large Language Model, dual constraints, interactive benchmark

Category: Natural Language Processing

Research Objective:

– The paper introduces AdaPlanBench, a benchmark for evaluating Large Language Models (LLMs) on their capacity to adaptively plan and re-plan under progressively revealed world and user constraints.

Research Methods:

– Utilizes a set of 307 household tasks and a scalable constraint construction pipeline to augment each task with dual constraints.

– LLM agents interact through a multi-turn protocol where constraints are revealed only when violated, requiring iterative plan revisions.

Research Conclusions:

– Experiments with ten leading LLMs indicate that adaptive planning under dual constraints is challenging, with a maximum accuracy of 67.75%.

– Performance is particularly degraded by user constraints and is affected by weaker physical grounding and reduced effectiveness.

Paper link: https://huggingface.co/papers/2606.05622

51. ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?

Keywords: Role-playing language agents, ArcANE, Character Arc, Narrative Evaluation, psychological trajectory

Category: Natural Language Processing

Research Objective:

– The study aims to develop and evaluate benchmarks for Role-playing language agents that focus on dynamic character development and psychological trajectory alignment rather than static factual recall.

Research Methods:

– Introduced ArcANE, a benchmark evaluating character development across 17 novels and 80 principal characters, using an approach that segments narratives into phases along a psychological axis to test model responses.

Research Conclusions:

– ArcANE-conditioned models demonstrate superior performance compared to other context strategies, especially in scenarios unexplored in source texts, with fine-tuned models (ArcANE-8B/32B) further enhancing performance outside source text contexts.

Paper link: https://huggingface.co/papers/2606.05553

Keywords: Audio Interaction Model, Large Audio Language Models, SoundFlow, Streaming Audio Models, Real-time ASR

Category: Human-AI Interaction

Research Objective:

– Develop a unified streaming audio model that combines offline task execution with real-time audio instruction following in an end-to-end framework.

Research Methods:

– The development of SoundFlow, facilitating an always-on perceive-decide-respond loop through streaming-native data construction and comprehension-aware training.

– Creation of StreamAudio-2M, a comprehensive streaming corpus, and Proactive-Sound-Bench to evaluate proactive audio intervention.

Research Conclusions:

– Audio-Interaction preserves competitive performance on mainstream audio tasks while enabling new real-time capabilities such as ASR, streaming audio instruction, and proactive help.

Paper link: https://huggingface.co/papers/2606.05121

2. Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories

Keywords: Deep-research agents, span-level error localization, error spans, claim-centric auditing, trajectory evidence

Category: Knowledge Representation and Reasoning

Research Objective:

– Develop a framework to audit deep-research agents by identifying error spans in their reasoning paths, enhancing reliability assessment beyond final answer evaluation.

Research Methods:

– Collect 2,790 trajectories from agent frameworks, backbone models, and benchmarks, convert logs into semantic spans; annotate error spans through LLM-assisted expert review.

– Build TELBench, a benchmark for identifying error spans and propose DRIFT, a claim-centric auditing framework to track agent claims and their support in trajectory evidence.

Research Conclusions:

– DRIFT improves span-level error localization and first-error accuracy by up to 30 percentage points, providing a process-level view of reliability in deep-research agents.

Paper link: https://huggingface.co/papers/2606.02060

3. OVO-S-Bench: A Hierarchical Benchmark for Streaming Spatial Intelligence in Multimodal LLMs

Keywords: streaming spatial intelligence, multimodal language models, OVO-S-Bench, allocentric mapping, chain-of-thought reasoning

Category: Multi-Modal Learning

Research Objective:

– To introduce a comprehensive benchmark, OVO-S-Bench, for evaluating streaming spatial intelligence in multimodal language models, particularly in applications such as robotics, augmented reality, and autonomous driving.

Research Methods:

– Creation of a fully human-annotated benchmark consisting of 1,680 questions over 348 videos, with annotations made by 12 trained annotators who also participated in blind cross-reviews. The benchmark features questions at four levels of abstraction and uses precise timing for queries and evidence intervals.

Research Conclusions:

– Existing MLLMs, especially those emphasizing streaming and spatial tuning, underperform in comparison to their backbones, with major challenges in allocentric mapping. Furthermore, ungrounded chain-of-thought reasoning exacerbates spatial errors. The study identifies significant bottlenecks and sets a challenging test environment for advancing streaming spatial MLLMs.

Paper link: https://huggingface.co/papers/2606.03890

4. M^3Eval: Multi-Modal Memory Evaluation through Cognitively-Grounded Video Tasks

Keywords: Multi-modal models, Memory, Video understanding, Cognitive psychology, Disentangled representations

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the memory capabilities of multi-modal models, particularly in maintaining disentangled representations and addressing human-like interference patterns to enhance video understanding systems.

Research Methods:

– Introduction of M^3Eval, a comprehensive evaluation framework grounded in cognitive psychology, with carefully constructed tasks to assess different memory dimensions in multi-modal models.

Research Conclusions:

– Multi-modal models show consistent weaknesses in maintaining disentangled representations and demonstrate interference patterns different from human memory. These models also have more reliable memory in the spatial domain rather than the temporal domain, with limited symbolic memory abilities. The findings emphasize the need for better memory mechanisms in multi-modal models.

Paper link: https://huggingface.co/papers/2606.05008

5. Benchmarks are Not Enough: RAMP for Runtime Assessing of Agentic Models in Production Systems

Keywords: RAMP, software engineering agents, runtime assessment, compiler-construction workloads, failure propagation

Category: AI Systems and Tools

Research Objective:

– The objective of this research is to develop the RAMP framework to provide a more realistic assessment of long-horizon software engineering agents, capturing the complexity of real-world production environments.

Research Methods:

– The research utilizes RAMP, built on the YatCC platform, to introduce compiler-construction workloads with serial dependencies and toolchain interactions, assessing performance through a unified runtime assessment architecture.

Research Conclusions:

– The findings reveal significant capability degradation in software engineering agents across serial workflows, highlighting systematic failure propagation and substantial resource inefficiencies, underlining the need for continuous, production-grounded evaluation methods.

Paper link: https://huggingface.co/papers/2605.27492

6. Streaming Communication in Multi-Agent Reasoning

Keywords: StreamMA, multi-agent reasoning, latency, pipelining, step-level scaling law

Category: Knowledge Representation and Reasoning

Research Objective:

– Introduce StreamMA to enable efficient multi-agent reasoning by streaming intermediate results to improve latency and effectiveness.

Research Methods:

– Utilization of a “generate-then-transfer” paradigm with formal analysis of stream, serial, and single protocols across multiple reasoning benchmarks and LLMs.

Research Conclusions:

– StreamMA reduces latency and enhances effectiveness by leveraging reliable early steps and pipelining; shows significant performance improvement over baselines.

– A new “step-level scaling law” is discovered, enhancing both effectiveness and efficiency by increasing per-agent steps.

Paper link: https://huggingface.co/papers/2606.05158

7. MMG2Skill: Can Agents Distill In-the-Wild Guides into Self-Evolving Skills?

Keywords: guide-to-skill learning, closed-loop framework, vision-language model, trajectory-level feedback, MMG2Skill

Category: Multi-Modal Learning

Research Objective:

– The research aims to convert web-based procedural guides into executable skills to enhance agent performance in various tasks such as GUI control, gameplay, and card play.

Research Methods:

– The MMG2Skill framework formalizes guide-to-skill learning by compiling guides into structured skills and refining them through closed-loop learning, leveraging a fixed vision-language model (VLM) agent, and trajectory-level root-cause feedback.

Research Conclusions:

– The framework consistently outperforms baseline agents with substantial macro-average performance gains across different models and settings. Structured skill construction and trajectory-driven revisions are crucial for these improvements, and early stopping mechanisms can efficiently reduce performance regression in successful tasks.

Paper link: https://huggingface.co/papers/2606.01993

8. MemTrain: Self-Supervised Context Memory Training

Keywords: Memory capabilities, Self-supervised training, Long-horizon language models, MemTrain, GRPO

Category: Natural Language Processing

Research Objective:

– Enhance the context-memory capability of long-horizon language model agents for improved downstream reasoning performance.

Research Methods:

– Introduced MemTrain, a self-supervised training framework utilizing two proxy tasks over unlabeled Wikipedia corpora, coupled with optimization through GRPO.

Research Conclusions:

– MemTrain framework significantly improves memory-intensive reasoning performance, achieving up to 17.67-point gains in benchmarks compared to direct task-specific post-training.

Paper link: https://huggingface.co/papers/2606.03197

9. ZipSplat: Fewer Gaussians, Better Splats

Keywords: Token-based feed-forward, 3D Gaussian Splatting, Pose-free imaging, Multi-view backbone, K-means clustering

Category: Computer Vision

Research Objective:

– To propose a new method, ZipSplat, that decouples 3D Gaussian placement from the pixel grid, enabling efficient scene reconstruction with fewer Gaussians and superior performance on pose-free imaging tasks.

Research Methods:

– Utilizes a token-based feed-forward model with a multi-view backbone to extract dense visual tokens, compresses them using k-means clustering, refines tokens with cross- and self-attention, and decodes them into Gaussians with a lightweight MLP.

Research Conclusions:

– ZipSplat sets a new state of the art in DL3DV and RealEstate10K benchmarks with significantly fewer Gaussians than previous methods, improving performance metrics by surpassing the best pose-free baseline by 2.1dB and 1.2dB PSNR respectively.

– It generalizes zero-shot to datasets like Mip-NeRF360 and ScanNet++, outperforming all comparable baselines.

Paper link: https://huggingface.co/papers/2606.05102

10. AAD-1: Asymmetric Adversarial Distillation for One-Step Autoregressive Video Generation

Keywords: AAD-1, Asymmetric Adversarial Distillation, autoregressive, motion collapse, training instability

Category: Generative Models

Research Objective:

– The main goal is to enhance one-step autoregressive image-to-video generation by resolving issues related to motion collapse and training instability while leveraging Asymmetric Adversarial Distillation.

Research Methods:

– The research employs a novel framework, AAD-1, that introduces an asymmetric design between the generator and discriminator, along with a phased training strategy to maintain stability and prevent motion collapse.

Research Conclusions:

– Extensive experiments on VBench confirm that AAD-1 delivers state-of-the-art results in one-step autoregressive video generation, achieving improved performance over current methods.

Paper link: https://huggingface.co/papers/2606.03972

11. AutoLab: Can Frontier Models Solve Long-Horizon Auto Research and Engineering Tasks?

Keywords: AutoLab, long-horizon iterative optimization, benchmark, persistent iteration, time awareness

Category: AI Systems and Tools

Research Objective:

– The study aims to evaluate the long-horizon iterative optimization capabilities of frontier models across diverse domains, focusing on the importance of persistence and time awareness over initial performance.

Research Methods:

– The introduction of AutoLab as a new benchmark for ultra long-horizon closed-loop optimization with 36 tasks spanning four domains. This benchmark challenges models to improve from a suboptimal baseline under a strict time budget.

Research Conclusions:

– AutoLab demonstrates that the key to success lies in an agent’s ability to persistently iterate, benchmark, and utilize empirical feedback. Time awareness and persistent iteration are crucial, as shown by the performance of claude-opus-4.6, while many other models underperform.

Paper link: https://huggingface.co/papers/2606.05080

12. AUDITFLOW: Executable Symbolic Environments for Structured Financial Reporting Verification

Keywords: Language-model agents, AuditFlow, US-GAAP taxonomy, Evidence Verification

Category: AI in Finance

Research Objective:

– The primary objective is to develop a framework, AuditFlow, for structured financial audit verification that separates adaptive search from deterministic verification.

Research Methods:

– A graph-grounded multi-agent framework is proposed, utilizing a static US-GAAP taxonomy graph and a dynamic XBRL filing graph, enabling typed tool access for fact retrieval, taxonomy traversal, numerical checking, and rule evaluation.

Research Conclusions:

– AuditFlow achieved 82.09% joint audit accuracy in tests with GPT-5.5, significantly outperforming existing baselines. The study demonstrated the necessity of a symbolic environment for reliable audit verification, as its removal drastically reduced accuracy.

Paper link: https://huggingface.co/papers/2606.03031

13. GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors

Keywords: GRAIL, 3D asset composition, sim-to-real transfer, humanoid robot, video foundation models

Category: Robotics and Autonomous Systems

Research Objective:

– The objective of this research is to enable effective sim-to-real transfer for robot control using diverse humanoid manipulation and locomotion data generated through 3D asset composition and video foundation models.

Research Methods:

– The research utilizes a digital generation pipeline, GRAIL, which composes 3D assets, simulator-ready scenes, and integrates video foundation models to synthesize interactions. This setup allows for model-based object tracking and interaction-aware optimization without relying on physical setups.

Research Conclusions:

– The study concludes that using GRAIL-generated data alone, task-general trackers can be trained and implemented successfully, achieving 84% real-world success on diverse object pick-up and 90% success on stair-climbing with the Unitree G1 humanoid robot.

Paper link: https://huggingface.co/papers/2606.05160

14. WebRISE: Requirement-Induced State Evaluation for MLLM-Generated Web Artifacts

Keywords: WebRISE, MLLM-generated, Interaction Contract Graphs, user-intent transitions, implicit constraints

Category: Multi-Modal Learning

Research Objective:

– The research introduces WebRISE to evaluate web artifacts generated by MLLMs through interaction contracts capturing user intent transitions and requirement checks.

Research Methods:

– WebRISE compiles task requirements into Interaction Contract Graphs that span multiple input modalities and include observable states, transitions, and DOM/visual assertions.

Research Conclusions:

– WebRISE outperforms traditional methods in error detection, revealing significant model performance gaps. Even the strongest MLLMs show limited transition validity and requirement coverage, with video providing the strongest interaction signal.

Paper link: https://huggingface.co/papers/2606.03220

15. KletterMix: Climbing Toward High-Quality German Pretraining Data

Keywords: KletterMix, German corpus, language model pretraining, translation quality, downstream evaluations

Category: Natural Language Processing

Research Objective:

– To introduce KletterMix, a high-quality German corpus for language model pretraining and assess its impact on German-language tasks.

Research Methods:

– Development of KletterMix via translation of an English corpus while maintaining structure, metadata, and diversity.

– Evaluation through controlled pretraining and comparison to existing German corpora.

Research Conclusions:

– KletterMix, by preserving much of the semantic and stylistic richness through careful translation, demonstrates measurable improvements in German-language downstream tasks, enhancing the German pretraining data ecosystem.

Paper link: https://huggingface.co/papers/2606.03773

16. Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation

Keywords: FiRe-OPD, On-Policy distillation, optimization stability, token selection, supervision signals

Category: Natural Language Processing

Research Objective:

– The research aims to enhance On-Policy Distillation (OPD) in large language models by introducing FiRe-OPD, a method that filters low-quality trajectories and applies soft reweighting to improve optimization stability and the selection of informative tokens.

Research Methods:

– The proposed FiRe-OPD method involves filtering trajectories to eliminate low-quality samples and employing a soft-weighting mechanism to highlight informative tokens, contrasting with hard token selection, and is evaluated in various OPD settings including single and multi-teacher contexts.

Research Conclusions:

– FiRe-OPD demonstrates superior performance in OPD optimization compared to recent token-level methods, achieving significant improvements in benchmarks such as AIME 2024 and Miner, and the method’s effectiveness is validated across different teaching scenarios.

Paper link: https://huggingface.co/papers/2606.02684

17. MapAgent: An Industrial-Grade Agentic Framework for City-scale Lane-level Map Generation

Keywords: MapAgent, autonomous driving, vectorized mapping, constraint-aware reasoning, production automation

Category: Robotics and Autonomous Systems

Research Objective:

– The objective is to develop an industrial-grade agentic architecture called MapAgent for automating the generation of specification-compliant lane maps, enhancing lane-level navigation for autonomous driving.

Research Methods:

– MapAgent integrates a vectorization backbone with constraint-aware reasoning and deterministic map editing to produce lane maps. It employs a Judge-Planner-Worker loop for specification inspection and error correction.

Research Conclusions:

– The integration of MapAgent into Baidu Maps has demonstrated its effectiveness, enabling lane-level map generation for over 360 cities with a production automation rate exceeding 95%.

Paper link: https://huggingface.co/papers/2606.04513

18. Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching

Keywords: Wide-baseline matching, multimodal large language models, ReasonMatch-Bench, Dynamic Correspondence Reinforcement Learning, data-generation pipeline

Category: Computer Vision

Research Objective:

– To improve spatial reasoning in multimodal large language models through a systematic evaluation and training framework.

Research Methods:

– Introduced ReasonMatch-Bench for testing wide-baseline matching capabilities.

– Developed a scalable data-generation pipeline to automatically extract training data from video-3D corpora.

– Proposed Dynamic Correspondence Reinforcement Learning to enhance training.

Research Conclusions:

– Existing MLLMs struggle with fine-grained wide-baseline correspondence tasks.

– The proposed framework and methods significantly improve model performance on ReasonMatch-Bench and related benchmarks.

Paper link: https://huggingface.co/papers/2606.03577

19. Self-Distilled Policy Gradient

Keywords: Self-distillation, Policy-gradient, Reinforcement Learning, KL regularization

Category: Reinforcement Learning

Research Objective:

– The paper aims to improve the stability and performance of reinforcement learning by combining self-distillation with verifier advantages and KL regularization.

Research Methods:

– It introduces the SDPG framework, which integrates on-policy self-distillation with a verifier-advantage approach and reference-policy KL regularization.

Research Conclusions:

– Empirical results demonstrate that SDPG enhances stability and performance compared to existing baselines like RLVR and self-distillation.

Paper link: https://huggingface.co/papers/2606.04036

20. Echo-Infinity: Learning Evolving Memory for Real-Time Infinite Video Generation

Keywords: Echo Infinity, evolving memory, autoregressive methods, Unified Relative RoPE, video diffusion transformers

Category: Generative Models

Research Objective:

– To develop Echo Infinity, an autoregressive framework for real-time infinite video generation utilizing a learnable evolving memory and Unified Relative RoPE.

Research Methods:

– Implementation of a dynamic and learnable memory system that abstracts and compresses any-length history at constant cost.

– Introduction of the Unified Relative RoPE Recipe to overcome finite RoPE constraints and improve training efficiency with video diffusion transformers.

Research Conclusions:

– Echo Infinity achieves state-of-the-art performance in both short and long video generation, showcasing the potential for practical real-time infinite video generation.

Paper link: https://huggingface.co/papers/2606.04527

21. ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning

Keywords: ThoughtFold, Large Reasoning Models, redundant explorations, Chain-of-Thoughts, fine-grained preference learning

Category: Knowledge Representation and Reasoning

Research Objective:

– Address over-thinking in Large Reasoning Models by eliminating redundant explorations in chain-of-thought reasoning processes using fine-grained preference learning.

Research Methods:

– Utilize fine-grained preference learning and an introspective strategy to identify and penalize redundant explorations, using masked preference optimization to streamline reasoning paths.

Research Conclusions:

– ThoughtFold significantly enhances efficiency by reducing token usage in reasoning models by approximately 56% while maintaining state-of-the-art accuracy.

Paper link: https://huggingface.co/papers/2606.03503

22. Qwen-Image-Flash: Beyond Objective Design

Keywords: Few-step distillation, Visual generative models, Training recipe, Teacher guidance, Task mixture

Category: Generative Models

Research Objective:

– The study aims to optimize few-step distillation for visual generative models by exploring improved training recipes beyond conventional distillation objectives.

Research Methods:

– The research investigates three critical factors: data composition, teacher guidance, and task mixture, using Qwen-Image-2.0 as a case study in text-to-image generation and instruction-guided image editing.

Research Conclusions:

– The findings indicate that a holistic approach in the training pipeline, which includes well-structured data, effective teacher guidance, and a balanced task mixture, is crucial for enhancing the effectiveness of few-step distillation in visual generative models.

Paper link: https://huggingface.co/papers/2606.03746

23. Reproducing, Analyzing, and Detecting Reward Hacking in Rubric-Based Reinforcement Learning

Keywords: reward hacking, rubric-based reinforcement learning, LLM-as-a-Judge, CHERRL, controller hacking environment

Category: Reinforcement Learning

Research Objective:

– Introduce CHERRL, a controlled environment for studying and analyzing reward hacking in rubric-based reinforcement learning using LLM judges.

Research Methods:

– Utilize CHERRL to inject known biases into LLM-as-a-Judge, allowing stable reproduction and explicit observation of reward hacking, and precise identification of its onset.

Research Conclusions:

– CHERRL provides a clean experimental testbed for exploring mechanisms and mitigations of reward hacking, facilitating automatic detection from training logs, and examining judge biases from discoverability and exploitability perspectives.

Paper link: https://huggingface.co/papers/2606.04923

24. Cosmos 3: Omnimodal World Models for Physical AI

Keywords: Cosmos 3, omnimodal world models, mixture-of-transformers architecture, Physical AI, embodied agents

Category: Multi-Modal Learning

Research Objective:

– Introduce Cosmos 3, an omnimodal world model designed to jointly process and generate different data types through a unified mixture-of-transformers architecture.

Research Methods:

– Develop a unified framework that integrates various modalities, including language, image, video, audio, and action sequences, achieving state-of-the-art performance across multiple tasks.

Research Conclusions:

– Cosmos 3 establishes a new benchmark for understanding and generation tasks and is recognized as the best open-source Text-to-Image, Image-to-Video, and policy models. The project’s resources are made available to the community to accelerate research in Physical AI.

Paper link: https://huggingface.co/papers/2606.02800

Keywords: task-specialized language models, multi-hop reasoning, question answering, context faithfulness, structured reasoning traces

Category: Natural Language Processing

Research Objective:

– To introduce the Optimal Cognitive Core (OCC) as a family of task-specialized small language models optimized for faithful question answering, requiring robust multi-hop reasoning without relying on memorized knowledge.

Research Methods:

– Development of a novel training pipeline to synthesize multi-context, multi-hop QA data at scale, resulting in over three million examples designed to enhance multi-hop reasoning and context faithfulness.

Research Conclusions:

– The OCC-RAG models, capable of producing structured reasoning traces with source citations, demonstrate that compact task-specialized language models can match or exceed the performance of larger general-purpose models in multi-hop reasoning and faithfulness benchmarks.

Paper link: https://huggingface.co/papers/2606.00683

2. Trust Region On-Policy Distillation

Keywords: Trust Region, On-Policy Distillation, Distribution Mismatch, Outlier Estimation, Off-Policy Guidance

Category: Natural Language Processing

Research Objective:

– The study aims to enhance the reliability and stability of On-Policy Distillation in large language models by addressing distribution mismatches between teacher and student models.

Research Methods:

– TrOPD utilizes trust regions to ensure reliable supervision, implements credit assignment strategies, and incorporates outlier estimation techniques, such as gradient clipping and forward-KL estimation, alongside off-policy guidance using teacher prefixes.

Research Conclusions:

– The proposed TrOPD consistently outperforms existing OPD baselines in areas such as mathematical reasoning, code generation, and general-domain benchmarks, demonstrating improved reliability and performance.

Paper link: https://huggingface.co/papers/2606.01249

3. KVarN: Variance-Normalized KV-Cache Quantization Mitigates Error Accumulation in Reasoning Tasks

Keywords: KVarN, KV-cache quantization, Hadamard rotation, dual-scaling variance normalization, autoregressive decoding

Category: Natural Language Processing

Research Objective:

– Introduce KVarN, a new calibration-free KV-cache quantizer to reduce error accumulation in autoregressive decoding for large language models.

Research Methods:

– Utilizes Hadamard rotation and dual-scaling variance normalization to address and correct token-scale errors in KV-cache quantization.

Research Conclusions:

– KVarN sets a new state-of-the-art for KV-cache quantization on generative benchmarks like MATH500, AIME24, and HumanEval, operating at 2-bit precision, and is shown to significantly minimize error accumulation.

Paper link: https://huggingface.co/papers/2606.03458

4. World Models Meet Language Models: On the Complementarity of Concrete and Abstract Reasoning

Keywords: Controlled Concrete Reasoning, World Models, Multimodal Large Language Models, Visual Simulation, Privileged-Future On-Policy Self-Distillation

Category: Multi-Modal Learning

Research Objective:

– To enhance prediction accuracy and robustness by combining visual simulation with abstract reasoning using privileged future information.

Research Methods:

– Developed a method called Privileged-Future On-Policy Self-Distillation (PF-OPSD) which employs ground-truth future videos as privileged context for training.

Research Conclusions:

– PF-OPSD outperforms baseline methods by 10.6% and 10.9% on constructed benchmarks VRQABench and OpenWorldQA, enhancing robustness against noisy or conflicting rollouts.

Paper link: https://huggingface.co/papers/2606.03603

5. MIRA: Mid-training Rubric Anchoring for Source-Aware Data Selection

Keywords: MIRA, Mid-training, data selection, source-aware filtering, self-anchored rubric discovery

Category: Natural Language Processing

Research Objective:

– The objective is to develop MIRA, a source-aware filtering framework for mid-training data selection in LLM development, focusing on balancing scalability and semantic accuracy across heterogeneous data sources using self-anchored rubric discovery.

Research Methods:

– MIRA employs self-anchored rubric discovery to build rubrics during data selection, allowing it to evaluate source groups effectively and use scalable student scorers for full-corpus filtering.

Research Conclusions:

– MIRA improves data selection by outperforming baseline methods in nine code benchmarks, matching the performance of full-corpus runs while using only half the tokens, thereby demonstrating efficiency and effectiveness.

Paper link: https://huggingface.co/papers/2605.30288

6. Benchmarking Visual State Tracking in Multimodal Video Understanding

Keywords: Visual State Tracking, Multimodal Large Language Models, Continuous Perception, VSTAT, Video Understanding

Category: Multi-Modal Learning

Research Objective:

– The objective is to evaluate and improve visual state tracking in Multimodal Large Language Models (MLLMs) which struggle particularly with video content.

Research Methods:

– A new benchmark called VSTAT is introduced, consisting of 834 video clips with 1,500 questions, demanding continuous perception and integration across video streams.

Research Conclusions:

– Despite strong performance on other benchmarks, MLLMs perform poorly on VSTAT, failing to visually perceive tracked events. Recent agentic approaches do not alleviate these challenges effectively.

Paper link: https://huggingface.co/papers/2606.03920

7. NVIDIA OmniDreams: Real-Time Generative World Model for Closed-Loop Autonomous Vehicle Simulation

Keywords: OmniDreams, generative world model, action-conditioned video, photorealistic sensor generation, autonomous driving

Category: Generative Models

Research Objective:

– The main objective is to develop OmniDreams, a generative world model trained from the Cosmos diffusion model, to enable real-time, action-conditioned video generation for evaluating autonomous driving policies in complex, unseen scenarios.

Research Methods:

– Utilization of closed-loop simulation systems, mid- and post-training on 21k hours of driving scenarios, and integration with Alpamayo 1 policy model and AlpaSim orchestrator for creating a reactive simulation environment.

Research Conclusions:

– OmniDreams successfully synthesizes complex, unobserved driving scenarios, supports scalable and comprehensive autonomous driving policy evaluation, and demonstrates potential as a backbone for policy architectures, outperforming existing models in preliminary tests using fewer parameters.

Paper link: https://huggingface.co/papers/2606.03159

8. Ψ-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues

Keywords: Persuasive Conversation, Personalized Agents, User Profiles, LLMs, Persona-Sensitive Influencing

Category: Natural Language Processing

Research Objective:

– To systematically evaluate the proactive personalization of language models in realistic interactions, focusing on persuasion through conversation.

Research Methods:

– Introduction of Ψ-Bench, a benchmark designed to assess the ability of Large Language Models (LLMs) to influence users using conversation in scenarios with embedded user profiles.

Research Conclusions:

– Findings indicate that while LLMs can generate coherent arguments, they show limited effectiveness in persuasion. Providing access to user-specific profiles significantly enhances performance by 18.24%. The study emphasizes the importance of persona-sensitive influencing as a direction for developing more proactive personalized LLM agents.

Paper link: https://huggingface.co/papers/2606.02754

9. Small RL Controller, Large Language Model: RL-Guided Adaptive Sampling for Test-Time Scaling

Keywords: Adaptive sampling, Markov decision process, Reinforcement learning, Lagrangian relaxation, Large language models

Category: Reinforcement Learning

Research Objective:

– The paper aims to optimize adaptive sampling for large language models using an MDP framework to balance correctness, latency, and computational cost.

Research Methods:

– Adaptive sampling is formulated as a Markov decision process and optimized via reinforcement learning to train a lightweight sampling controller.

Research Conclusions:

– The proposed method allows for efficient trade-offs between correctness, sampling rounds, and computation cost, showing superiority over existing baselines like ASC and ESC.

Paper link: https://huggingface.co/papers/2606.03102

10. PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training

Keywords: PaddleOCR-VL-1.6, document parsing, data optimization, post-training, OmniDocBench

Category: Computer Vision

Research Objective:

– To enhance document parsing performance by improving on the previous PaddleOCR-VL-1.5 model through targeted data optimization and a progressive post-training approach.

Research Methods:

– Introduced a region-aware data optimization framework to identify and enhance weak areas from previous models.

– Employed a progressive post-training strategy based on curated data selection and reinforcement learning to achieve superior model performance.

Research Conclusions:

– PaddleOCR-VL-1.6 achieved a state-of-the-art score of 96.33% on OmniDocBench v1.6, showcasing its competitiveness against leading VLMs and establishing a practical post-training recipe for the PaddleOCR-VL series.

Paper link: https://huggingface.co/papers/2606.03264

11. Decentralized Instruction Tuning: Conflict-Aware Splitting and Weight Merging

Keywords: Instruction tuning, large language models, gradient interference, decentralized training, weight merging

Category: Natural Language Processing

Research Objective:

– The study aims to enhance instruction tuning of large language models by mitigating the issues of gradient interference and bandwidth-heavy synchronization through a decentralized training method.

Research Methods:

– The research employs a decentralized training approach that involves independently training partitions of mixed datasets, resolving gradient conflicts, and merging results via a weighted averaging strategy.

Research Conclusions:

– The proposed method, MERIT, effectively improves tuning performance on large language models like Qwen2.5-VL-3B and scales to larger models, showing comparable or superior performance to centralized methods with reduced communication costs.

Paper link: https://huggingface.co/papers/2606.01717

12. OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification

Keywords: OmniOPD, On-Policy Distillation, semantic similarity, black-box teachers, token-level feedback

Category: Reinforcement Learning

Research Objective:

– The objective of the research is to improve the limitations of standard On-Policy Distillation (OPD) by using a chunk-level semantic similarity approach instead of token-level logits, with a focus on black-box teachers.

Research Methods:

– The proposed method involves using a logit-free, chunk-level supervision signal that incorporates Monte Carlo rollouts and a continuous semantic similarity metric over multi-token chunks. This method is further enhanced using a peak-entropy scheduler and a Dirichlet-Multinomial Bayesian prior to ensure stability and prevent policy collapse.

Research Conclusions:

– OmniOPD outperforms the standard OPD method by up to +28.64% on competitive benchmarks like math, confirming its effectiveness. Additionally, it yields an additional +9.54% improvement when paired with stronger black-box teachers, surpassing the performance of self-exploratory Reinforcement Learning, thus proving its superiority in extracting reliable learning signals.

Paper link: https://huggingface.co/papers/2606.01476

13. Value-Aware Stochastic KV Cache Eviction for Reasoning Models

Keywords: Value-aware Stochastic KV Cache Eviction, Reasoning Models, Cache Diversity, KV Cache Eviction, FlashAttention2

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to improve the accuracy of reasoning models under compression by introducing a new KV cache eviction method that protects large-magnitude states and promotes diverse eviction decisions.

Research Methods:

– The researchers identified critical factors affecting KV cache eviction accuracy, including the impact of large-magnitude value states and the introduction of stochasticity for increasing cache diversity. They proposed the Value-aware Stochastic KV Cache Eviction (VaSE) method.

Research Conclusions:

– VaSE enhances the average accuracy of Qwen3 models with 4x KV cache compression across various reasoning tasks compared to state-of-the-art selection methods, outperforming the strongest eviction method by over 4%, effectively balancing efficiency and accuracy.

Paper link: https://huggingface.co/papers/2606.03928

14. Domain-Specific Data Synthesis for LLMs via Minimal Sufficient Representation Learning

Keywords: Domain-specific data synthesis, Inductive paradigm, Reference examples, Prompt tuning, Synthetic data distribution

Category: Machine Learning

Research Objective:

– The objective is to address the challenge of domain-specific data synthesis using an inductive approach, learning domain representations from reference examples to improve code benchmark performance.

Research Methods:

– The researchers developed a novel framework, DOMINO, which integrates prompt tuning with a contrastive disentanglement objective to extract domain-level patterns from reference examples, mitigating overfitting.

Research Conclusions:

– DOMINO expands the support of synthetic data distribution, ensuring diversity and improving Pass@1 accuracy by up to 4.63% on coding benchmarks with implicit domain definitions, thus proving its effectiveness and robustness.

Paper link: https://huggingface.co/papers/2605.30039

15. αDepth: Learning Single-Pass Soft Boundary Decomposition for Stereo Conversion

Keywords: αDepth, Circular Alpha Representation, stereo conversion, soft boundaries, layered representation

Category: Computer Vision

Research Objective:

– To address the challenges posed by soft boundaries in stereo conversion through a novel layered representation approach, thereby improving depth modeling accuracy in complex scenes.

Research Methods:

– Introduces αDepth, which utilizes Circular Alpha Representation (CAR) to allow for local boundary decomposition and efficient scene-level inference, overcoming the limitations of traditional matting techniques.

Research Conclusions:

– αDepth achieves state-of-the-art performance in stereo conversion, effectively eliminating issues like background bleeding and structural distortions at soft boundaries without needing manual intervention.

Paper link: https://huggingface.co/papers/2606.00386

16. Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models

Keywords: YOLO26, real-time vision, NMS-free, segmentation, pose estimation

Category: Computer Vision

Research Objective:

– To develop a unified real-time vision model family, YOLO26, that overcomes the limitations of existing YOLO models by offering NMS-free inference, improved training strategies, and multi-task capabilities.

Research Methods:

– Utilizes a dual-head design for native NMS-free end-to-end inference, employs a hybrid Muon-SGD optimizer and Progressive Loss for training, and ensures small object detection through STAL.

Research Conclusions:

– YOLO26 advances accuracy and efficiency across tasks like detection, segmentation, and pose estimation, reaching 40.9-57.5 mAP on COCO with minimized latency, enhancing the performance of real-time detectors.

Paper link: https://huggingface.co/papers/2606.03748

17. Mitigating Perceptual Judgment Bias in Multimodal LLM-as-a-Judge via Perceptual Perturbation and Reward Modeling

Keywords: Perceptual Judgment Bias, Multimodal Large Language Models, Visual Perturbations, Perceptual Fidelity

Category: Multi-Modal Learning

Research Objective:

– The study aims to address the Perceptual Judgment Bias in multimodal large language models, where models prefer textual plausibility over visual evidence.

Research Methods:

– The researchers propose a training framework using a Perceptually Perturbed Judgment Dataset and a structured GRPO-based reward combined with a batch-ranking objective to improve perceptual fidelity and evaluation consistency.

Research Conclusions:

– The approach significantly enhances perceptual fidelity, ranking coherence, and alignment with human evaluation in multimodal judge benchmarks, establishing a pathway for robust and interpretable multimodal judges.

Paper link: https://huggingface.co/papers/2606.02578

18. Prior Availability in Industrial Visual Sim-to-Real: A Review of CAD-Guided and CAD-Unavailable Regimes

Keywords: domain-gap, Industrial visual sim-to-real, CAD-available, CAD-unavailable, boundary-prior settings

Category: Computer Vision

Research Objective:

– The objective of the research is to reframe Industrial visual sim-to-real as a domain-gap problem categorized by prior availability, for robust deployment across varied industrial conditions.

Research Methods:

– The study distinguishes between CAD-available, CAD-unavailable, and boundary-prior settings, using empirical anchors on datasets such as T-LESS/BOP, MVTec AD, and VisA.

Research Conclusions:

– The findings suggest that simply counting CAD renders does not ensure successful transfer; source distribution design and detector capacity are more significant, emphasizing the importance of prior availability to support deployment decisions.

Paper link: https://huggingface.co/papers/2605.30581

19. WALL-WM: Carving World Action Modeling at the Event Joints

Keywords: video-action learning, Vision-Language-Action, semantic events, event-grounded pretraining, state-of-the-art performance

Category: Multi-Modal Learning

Research Objective:

– The research aims to enhance video-action learning by transitioning from fixed action chunks to semantic events, leading to more flexible and scalable Vision-Language-Action training and inference.

Research Methods:

– The WALL-WM model employs event-grounded Vision-Language-Action pretraining and utilizes a data ecosystem comprising event-level captions and cluster-balanced sampling to support scalable learning.

– Two inference modes are introduced: the event mode for variable-length execution and the unified mode with Staircase Decoding for fixed-length chunk inference.

Research Conclusions:

– WALL-WM demonstrates broad generalization across languages, scenes, and tasks, achieving state-of-the-art performance in large-scale real-world generalization evaluations.

Paper link: https://huggingface.co/papers/2606.01955

20.

Paper link: 

21. Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations

Keywords: Deception Detection, Linear Probes, Distributional Shift, AUROC, Cross-Domain Transfer

Category: Natural Language Processing

Research Objective:

– To assess why linear probes for deception detection in large language models fail under distributional shifts despite high performance on clean benchmark data.

Research Methods:

– Conducted systematic tests across the Gemma 3 model family, examining four hypotheses about deception encoding, and analyzed multi-dimensional probes with style-augmented data.

Research Conclusions:

– Probes obtain near-perfect AUROC on clean data but fail on stylistic shifts; however, style-augmented probes regain high detection accuracy.

– Single-direction and entropy-proxy hypotheses are rejected, with deception encoded in multi-dimensional, distributed sub-threshold features.

– Probe fragility is attributed to distributional narrowness rather than architectural limitations, as style-augmented probes recover detection effectiveness.

Paper link: https://huggingface.co/papers/2605.27958

22. BA-T: An Iterative Transformer for Two-View Bundle Adjustment

Keywords: Iterative Transformer, 3D reconstruction, cross-view consistency, bundle adjustment, lightweight design

Category: Computer Vision

Research Objective:

– To improve 3D reconstruction accuracy and cross-view consistency using an iterative Transformer architecture inspired by bundle adjustment.

Research Methods:

– Implement an iterative Transformer called BA-T that uses structured updates as a repeatable layer in implicit token space, refining predictions through latent residual in a single lightweight layer.

Research Conclusions:

– BA-T enhances pose and reconstruction accuracy across iterations, achieves stronger cross-view consistency than conventional models, and matches or surpasses larger models using only 16% of their decoder parameters.

Paper link: https://huggingface.co/papers/2606.03287

23. AURA: Action-Gated Memory for Robot Policies at Constant VRAM

Keywords: Embodied AI, AURA-Mem, KV-cache, Recurrent Memory, Action-Utility

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces AURA-Mem, a recurrent memory system designed to adapt to the constraints of embodied AI by reducing memory writes and efficiently managing memory resources.

Research Methods:

– AURA-Mem employs a frozen vision-language-action backbone and a learned gate mechanism to determine when memory writes are necessary, based on action-impacting observations.

Research Conclusions:

– AURA-Mem significantly outperforms traditional KV-cache systems by reducing memory writes by up to 9.19 times while maintaining accuracy, especially in bandwidth-limited environments typically encountered in robotics.

Paper link: https://huggingface.co/papers/2606.02775

24. ClawHub Security Signals: When VirusTotal, Static Analysis, and SkillSpector Disagree

Keywords: AI agents, layered security governance, scanner disagreement, VirusTotal, NVIDIA SkillSpector

Category: AI Systems and Tools

Research Objective:

– This research focuses on evaluating the discrepancies in detection rates across different types of scanners and attack surfaces within agent skills, which amplify the capabilities of AI agents.

Research Methods:

– The study utilizes a sanitized dataset named ClawHub Security Signals encompassing 67,453 versions of agent skills, analyzing disagreements among three scanning tools: VirusTotal, static heuristic analysis, and NVIDIA SkillSpector.

Research Conclusions:

– The extensive scanner disagreement highlights the need for layered security governance instead of relying on single-scanner decisions, with implications for agent-skill security requiring tailored security triage models.

Paper link: https://huggingface.co/papers/2606.01494

25. A Multi-AI-agent Framework Enabling End-to-end Finite Element Analysis for Solid Mechanics Problems

Keywords: Multi-agent framework, Large language models, Finite element analysis, Solid mechanics, AI-empowered optimization

Category: AI Systems and Tools

Research Objective:

– The study aims to automate finite element analysis for solid mechanics using AbaqusAgent, which uses large language models to convert natural-language instructions into executable simulations.

Research Methods:

– AbaqusAgent deploys a multi-agent framework composed of six agents handling pre-processing and post-processing steps in FEA, successfully validating 50 solid mechanics problems with an 86% success rate.

Research Conclusions:

– AbaqusAgent improves the efficiency and accessibility of FEA, enhances human-simulation interaction, and integrates with AI-driven workflows for optimization and material characterization, thus advancing computational mechanics education.

Paper link: https://huggingface.co/papers/2606.00138

26. Conditional Hypothesis Generation for LLM-Based Text Analysis with Researcher-Specified Covariates

Keywords: Conditional hypothesis generation, covariates, stratum imbalance, sign reversal, computational social science

Category: Natural Language Processing

Research Objective:

– Introduce a conditional hypothesis generation framework that incorporates covariates to identify meaningful language variations across subgroups.

Research Methods:

– Propose two econometrics-inspired methods: feature–covariate interactions for detecting sign reversals and within-stratum demeaning with inverse-frequency reweighting for equalizing underrepresented strata.

Research Conclusions:

– Demonstrated improved performance over global baselines through synthetic experiments and expert evaluations on real-world datasets, showing that covariate-aware generation surfaces more useful hypotheses within relevant subgroups.

Paper link: https://huggingface.co/papers/2606.03029

27. MERIT: Learning Disentangled Music Representations for Audio Similarity

Keywords: MERIT framework, conditional audio generation, source-separated stems, disentangled music representations, factor-specific music representations

Category: Generative Models

Research Objective:

– Introduce MERIT, a framework designed to learn disentangled music representations focusing on melody, rhythm, and timbre to allow for nuanced musical queries.

Research Methods:

– Employ a novel training strategy using conditional audio generation and source-separated stems to encourage single-factor variation in training data.

Research Conclusions:

– Demonstrated strong factor-wise disentanglement where each representation head responds primarily to its intended perceptual dimension, applicable across both synthetic and real-world audio.

Paper link: https://huggingface.co/papers/2605.27346

28. Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces

Keywords: Chain-of-thought, Fine-tuning, Post-conclusion continuation, Harmful continuation, Uncertainty

Category: Knowledge Representation and Reasoning

Research Objective:

– Investigate the impact of post-conclusion continuations in answer-correct long chain-of-thought (CoT) traces on the fine-tuning outcomes of reasoning-oriented language models.

Research Methods:

– Utilized a delete-only editor to execute answer-preserving suffix removal, comparing CoT-based supervised fine-tuning (SFT) on original and processed traces to assess training effects.

Research Conclusions:

– Post-conclusion continuations in CoT traces negatively affect training, causing uncertainty-geometry mismatches. The introduction of Harmful Continuation Cut (HCC) as a boundary proxy offers a solution to address this issue.

Paper link: https://huggingface.co/papers/2605.29288

29. PlatonicNav: Unveiling Semantic Correspondence in Navigation with Platonic Topological Maps

Keywords: Embodied visual navigation, Semantic maps, Vision-only approach, Training-free framework, PlatonicNav

Category: Robotics and Autonomous Systems

Research Objective:

– To develop a training-free framework for embodied navigation using a vision-only approach that creates semantic maps and achieves language grounding through blind matching without utilizing paired vision-language data.

Research Methods:

– Introduce PlatonicNav, a framework that extends the Platonic Representation Hypothesis by employing a self-supervised visual encoder to craft a Platonic Topological Map which fuses geometric and semantic data.

Research Conclusions:

– Extensive experiments indicate that PlatonicNav can generalize across tasks, modalities, and embodiments without explicit cross-modal training, as evidenced by benchmarks and deployments.

Paper link: https://huggingface.co/papers/2606.01788

30. Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams

Keywords: Adaptive Auto-Harness, LLM agents, evolution loss, adaptation loss, stateful multi-agent evolver

Category: AI Systems and Tools

Research Objective:

– The primary goal of the research is to address dynamic task streams in auto-harness systems through a framework that decomposes performance gaps into evolution and adaptation losses for sustained performance improvement.

Research Methods:

– Implementation of Adaptive Auto-Harness, which involves a stateful multi-agent evolver, a harness tree with solve-time routing, and human-steering hooks to manage and adapt to evolving task environments.

Research Conclusions:

– The Adaptive Auto-Harness framework outperforms existing auto-harness baselines by effectively utilizing construction, routing, and targeted human steering across various streams such as prediction-market, security-competition, and event-forecasting.

Paper link: https://huggingface.co/papers/2606.01770

31. Decoupled Residual Denoising Diffusion Models for Unified and Data Efficient Image-to-Image Translation

Keywords: Decoupled Residual Denoising Diffusion, unified I2I translation, domain harmonization, data efficiency, diffusion models

Category: Generative Models

Research Objective:

– To propose Decoupled Residual Denoising Diffusion models (DRDD) that enhance data efficiency and performance in unified image-to-image (I2I) translation by separating noise diffusion for domain harmonization from residual diffusion for semantic mapping.

Research Methods:

– DRDD introduces two sequential and independent diffusion stages: stochastic noise diffusion for domain harmonization and manifold lifting, and deterministic residual diffusion for semantic mapping within a fixed-noise domain.

Research Conclusions:

– DRDD is compatible with mainstream diffusion models and provides robust, unified I2I translation even with limited paired data, offering substantial improvements in data efficiency.

Paper link: https://huggingface.co/papers/2606.01048

32. Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching

Keywords: Bootstrap Your Generator, unpaired training, flow matching, gradient routing, image and video editing

Category: Generative Models

Research Objective:

– To develop a framework, Bootstrap Your Generator (ByG), enabling unpaired training of flow matching editing models, leveraging base model knowledge for improved generalization without extensive datasets.

Research Methods:

– Utilization of instruction-following cues with cycle-consistency for structure preservation and routing gradients from downstream losses over clean predictions to noisy training states.

Research Conclusions:

– ByG demonstrates state-of-the-art performance in data-scarce settings, effectively generalizes to new domains, and outperforms supervised baselines by bridging the train-inference gap and extracting robust semantic cues.

Paper link: https://huggingface.co/papers/2606.03911

33. Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

Keywords: Deep Learning, Sleep Paradigm, Knowledge Seeding, Memory Consolidation, Reinforcement Learning

Category: Machine Learning

Research Objective:

– The study introduces a novel Sleep paradigm in deep learning, inspired by human learning processes, to enhance long-term learning and self-improvement.

Research Methods:

– Utilizes a two-stage process comprising Memory Consolidation through Knowledge Seeding and a Dreaming phase involving Reinforcement Learning, simulating sleep to consolidate memories and self-improve.

Research Conclusions:

– Experiments demonstrate the effectiveness of the Sleep paradigm in improving continual learning, knowledge incorporation, and few-shot generalization, highlighting its importance in enhancing long-term learning capabilities.

Paper link: https://huggingface.co/papers/2606.03979

34. TRON: Targeted Rule-Verifiable Online Environments for Visual Reasoning RL

Keywords: TRON, Scalable Reinforcement Learning, Visual Reasoning, Online Environment Substrate, Multimodal Reasoning

Category: Reinforcement Learning

Research Objective:

– The paper introduces TRON, which aims to provide scalable and controllable reinforcement learning specifically for visual reasoning through an online environment that generates limitless diverse training instances with verifiable answers.

Research Methods:

– The study presents an online environment substrate where training instances are generated on demand by a controllable generator-verifier program. This allows the creation of an unbounded stream of instances tailored to the difficulty requirement of ongoing training. The TRON suite includes 520 environments divided into five ability buckets, supporting both holistic and specialized training models.

Research Conclusions:

– The implementation of TRON and its method notably enhances performance on ten multimodal reasoning benchmarks, showcasing efficacy across various models such as Qwen3-VL-4B, Qwen2.5-VL-7B, and MiMo-VL-7B-SFT.

Paper link: https://huggingface.co/papers/2606.01599

35. AutoMedBench: Towards Medical AutoResearch with Agentic AI Models

Keywords: AutoMedBench, autonomous agents, medical-AI research, validation, verification

Category: AI in Healthcare

Research Objective:

– The objective of AutoMedBench is to establish a comprehensive benchmark for autonomous medical-AI research. It aims to evaluate the performance of medical-AI agents across various workflow stages, with a particular focus on the validation stage.

Research Methods:

– The study organizes agent performance evaluation into a five-stage workflow: Plan, Setup, Validate, Inference, and Submit. It includes long-horizon tasks in medical imaging and multimodal inference, evaluated under two difficulty tiers. Stage-level analysis is performed to identify strengths and weaknesses in the workflow.

Research Conclusions:

– Findings highlight that the validation stage is the weakest link in the medical-AI workflow process, while the setup stage is the strongest. The study reveals that current agents excel in making pipelines executable but struggle with reliability verification, as evidenced by post-run error analysis focusing on verification and submission failures.

Paper link: https://huggingface.co/papers/2606.01961

36. A Local Perturbation Theory for Cross-Domain Interference and Recovery in Multi-Domain RL

Keywords: Multi-domain reinforcement learning, Reinforcement learning, Large language models, Catastrophic forgetting, Conflict subspace

Category: Reinforcement Learning

Research Objective:

– To investigate and address performance degradation in multi-domain reinforcement learning (RL) within large language models (LLMs) due to shared computational pathways.

Research Methods:

– The study applies targeted refresh and rollback techniques and proves the impact of domain-specific training using a local perturbation model and analysis of second-order damage terms in a conflict subspace.

Research Conclusions:

– The research demonstrates that sparse, low-dimensional parameter changes can experience interference in multi-domain RL, leading to performance loss. Techniques like domain refresh and rollback aid in recovering capabilities in specific tasks with limited collateral damage, providing a mechanistic account of interference and recovery.

Paper link: https://huggingface.co/papers/2606.02398

37. Humanoid-GPT: Scaling Data and Structure for Zero-Shot Motion Tracking

Keywords: Humanoid-GPT, GPT-style Transformer, zero-shot generalization, motion corpus, dynamic behaviors

Category: Robotics and Autonomous Systems

Research Objective:

– Introduce Humanoid-GPT, a Transformer model designed for whole-body control using a large-scale motion dataset.

Research Methods:

– Utilize causal attention in a generative Transformer pre-trained on a 2B-frame retargeted corpus including major mocap datasets and in-house recordings.

Research Conclusions:

– The model achieves robust zero-shot generalization to unseen motions and tasks, establishing a new performance frontier in tracking dynamic and complex behaviors.

Paper link: https://huggingface.co/papers/2606.03985

38. From Activation to Causality: Discovery of Causal Visual Representations in the Human Brain

Keywords: BrainCause framework, generative and brain models, causal testing, neural representations, image-to-fMRI encoding model

Category: Multi-Modal Learning

Research Objective:

– The study aims to identify valid neural representations of visual concepts in the human brain using the BrainCause framework, addressing the insufficient evidence provided by activation alone.

Research Methods:

– The BrainCause framework combines generative and brain models to create controlled stimuli and conduct targeted causal testing. It uses an image-to-fMRI encoding model to predict brain responses and identify specific neural representations.

Research Conclusions:

– The approach demonstrates that activation alone is insufficient to confirm concept representation, as many localizations could be false positives without causal validation. It recovers known functional localizations and identifies new candidate representations across multiple concepts.

Paper link: https://huggingface.co/papers/2605.23895

Keywords: multi-agent framework, scientific figures, editable SVGs, CraftBench, AI Systems and Tools

Category: AI Systems and Tools

Research Objective:

– To develop a multi-agent framework that generalizes figure generation across different types and input conditions, producing editable output formats.

Research Methods:

– Implemented Crafter, a multi-agent harness, and CraftEditor for converting raster outputs into editable SVGs; introduced CraftBench for benchmarking with human quality annotation.

Research Conclusions:

– Experiments demonstrate Crafter’s superior performance compared to standalone generators, and CraftEditor’s successful conversion into editable SVGs that outperform all baselines.

Paper link: https://huggingface.co/papers/2605.30611

2. A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks

Keywords: Automated benchmark generation, Adaptive Contrastive n-gram model, Tool Sequence Evolution, Task Synthesis, τ^c-Bench

Category: AI Systems and Tools

Research Objective:

– The research aims to address the limitations of existing benchmarks by proposing TASTE, an automated method for generating challenging tasks with broader tool-use coverage.

Research Methods:

– TASTE employs an Adaptive Contrastive n-gram model to generate valid tool sequences and uses clustering to select representative sequences. It then refines these sequences through iterative difficulty evolution.

Research Conclusions:

– The study concludes that existing high scores on benchmarks like τ^2-Bench may reflect saturation. TASTE-generated tasks increase difficulty and expand the tool combinations, enabling more robust evaluation of agent capabilities.

Paper link: https://huggingface.co/papers/2605.28556

3. Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses

Keywords: Reinforcement Learning, Search Agents, Harness-1, Stateful Search, Curated Recall

Category: Reinforcement Learning

Research Objective:

– To enhance retrieval performance of a 20B search agent by separating semantic decision-making from environmental bookkeeping within a stateful search framework.

Research Methods:

– A 20B search agent, known as Harness-1, is trained using reinforcement learning. The agent operates within a stateful search harness that handles environmental bookkeeping such as working memory, candidate selection, evidence curation, and verification.

Research Conclusions:

– Harness-1 demonstrates superior performance across eight retrieval benchmarks, achieving a 0.730 average curated recall. The approach outperforms smaller search agents and competes with larger models, showing strong results in transfer benchmarks which suggest generalizability beyond training domains.

Paper link: https://huggingface.co/papers/2606.02373

4. Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding

Keywords: Domino, Speculative Decoding, Parallel Backbone, Causal Dependency

Category: Natural Language Processing

Research Objective:

– This paper introduces Domino, a speculative decoding framework aimed at enhancing LLM inference speed by decoupling causal dependency modeling from costly autoregressive drafting methods.

Research Methods:

– The method involves using a parallel draft backbone to generate initial distributions, followed by a lightweight causal refinement step, supported by a base-anchored training curriculum to stabilize encoding and optimize distribution.

Research Conclusions:

– Experiments demonstrate that Domino achieves significant speedups in both end-to-end execution and throughput, with improvements of up to \(5.49\times\) and \(5.8\times\) respectively, when applied to different backend systems.

Paper link: https://huggingface.co/papers/2605.29707

5. NITP: Next Implicit Token Prediction for LLM Pre-training

Keywords: Next Implicit Token Prediction, language model, dense continuous supervision, implicit semantic content, optimization landscape

Category: Natural Language Processing

Research Objective:

– The study introduces Next Implicit Token Prediction (NITP) to improve language model generalization by incorporating dense continuous supervision directly in the representation space.

Research Methods:

– NITP enhances training by predicting the implicit semantic content of the next token using shallow-layer representations from the same model, providing self-supervised targets.

Research Conclusions:

– Empirical results show significant performance improvements across various model sizes with minimal computational overhead, demonstrating notable gains in benchmark tasks like MMLU-Pro, C3, and CommonsenseQA.

Paper link: https://huggingface.co/papers/2605.24956

6. Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?

Keywords: Target Viewpoint Reproduction, visual history, post-training framework, Multi-turn GRPO

Category: Reinforcement Learning

Research Objective:

– The study aims to address the challenge of Target Viewpoint Reproduction (TVR), requiring foundation models to actively adjust 3D viewpoints to match target images, thus evaluating and enhancing spatial intelligence capabilities.

Research Methods:

– The research introduces a unified TVR post-training framework utilizing expert-trajectory SFT, rationale-supervised CoT-SFT, offline Single-turn GRPO, and on-policy Multi-turn GRPO from live simulator rollouts to improve model performance.

Research Conclusions:

– The study identifies spatial intelligence gaps in foundation models and demonstrates that visual-action SFT significantly boosts success rates, using TVRBench as a benchmark to measure and train models for active perception and action in 3D environments, achieving a performance increase to 50.8% in a 9B open-source model and 51.4% through Multi-turn GRPO.

Paper link: https://huggingface.co/papers/2606.01247

7. X-Stream: Exploring MLLMs as Multiplexers for Multi-Stream Understanding

Keywords: X-Stream, multi-stream streaming understanding, MLLMs, multi-modal large language models, concurrent streams

Category: Multi-Modal Learning

Research Objective:

– Introduce X-Stream, the first benchmark for evaluating the understanding of concurrent multi-streams in real-world applications like live sports broadcasting and autonomous driving.

Research Methods:

– Utilized a dual-verification pipeline to ensure comprehensive evaluation over 4,220 QA pairs across 932 videos in multi-window, multi-view, and multi-device scenarios; evaluated MLLMs using Signal Multiplexing Theory.

Research Conclusions:

– The study demonstrates that current state-of-the-art MLLMs struggle with concurrent streams, achieving only 50% efficiency, and highlights the limitations of existing multiplexing schemes, offering guidance for future multi-stream systems development.

Paper link: https://huggingface.co/papers/2606.02482

8. Which Pretraining Paradigm Better Serves Spatial Intelligence? An Empirical Comparison of Vision-Language and Video Generation Models

Keywords: Vision-Language Models, Video Generation Models, Spatial Intelligence, Semantic Tagging, 3D Geometry Prediction

Category: Multi-Modal Learning

Research Objective:

– To systematically compare Vision-Language Models (VLMs) and Video Generation Models (VGMs) for spatial intelligence tasks by evaluating their capabilities in semantic tagging, instance grouping, and 3D geometry prediction.

Research Methods:

– Conduct a frozen-feature probing study to analyze the representation strengths of VLMs and VGMs across key spatial intelligence axes.

Research Conclusions:

– Demonstrated complementarity between VLMs and VGMs wherein VLMs excel in semantic tagging and instance grouping, while VGMs perform better in dense geometry and camera motion prediction. Combining the features from both model families showed improvements in both geometry and semantics, indicating a promising direction for enhancing spatial-intelligence models.

Paper link: https://huggingface.co/papers/2605.28132

9. ESPO: Early-Stopping Proximal Policy Optimization

Keywords: Early-Stopping Proximal Policy Optimization, mathematical reasoning, trajectory failure, reinforcement learning, surrogate regret

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance mathematical reasoning in large language models by implementing a method to detect and terminate failed trajectories early, thereby improving performance and reducing computational waste.

Research Methods:

– The research introduces a method called ESPO (Early-Stopping Proximal Policy Optimization), which halts unsuccessful trajectories early, utilizing a surrogate regret computed from logits, and treats truncated trajectories as absorbing failure states with terminal reward.

Research Conclusions:

– ESPO demonstrates superior performance compared to PPO on benchmarks such as AIME, AMC, and MATH-500, while achieving a computational efficiency of over 20% in saving rollout tokens.

Paper link: https://huggingface.co/papers/2605.29860

10. MCP-Persona: Benchmarking LLM Agents on Real-World Personal Applications via Environment Simulation

Keywords: personalized tools, MCP-Persona, agent performance, social media platforms, state-of-the-art agents

Category: Natural Language Processing

Research Objective:

– The study aims to evaluate agent performance on personalized tools interacting with individual accounts and local databases through the MCP-Persona benchmark.

Research Methods:

– Implementation of the MCP-Persona benchmark focused on real-world personalized MCP tools across diverse applications, including social media and enterprise collaboration suites.

Research Conclusions:

– Findings reveal significant challenges faced by current state-of-the-art agents in personalized tool use, emphasizing the importance of MCP-Persona in identifying and addressing these limitations.

Paper link: https://huggingface.co/papers/2606.02470

11. Joint Agent Memory and Exploration Learning via Novelty Signals

Keywords: Joint Agent Memory and Exploration Learning (JAMEL), novelty-driven interaction, autonomous agents, exploration policy, latent memory

Category: Reinforcement Learning

Research Objective:

– The research introduces the Joint Agent Memory and Exploration Learning (JAMEL) framework to enhance exploration capabilities in open-ended environments while reducing computational costs.

Research Methods:

– JAMEL trains memory and exploration policies simultaneously using novelty-driven interactions and deterministic novelty signals like code coverage for supervision.

Research Conclusions:

– JAMEL successfully generalizes to new environments, outperforms open-weight baselines in exploration depth, and competes with closed-source models, achieving efficiency in token consumption.

Paper link: https://huggingface.co/papers/2606.01528

12. Speculative Pipeline Decoding: Higher-Accruacy and Zero-Bubble Speculation via Pipeline Parallelism

Keywords: Speculative Pipeline Decoding, pipeline parallelism, decoding latency, speculative decoding, theoretical speedup

Category: Natural Language Processing

Research Objective:

– The study aims to introduce a novel framework, Speculative Pipeline Decoding, to accelerate large language model inference by utilizing pipeline parallelism for parallel token processing and reducing decoding latency.

Research Methods:

– Partitioning the target large language model (LLM) into multiple pipeline stages allows parallel token processing. A speculation module aggregates intermediate features at various pipeline depths to predict the next token, ensuring bounded difficulty and high acceptance rates while eliminating latency bubbles.

Research Conclusions:

– The Speculative Pipeline Decoding framework demonstrates significantly higher theoretical speedup compared to mainstream baselines, offering a highly scalable solution for LLM decoding acceleration.

Paper link: https://huggingface.co/papers/2605.30852

13. Skill is Not One-Size-Fits-All: Model-Aware Skill Alignment for LLM Agents

Keywords: MASA, Hierarchical Skill Evolution, Model-Agnostic, Skill Effectiveness, Inference Cost

Category: Reinforcement Learning

Research Objective:

– To propose MASA (Model-Aware Skill Alignment), a framework for adapting skills to various backbones without altering agent weights to enhance performance in long-horizon interactive tasks.

Research Methods:

– Developed a two-stage process involving hierarchical skill evolution and a lightweight model-conditioned skill rewriter for fast adaptation.

– Utilized hill climbing and UCB-driven tree search guided by environment feedback and model capability profiles.

Research Conclusions:

– MASA achieves superior performance across tested environments and backbones, with significant gains over existing models.

– The framework allows the rewriter to generalize to unseen tasks efficiently, outperforming larger models with reduced inference cost.

Paper link: https://huggingface.co/papers/2605.30723

14. Brain-IT-VQA: From Brain Signals to Answers

Keywords: Brain-IT, VQA, fMRI, NSD-VQA, Transformer

Category: Multi-Modal Learning

Research Objective:

– Introduce the Brain-IT-VQA framework to decode visual content from fMRI signals and improve visual question answering.

Research Methods:

– Utilize a transformer-based architecture to decode language tokens from brain activity and integrate them with a language model.

Research Conclusions:

– The Brain-IT-VQA model significantly outperforms previous fMRI-based captioning and VQA approaches. The NSD-VQA dataset provides a new benchmark for reliable and interpretable evaluation, enabling the study of various visual and semantic information decodable from fMRI responses.

Paper link: https://huggingface.co/papers/2605.29588

15. AFUN: Towards an Affordance Foundation Model for Functionality Understanding

Keywords: Affordance understanding, RGB-D observation, AI Native

Category: Robotics and Autonomous Systems

Research Objective:

– To develop an affordance understanding model that predicts functional masks and 3D motion curves from RGB-D observations and language descriptions, enabling generalizable robot manipulation across diverse environments.

Research Methods:

– Utilization of a large-scale standardized data pipeline that converts heterogeneous robot, human, simulation, and real-world scan data into a unified affordance schema with language, masks, and object-centric 3D motion labels.

Research Conclusions:

– The model outperforms baseline methods in affordance segmentation, contact-point prediction, and 3D motion testing, demonstrating adaptability to real-world affordance tasks without the need for finetuning.

Paper link: https://huggingface.co/papers/2606.02551

16. PARCEL: Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding

Keywords: Vision-Language Model, Feature Extraction, Visual-Token Budgets, Elastic Queries, Spatial Grounding

Category: Computer Vision

Research Objective:

– Introduce PARCEL, a vision-language model architecture that dynamically partitions feature extraction to enhance efficiency and performance across various visual-token budgets.

Research Methods:

– Implementation of a visual tokenization architecture, using Pool-Anchored Resampling with Conditioned Elastic Queries, to address the computational challenges faced by Large Vision-Language Models (LVLMs).

Research Conclusions:

– PARCEL consistently outperforms existing baselines across 27 benchmarks, improving the performance-efficiency Pareto frontier and maintaining the “train once, deploy anywhere” paradigm.

Paper link: https://huggingface.co/papers/2605.30126

17. RoboStressBench: Benchmarking VLM Robustness to Physical Visual Stress in Embodied Scenes

Keywords: RoboStressBench, Vision-Language Models, visual stress, embodied AI, visual perception

Category: Multi-Modal Learning

Research Objective:

– The study aims to evaluate the robustness of Vision-Language Models (VLMs) to physical visual stress within embodied AI systems by introducing a principled benchmark called RoboStressBench.

Research Methods:

– The research decomposes visual stress into four dimensions: Material, Viewpoint, Lighting, and Geometry, inspired by the physical rendering equation. Comprehensive evaluations of state-of-the-art VLMs are conducted to identify stress-specific failure modes.

Research Conclusions:

– The study finds that different physical factors degrade different embodied capabilities. The introduced RoboStressBench provides a robust evaluation framework for diagnosing and improving VLM perception under real-world physical stress, enhancing the reliability of embodied AI systems.

Paper link: https://huggingface.co/papers/2606.00828

18. Multi-Agent Computer Use

Keywords: Multi-agent computer use, Task decomposition, Parallel execution, Directed acyclic graph, Agent coordination

Category: AI Systems and Tools

Research Objective:

– To evaluate and build multi-agent computer use (MACU) systems that improve upon single-agent approaches in handling complex, long-horizon tasks.

Research Methods:

– Proposing a general multi-agent setup where a manager model decomposes tasks into a directed acyclic graph (DAG) for parallel subagent execution, allowing dynamic task decomposition and consistent re-planning.

Research Conclusions:

– The MACU system outperforms single-agent baselines by 3.4-25.5% across different benchmarks, demonstrating enhanced test-time scaling and efficiency in complex long-horizon tasks. These improvements highlight multi-agent coordination as a promising approach for scaling computer use agents.

Paper link: https://huggingface.co/papers/2606.01533

19. HakushoBench: A Japanese Chart and Table VQA Benchmark from Governmental White Papers

Keywords: HakushoBench, Vision-Language Models, Japanese, Benchmark, Complex Visual Data

Category: Multi-Modal Learning

Research Objective:

– The primary objective is to evaluate vision-language models’ ability to understand complex Japanese chart and table visual data derived from governmental documents, filling the gap in non-English datasets.

Research Methods:

– Researchers utilized governmental white papers as a scalable source to create a new benchmark, HakushoBench, incorporating 2,053 images with annotated QA pairs to test deep and holistic understanding of visual data.

Research Conclusions:

– Experiments revealed that current open-weight models struggle with the HakushoBench, achieving a maximum accuracy of 58.6%, indicating significant room for improvement in understanding complex chart and table data.

Paper link: https://huggingface.co/papers/2606.01132

20. SOCO: Benchmarking Semantic Object Correspondence in Vision Foundation Models

Keywords: Semantic Object Correspondence, Vision Foundation Models, Keypoint Annotations, Vision-Language Models, Downstream Tasks

Category: Computer Vision

Research Objective:

– To evaluate structured object understanding in vision models by using the Semantic Object Correspondence (SOCO) benchmark.

Research Methods:

– Introducing the SOCO benchmark with taxonomy and consistent keypoint annotations across 100 categories to assess semantic correspondence.

Research Conclusions:

– Vision foundation models encode strong semantic structures but have limitations in correspondence transfer and object-part capturing.

– Vision-language models excel in text-prompted part localization but struggle with cross-image visual matching.

– Correspondence performance correlates more strongly with dense downstream tasks than with ImageNet classification.

Paper link: https://huggingface.co/papers/2605.31597

21. ACL-Verbatim: hallucination-free question answering for research

Keywords: VerbatimRAG, ModernBERT, extractive question answering, large language models, hallucinations

Category: Natural Language Processing

Research Objective:

– To develop a VerbatimRAG-based extractive question answering system that improves accurate information retrieval from research papers using a novel ground truth dataset and the ModernBERT model.

Research Methods:

– Application of the extractive question answering system to ACL Anthology research papers.

– Creation of a novel ground truth dataset for mapping user queries to relevant text spans in research papers and training extractive models using human-annotated synthetic queries.

Research Conclusions:

– The ModernBERT token classifier, with 150M parameters and trained with silver supervision from synthesized queries, achieves superior word-level F1 scores (53.6) compared to the leading evaluated LLM extractor (48.7).

Paper link: https://huggingface.co/papers/2605.21102

22. LongAttnComp: Cross-Family Context Compression for Long-Context Reasoning

Keywords: LongAttnComp, context compression, token-level chunking, positional reordering, two-stage fine-tuning

Category: Natural Language Processing

Research Objective:

– Adapt LongAttnComp for long-context processing to enhance context length and inference efficiency, addressing the bottleneck in real-world applications requiring 100k+ tokens.

Research Methods:

– Implement lightweight cross-attention scoring layer fine-tuning, token-level chunking, token-budget top-p algorithm, positional reordering, and format-agnostic query parser.

– Deploy a two-stage fine-tuning recipe using NIAH-style data and extending with multi-hop and reasoning data.

Research Conclusions:

– LongAttnComp matches or exceeds full-context accuracy on InfiniteBench Code-Debug, surpasses training-free baselines, and successfully transfers across various models.

– The two-stage recipe effectively closes the Stage 1 gap in multi-document reasoning while maintaining Code-Debug performance.

Paper link: https://huggingface.co/papers/2606.01336

23. Silent Failures in Physical AI: A Literature Review of Runtime Action Authorization for Autonomous Systems

Keywords: Physical AI systems, Runtime Assurance, Safety Mechanisms, Black-box Models, Robotics

Category: Robotics and Autonomous Systems

Research Objective:

– To address the safety challenges in Physical AI systems, focusing on the development of comprehensive runtime guardrail mechanisms to ensure safe operations amidst black-box models.

Research Methods:

– The study involves a synthesis of various streams like embodied foundation models, world models, and robotics simulation to formulate a bounded problem space for runtime authorization and safety verification.

Research Conclusions:

– The analysis reveals a significant gap in existing models’ ability to provide complete runtime safety assurance, proposing a taxonomy of runtime guardrail functions and evaluation requirements for Physical AI assurance mechanisms.

Paper link: https://huggingface.co/papers/2606.00090

24. Adapting Multilingual Embedding Models to Turkish via Cross-Lingual Tokenizer Surgery and Offline Distillation

Keywords: Turkish-focused sentence embedding, L2-normalized vectors, transformer backbone, embedding distillation, cosine similarity objective

Category: Natural Language Processing

Research Objective:

– Develop a Turkish-focused sentence embedding model called embeddingmagibu-200m with higher performance and reduced computational costs compared to larger models.

Research Methods:

– Implement a three-stage adaptation pipeline: (1) Create a Turkish-optimized multilingual tokenizer with a pruned vocabulary, (2) Clone a teacher embedding model maintaining the transformer backbone, and (3) Conduct offline embedding distillation from precomputed vectors using a cosine similarity objective.

Research Conclusions:

– The student model, with approximately 200M parameters, exhibits superior performance, with Pearson/Spearman correlations surpassing the teacher model, and achieves a competitive mean score on TR-MTEB tasks while utilizing 33% fewer parameters.

Paper link: https://huggingface.co/papers/2605.29992

25. TVIR: Building Deep Research Agents Towards Text–Visual Interleaved Report Generation

Keywords: multimodal deep research, TVIR, hierarchical multi-agent framework, Textual Assessment, Visual Assessment

Category: Multi-Modal Learning

Research Objective:

– The main goal is to evaluate and improve the factual reliability and visual alignment of automated report generation systems through a new multimodal benchmark and agent framework.

Research Methods:

– Introduction of TVIR (Text–Visual Interleaved Report Generation), including a benchmark (TVIR-Bench) with 100 expert-curated tasks requiring visual elements and a hierarchical multi-agent framework named TVIR-Agent for effective report generation.

– Implementation of a dual-path evaluation framework combining both Textual Assessment and Visual Assessment.

Research Conclusions:

– TVIR-Agent demonstrates strong overall performance across nine deep research systems, emphasizing the necessity of explicit multimodal design and evaluation in evidence-driven report generation.

Paper link: https://huggingface.co/papers/2606.02320

26. MindZero: Learning Online Mental Reasoning With Zero Annotations

Keywords: MindZero, self-supervised reinforcement learning, Theory of Mind, mental state hypotheses, multimodal large language models

Category: Reinforcement Learning

Research Objective:

– The primary aim is to develop MindZero, a self-supervised reinforcement learning framework, enabling multimodal large language models to efficiently perform online mental reasoning without needing explicit mental state annotations.

Research Methods:

– The framework trains models to generate mental state hypotheses that maximize the likelihood of observed actions using a planner, thereby internalizing model-based reasoning into fast single-pass inference.

Research Conclusions:

– MindZero significantly enhances multimodal large language models’ Theory of Mind capabilities, outperforming traditional model-based methods in accuracy and efficiency in mental reasoning tasks across gridworld and household domains.

Paper link: https://huggingface.co/papers/2606.00240

27. Compositional Text-to-Image Generation Via Region-aware Bimodal Direct Preference Optimization

Keywords: BiDPO, text-to-image models, preference-based fine-tuning, compositional fidelity, region-level guidance

Category: Generative Models

Research Objective:

– To enhance text-to-image models’ capability of generating images from complex compositional prompts through the proposed BiDPO framework.

Research Methods:

– Implementing a pipeline to create a large-scale preference dataset called BiComp with strict quality control.

– Extending Diffusion DPO to jointly optimize image and text preferences.

– Employing region-level guidance to focus on relevant compositional regions.

Research Conclusions:

– BiDPO significantly improves compositional fidelity, outperforming previous methods across multiple benchmarks.

– Demonstrates the potential of preference-based fine-tuning as a flexible and scalable approach for complex text-to-image tasks.

Paper link: https://huggingface.co/papers/2605.28615

28. Can Predicted Dynamics Exist in the Physical World?

Keywords: Physical Admissibility, Prediction-Control Interface, Receiver Operating Characteristic Curve, RMSE, AI Systems

Category: AI Systems and Tools

Research Objective:

– The study aims to establish a prediction-control interface for ensuring physical admissibility in AI systems by filtering out invalid proposals while maintaining high system performance.

Research Methods:

– The methodology involves evaluating decoded proposals using kinematic, dynamic, and direct-to-composed horizon conditions before execution to determine physical admissibility.

Research Conclusions:

– The research demonstrates that controlled falsification achieves high AUC values, showcasing the effectiveness of the physical admissibility gate in preventing invalid proposals while preserving significant mean progress.

Paper link: https://huggingface.co/papers/2606.00089

29. AI, Take the Wheel: What Drives Delegation and Trust in Human-Computer Cooperative Question Answering?

Keywords: Human-AI collaboration, trust, delegation choice, adoption choice, confirmation bias

Category: Human-AI Interaction

Research Objective:

– To understand when, why, and how humans decide to rely on AI in question-answering tasks.

Research Methods:

– Evaluation of 24 human-AI collaborative matches in a question-answering game with 387 delegation and 1440 adoption decisions recorded.

Research Conclusions:

– Human-AI collaboration can yield better performance than either party alone, but suboptimal decisions occur due to under-reliance on correct AI suggestions and over-reliance on misleading AI outputs. Recommendations include utilizing calibrated confidence, evidence-grounded explanations, and mechanisms to refine user trust.

Paper link: https://huggingface.co/papers/2605.28255

30. The Chain Holds, the Answer Folds: Trace-Answer Dissociation in Reasoning Models Under Adversarial Pressure

Keywords: reasoning models, chain-of-thought, unfaithful capitulation, adversarial conditions, multi-turn dialogue

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective is to identify and study a new failure mode in reasoning models termed unfaithful capitulation, where correct reasoning chains flip to incorrect answers under adversarial conditions.

Research Methods:

– Controlled experiments were conducted across multiple datasets and models using a 2×2 latent-versus-behavioral framework to isolate the failure mode that traditional metrics miss. Data includes three datasets—MT-Consistency, MMLU-Pro, GSM8K.

Research Conclusions:

– Findings show that reasoning models can maintain factual correctness in reasoning chains but still produce incorrect final answers due to adversarial pressures. The identified failure mode is corroborated by independent GPT-4o judging. The effect varies across different models and datasets.

Paper link: https://huggingface.co/papers/2605.29087

31. FreeForm: Reduced-Order Deformable Simulation from Particle-Based Skinning Eigenmodes

Keywords: Reduced-order simulation, Reproducing Kernel Particle Method, deformable hyperelastic objects, neural fields, AI Native

Category: Robotics and Autonomous Systems

Research Objective:

– The research aims to introduce a novel formulation for the mesh-free, reduced-order simulation of deformable hyperelastic objects, overcoming existing challenges with meshes and neural fields.

Research Methods:

– Employs the Reproducing Kernel Particle Method (RKPM) to construct reduced-order skinning weights by solving a generalized eigensystem on the elastic energy’s Hessian matrix for faster and more accurate simulations.

Research Conclusions:

– The proposed method achieves a 40x training speedup compared to neural fields and lower simulation error, effectively supporting various geometric representations and robot simulation applications.

Paper link: https://huggingface.co/papers/2605.29318

32. ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats

Keywords: Bilingual Benchmark, Chart Parsing, Human-Agent Collaborative Annotation, Canonical Semantic Spaces, Structure-Aware Metrics

Category: Computer Vision

Research Objective:

– To introduce ChartArena, a comprehensive bilingual benchmark for evaluating chart parsing models across various chart types and visual conditions, enabling fair comparison using a unified evaluation framework.

Research Methods:

– ChartArena covers eight chart families and evaluates them in three visual scenarios: digital, printed, and hand-drawn photos.

– Utilizes a human-agent collaborative annotation pipeline with multi-stage human verification to ensure annotation reliability.

– Employs a format-agnostic evaluation protocol mapping heterogeneous outputs into two canonical semantic spaces with structure-aware metrics.

Research Conclusions:

– Proprietary models like Gemini 3.1 Pro currently lead, but open-source systems are rapidly improving.

– Document parsing models perform well on numeric charts but struggle with diagrammatic structures.

– Chart parsers are limited to narrow chart families; radar charts and hand-drawn scenarios remain especially challenging.

– ChartArena reveals capability gaps and provides a unified foundation for future progress.

Paper link: https://huggingface.co/papers/2606.01348

33. Thinking in Blender: Staged Executable Inverse Graphics with Vision-Language Models

Keywords: Inverse graphics, Pretrained vision-language models, Blender program, Staged reconstruction, Task decomposition

Category: Computer Vision

Research Objective:

– The study investigates whether pretrained vision-language models can directly perform executable inverse graphics from a single image by reconstructing it into an editable Blender program without relying on other specialized models or techniques.

Research Methods:

– Introduction of Staged Executable Inverse Graphics (SEIG), a framework that progressively refines scene factors like geometry, materials, composition, and lighting directly in executable Blender code space.

Research Conclusions:

– Staged reconstruction significantly improves fidelity, emphasizing the importance of task decomposition for enhancing reconstruction fidelity with general-purpose vision-language models.

Paper link: https://huggingface.co/papers/2606.02580

34. τ_0-WM: A Unified Video-Action World Model for Robotic Manipulation

Keywords: unified video-action world model, robotic manipulation, video prediction, policy learning, action evaluation

Category: Robotics and Autonomous Systems

Research Objective:

– The study aims to integrate policy learning, video prediction, and action evaluation into a unified framework for improving robotic manipulation tasks.

Research Methods:

– Utilizes a shared video diffusion backbone and a model trained on 27,300 hours of diverse interaction data to predict and simulate future actions.

Research Conclusions:

– The τ_0-World Model demonstrates superior performance in challenging long-horizon and fine-grained robotic manipulation tasks compared to existing baselines.

Paper link: https://huggingface.co/papers/2606.01027

35. Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures

Keywords: Semantic Motion Anchors, Co-Speech Gesture Retrieval, Communicative Intent, 3D Gestures, Motion Primitives

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the alignment between spoken text and gesture representations, thereby enhancing retrieval accuracy and semantic relevance through the use of Semantic Motion Anchors.

Research Methods:

– The method involves discretizing 3D gestures into body-hand motion primitives, verbalizing them into structured descriptions, and grounding them in the transcript to provide auxiliary contrastive supervision, improving text-to-gesture retrieval metrics.

Research Conclusions:

– The proposed approach improves text-to-gesture retrieval R@1 by 8.2% over direct text-motion baselines and is preferred significantly in gesture generation tasks, demonstrating its efficacy in conveying communicative intent more clearly.

Paper link: https://huggingface.co/papers/2605.30608

36. Lost in Translation? Exploring the Shift in Grammatical Gender from Latin to Occitan

Keywords: Deep Learning Framework, Grammatical Gender, Historical Setting, Lexical Features, Contextual Features

Category: Natural Language Processing

Research Objective:

– Develop a deep learning framework to analyze the evolution of the grammatical gender system from Latin to Romance languages.

Research Methods:

– Introduce an interpretable deep learning framework to study both lexical and contextual factors.

– Evaluate and improve tokenizer performance in a low-resource historical setting.

– Analyze morphological features and part-of-speech categories for gender prediction.

Research Conclusions:

– Conventional tokenization strategies are insufficient for historical data, while the proposed tokenizer improves performance.

– The study provides insights into the distribution of gender information across lexical and sentential contexts.

– Publicly share code, datasets, and results for further research and validation.

Paper link: https://huggingface.co/papers/2605.09156

37.

Paper link: 

38. Who Annotates in NLP? A Large-scale Assessment of Human Annotation Reporting between 2018 and 2025

Keywords: Human annotation, NLP research, LLM-assisted extraction, annotation validity, annotation reporting

Category: Natural Language Processing

Research Objective:

– To conduct a large-scale audit of human annotation reporting practices in NLP and assess how well critical annotation details are documented over time.

Research Methods:

– An audit of annotation reporting using a unified taxonomy and an LLM-assisted extraction pipeline validated against a human-adjudicated gold standard.

Research Conclusions:

– Despite improvements in annotation reporting over time, gaps still exist in reproducibility and reliability. A scalable framework and minimum reporting recommendations are established to enhance human annotation practices.

Paper link: https://huggingface.co/papers/2606.02255

39. DOT-MoE: Differentiable Optimal Transport for MoEfication

Keywords: Differentiable Optimal Transport, Mixture of Experts, Large Language Models, Sparse MoE Models, Straight-Through Estimators

Category: Natural Language Processing

Research Objective:

– The objective is to address the inefficiencies in inference created by the scaling of Large Language Models (LLMs) and to propose DOT-MoE, a framework for efficient training of sparse Mixture of Experts (MoE) models through differentiable optimal transport.

Research Methods:

– The paper formulates dense layer decomposition as a Differentiable Optimal Transport (DOT) problem, utilizing balanced transport via Sinkhorn-Knopp iterations and Straight-Through Estimators (STE).

Research Conclusions:

– DOT-MoE significantly outperforms baseline methods such as structured pruning and heuristic clustering, retaining 90% performance of the dense model while reducing active parameters by 50%.

Paper link: https://huggingface.co/papers/2606.01666

40. The Hamilton-Jacobi Theory of Deep Learning

Keywords: Neural Networks, Hamilton–Jacobi, Residual Networks, Transformers, Viscous PDE

Category: Foundations of AI

Research Objective:

– The paper aims to formulate neural network training as a Hamilton–Jacobi initial-value problem, establishing exact connections to various neural network architectures.

Research Methods:

– It employs an analytical framework where gradient steps align with solving viscous Hamilton–Jacobi equations, creating links with residual networks, transformers, and recurrent networks.

Research Conclusions:

– The study reveals significant quantitative outcomes such as minimax optimal generalization rates, adversarial robustness, and interpretable scaling exponents, underlining the unifying role of a deformation parameter that harmonizes multiple theoretical perspectives.

Paper link: https://huggingface.co/papers/2605.28983

41. Review Arcade: On the Human Alignment and Gameability of LLM Reviews

Keywords: LLM-generated reviews, human reviews, ACL Rolling Review, paper scores, iterative revision

Category: Natural Language Processing

Research Objective:

– The research aims to evaluate large language model (LLM)-generated reviews from both the author and reviewer perspectives, focusing on their alignment with human reviews and the effectiveness of authors using LLMs to revise paper drafts iteratively.

Research Methods:

– Empirical experiments were conducted on papers from the 2025 ACL Rolling Review (ARR) to assess the alignment between LLM and human reviews and the impact of iterative draft-revision workflows.

Research Conclusions:

– Limited alignment was found between LLM and human reviews, with performance varying by prompts and models.

– Authors can use LLM feedback effectively to “game” the review process, potentially increasing paper scores by up to 35% in certain scenarios.

Paper link: https://huggingface.co/papers/2605.28897

42. Same Question, Different Source, Different Answer: Auditing Source-Dependence in Medical Multi-Source RAG

Keywords: Retrieval-Augmented Generation, Source-dependence, NLP Evaluation, Multi-source NLP

Category: Natural Language Processing

Research Objective:

– The paper aims to highlight the significance of analyzing inter-source relationships in multi-source NLP systems rather than focusing solely on answer correctness.

Research Methods:

– Deployment of a Retrieval-Augmented Generation system over an institutional corpus to evaluate the role of source-dependence in response generation.

– Creation of benchmarks such as TransplantQA and HERO-QA to assess system performance across different institutional sources.

– Utilization of a structured-output judge to score inter-source relationships using a validated taxonomy.

Research Conclusions:

– The research reveals that better retrieval techniques can uncover more significant disagreements between sources than previously recognized, implying that understanding source-dependence is crucial for deployed multi-source NLP systems.

– The proposed framework is domain-agnostic and applicable to other fields such as legal and educational applications of retrieval-augmented generation systems.

Paper link: https://huggingface.co/papers/2605.29084

43. Model-Based Quality Assessment for Massively Multilingual Parallel Data

Keywords: Multilingual Parallel-Data, Parallelism Assessment, Quality Estimation, Language Pairs, Embedding Models

Category: Natural Language Processing

Research Objective:

– To assess multilingual parallel-data using direction-specific approaches instead of universal metrics due to variability across language pairs.

Research Methods:

– Decomposing model-based assessment into parallelism assessment with multilingual embeddings and reference-free quality estimation (QE).

– Benchmarking embedding models on FLORES-200 and BOUQuET retrieval tasks covering 6,654 source-target directions.

– Evaluating nine reference-free evaluators on professional translations across 41,412 ordered source-target directions.

Research Conclusions:

– No model is universally reliable across translation directions.

– Naive QE ensembles tend to dilute strong model signals.

– Documented target-language coverage correlates with higher QE scores, indicating the need for direction-aware routing and calibration in multilingual parallel-data assessment.

Paper link: https://huggingface.co/papers/2606.00285

44. A Formally Verified Library of Mathematical Finance in Lean 4

Keywords: Lean 4 proof assistant, mathematical finance, machine-checked, risk-neutral pricing measure

Category: AI in Finance

Research Objective:

– Develop a broad library of mathematical finance within the Lean 4 proof assistant, incorporating over two hundred theorems without omissions.

Research Methods:

– Construct the L2 Itô integral as a bounded linear isometry; derive, rather than assume, the risk-neutral pricing measure.

– Classify and audit the faithfulness of results, ensuring transparency and accuracy in the relation of Lean statements to mathematical claims.

Research Conclusions:

– The project provides certified unification of known financial results instead of introducing new theory, contributing reusable verified foundations for mathematical finance.

Paper link: https://huggingface.co/papers/2606.01356

45. Geometric Latent Reasoning Induces Shorter Generations in LLMs

Keywords: Geometric Latent Reasoning, latent reasoning, geometric path-approximation, pretrained token-embedding space, Qwen3 models

Category: Natural Language Processing

Research Objective:

– The research aims to explore latent reasoning within a geometric path-approximation framework in pretrained token-embedding space to reduce generation length while maintaining accuracy.

Research Methods:

– Introduced Geometric Latent Reasoning (GLR) using a lightweight transition head to predict iterative direction updates in embedding space, utilizing textual chain-of-thought traces as anchors.

Research Conclusions:

– Evaluations demonstrate that geometric latent reasoning results in substantially shorter generations without explicit length objectives, suggesting a new tradeoff between latent computation budget, output length, and accuracy.

Paper link: https://huggingface.co/papers/2606.02248

46. Unified Neural Scaling Laws

Keywords: Unified Neural Scaling Law, deep neural networks, scaling behaviors, architectures, reinforcement learning

Category: Machine Learning

Research Objective:

– The research aims to introduce a Unified Neural Scaling Law that models and extrapolates deep neural network scaling behaviors across multiple simultaneous dimensions, including parameters, dataset size, training steps, and compute.

Research Methods:

– The study presents a functional form that captures how the evaluation metric varies with changes in several factors such as model parameters, dataset size, training steps, and hyperparameters for a range of architectures and tasks.

Research Conclusions:

– Compared to other functional forms, the introduced law provides considerably more accurate extrapolations of scaling behavior in diverse tasks including vision, language, math, and reinforcement learning.

Paper link: https://huggingface.co/papers/2605.26248

47. StressDream: Steering Video World Models for Robust Policy Evaluation and Improvement

Keywords: Video world models, policy evaluation, Vision-Language Model, diffusion-based, autonomous driving

Category: Generative Models

Research Objective:

– To enhance video world models by steering diffusion-based imaginations toward high-impact yet plausible outcomes for improved policy evaluation and improvement.

Research Methods:

– Utilization of optimized noise initialization with semantic and plausibility objectives to guide video imaginations.

– Implementation of a semantic objective via a Vision-Language Model to provide informative gradients about the generated video.

– Introduction of a plausibility objective to prevent out-of-distribution noise from resulting in implausible imaginations.

Research Conclusions:

– StressDream successfully directs video world model imaginations towards outcomes that are both high-impact and plausible, aiding in robust policy evaluation, particularly in scenarios like autonomous driving and robotic manipulation.

Paper link: https://huggingface.co/papers/2606.00267

48. Confidence-Adaptive SwiGLU for Mixture-of-Experts

Keywords: SwiGLU, Mixture-of-Experts, token-level routing confidence, Transformer MLPs, Confidence-Aware

Category: Machine Learning

Research Objective:

– The study aims to enhance Mixture-of-Experts models by developing Confidence-Aware SwiGLU, which dynamically adjusts expert gate sharpness based on token-level routing confidence.

Research Methods:

– The authors propose κ-SwiGLU, a variant of SwiGLU, which parameterizes gate sharpness as a learnable function dependent on router logit values, tested across MoE Transformer models with varying layers.

Research Conclusions:

– κ-SwiGLU shows improved mean CORE performance with minimal computational cost, suggesting that confidence-aware gate sharpness offers a promising improvement for MoE MLPs.

Paper link: https://huggingface.co/papers/2606.00761

49. EVA01: Unified Native 3D Understanding and Generation via Mixture-of-Transformers

Keywords: EVA01, 3D mesh integration, Multimodal Large Language Models, Mixture-of-Transformers, text-to-3D generation

Category: Multi-Modal Learning

Research Objective:

– The paper introduces EVA01, which enables native integration of 3D meshes into Multimodal Large Language Models (MLLMs) to enhance generation and editing capabilities.

Research Methods:

– EVA01 is constructed using a Mixture-of-Transformers architecture, separating the model into an Understanding Expert and a Generation Expert interconnected by shared global self-attention with hard modality routing.

Research Conclusions:

– EVA01 exhibits state-of-the-art performance in native text-to-3D generation fidelity and supports robust long-context, multi-turn geometric editing with identity preservation, offering significant advancements over traditional stateless reconstruction pipelines.

Paper link: https://huggingface.co/papers/2605.16745

50. SVI-Bench: A Dynamic Microworld for Strategic Video Intelligence

Keywords: Strategic Video Intelligence, Causal Reasoning, Strategic Planning, Multi-Agent Systems, Agentic Baselines

Category: Knowledge Representation and Reasoning

Research Objective:

– To address gaps in evaluating Strategic Video Intelligence by introducing a comprehensive benchmark named SVI-Bench, aimed at understanding complex cognitive tasks.

Research Methods:

– Developing a large-scale benchmark using team sports as dynamic microworlds, which includes 35K hours of video, 15M annotated actions, and other structured data, organized into a four-pillar task hierarchy.

Research Conclusions:

– Current models perform well in perceptual tasks but struggle significantly with higher cognitive levels, with agentic tasks being the most challenging, achieving just 5% accuracy in some cases.

Paper link: https://huggingface.co/papers/2605.31529

51. 3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code

Keywords: Vision-language models, Procedural 3D modeling, 3DCodeBench, Test-time scaling

Category: Generative Models

Research Objective:

– To evaluate vision-language models (VLMs) for their capacity to translate text and images into executable 3D code, using the 3DCodeBench benchmark.

Research Methods:

– Introduced 3DCodeBench to evaluate 12 VLMs on procedural 3D modeling tasks.

– Developed 3DCodeArena platform for human preference-based ranking of generated 3D outputs.

Research Conclusions:

– Major challenges identified include API mismatches and issues with disconnected 3D components.

– Highlighted the importance of high-quality procedural coding data and a robust execution environment for effective procedural 3D modeling.

Paper link: https://huggingface.co/papers/2606.01057

52. Not only where, But when: Temporal Scheduling for RLVR

Keywords: Reinforcement Learning, Verifiable Rewards, Temporal Scheduling, Policy Optimization, Credit Allocation

Category: Reinforcement Learning

Research Objective:

– The paper aims to enhance policy evolution and learning stability in reinforcement learning by integrating temporal scheduling with credit allocation criteria focused on verifiable rewards.

Research Methods:

– The study utilizes a novel approach of scheduling the credit allocation criteria temporally, prioritizing targeted tokens and gradually shifting towards general optimization to improve reinforcement learning dynamics.

Research Conclusions:

– Temporal scheduling leads to more stable and efficient learning dynamics, presenting a promising optimization dimension that better accommodates heterogeneous policy behaviors while improving policy evolution, as demonstrated through experiments on mathematical and general reasoning benchmarks.

Paper link: https://huggingface.co/papers/2605.25381

53. FineVerify: Scaling Test-Time Compute with Fine-Grained Self-Verification for Agentic Search

Keywords: Fine-grained self-verification, Agentic search, Language model agents, Checkable sub-questions

Category: Natural Language Processing

Research Objective:

– The objective of FineVerify is to improve accuracy in agentic search through decomposed sub-question checking and trajectory selection.

Research Methods:

– FineVerify integrates a fine-grained self-verification framework that decomposes each question into checkable sub-questions, verifies sampled candidates against each sub-question, and selects the candidate with the highest aggregated score.

Research Conclusions:

– FineVerify outperforms standard scaling baselines across four agentic search benchmarks and two models, considerably improving accuracy points.

– It not only enhances model accuracy but also provides interpretable verification traces for auditing errors in agentic search systems.

Paper link: https://huggingface.co/papers/2606.00660

54. Measuring the Depth of LLM Unlearning via Activation Patching

Keywords: Unlearning Depth Score, Large Language Model, AI Safety, Privacy Protection, Causal Approach

Category: Natural Language Processing

Research Objective:

– Introduce Unlearning Depth Score (UDS) to evaluate how thoroughly knowledge has been erased from large language models, addressing limitations of previous methods that overlook hidden knowledge in internal representations.

Research Methods:

– Utilizes activation patching to identify and measure erasure of target knowledge in model layers, assessing the unlearning on a 0-1 scale across different models and methods.

Research Conclusions:

– UDS demonstrated the highest faithfulness and robustness among 20 metrics, establishing itself as a reliable metric for unlearning evaluation, with guidelines provided for integration into benchmarking frameworks.

Paper link: https://huggingface.co/papers/2605.24614

55. RoboSemanticBench: Diagnosing Semantic Grounding in Action Prediction for VLA Models

Keywords: RoboSemanticBench, Vision-Language-Action models, action prediction, semantic grounding, robot fine-tuning

Category: Robotics and Autonomous Systems

Research Objective:

– The study introduces RoboSemanticBench to diagnose semantic grounding in action prediction, examining if vision-language-action models can semantically understand complex instructions to manipulate correct physical targets.

Research Methods:

– An embodied benchmark where a robot receives questions, observes candidate answer blocks, and must grasp the block corresponding to the correct answer to evaluate semantic understanding in robots.

Research Conclusions:

– There exists a gap between robots’ semantic competence and their action prediction, as VLA models often select the correct block at near-random rates despite successful object grasping.

Paper link: https://huggingface.co/papers/2606.02277

56. Off-the-Shelf LLMs as Process Scorers: Training-Free Alternative to PRMs for Mathematical Reasoning

Keywords: Chunk-Level Guided Generation, large language model, process scorer, PRM guided search, majority voting

Category: Natural Language Processing

Research Objective:

– To propose a new method, Chunk-Level Guided Generation, that improves reasoning accuracy in small model generation by using a large language model as a process scorer to select candidate chunks.

Research Methods:

– Utilizes a large off-the-shelf language model to score fixed-length candidate chunks without text generation, applying two selection rules: Likelihood-Guided Selection and Contrastive-Guided Selection.

Research Conclusions:

– Chunk-Level Guided Generation outperforms traditional methods like majority voting and is competitive with PRM guided search in several benchmarks, achieving higher accuracy in tasks like MATH and Minerva Math and producing shorter reasoning traces.

Paper link: https://huggingface.co/papers/2606.01682

57. MineExplorer: Evaluating Open-World Exploration of MLLM Agents in Minecraft

Keywords: Multimodal large language models, open-world exploration, Minecraft, multi-agent synthesis, task graphs

Category: Multi-Modal Learning

Research Objective:

– To evaluate the open-world exploration capabilities of multimodal large language models (MLLMs) using the developed MineExplorer benchmark within the context of Minecraft.

Research Methods:

– MineExplorer employs a multi-agent synthesis workflow that constructs task graphs, sandbox scenes, and milestone evaluators to design reliable instances for performance evaluation.

Research Conclusions:

– Human evaluation indicated that the multi-agent synthesis approach yields more reliable instances compared to single-agent baselines.

– MLLMs demonstrate strong performance in single-hop tasks, but their ability degrades in multi-hop tasks requiring coordination over longer trajectories.

– Larger models or enhanced thinking modes don’t always result in better performance, uncovering challenges in open-world exploration.

Paper link: https://huggingface.co/papers/2605.30931

58. Agent Skills Should Go Beyond Text: The Case for Visual Skills

Keywords: Multimodal skills, Visual-centric tasks, Reusable skills, Visual support, Spatial correspondence

Category: Multi-Modal Learning

Research Objective:

– The study aims to address the limitations of text-only skill-learning methods for visual-centric tasks by proposing a multimodal skill paradigm that integrates textual logic with visual support.

Research Methods:

– The paper introduces a system named SYSTEM, which automatically converts agent experience into reusable multimodal skills by preserving textual reasoning, spatial references, visual boundaries, and interaction patterns.

Research Conclusions:

– Experiments demonstrate that visual skills outperform text-only skills in tasks requiring spatial correspondence, visual evidence, and state-aware interaction. This supports the argument that agent skills should evolve beyond text to become multimodal for future agent development.

Paper link: https://huggingface.co/papers/2606.01414

59. Policy and World Modeling Co-Training for Language Agents

Keywords: PaW, Policy learning, World modeling, Reinforcement learning (RL), Language agent training

Category: Reinforcement Learning

Research Objective:

– To enhance language agent training by integrating policy learning and world modeling using on-policy RL rollouts without additional computational overhead.

Research Methods:

– Introduced a co-training framework called PaW that integrates auxiliary world modeling supervision with policy learning during RL.

– Utilized action-entropy-based WM data selection, noise-tolerant WM loss, and reward-adaptive loss balancing to ensure informative and stable WM supervision.

Research Conclusions:

– Experiments demonstrate that the PaW framework provides consistent improvements over existing strong RL baselines across various models and RL algorithms, highlighting that standard RL rollouts can effectively function as a source of world modeling supervision for training language agents.

Paper link: https://huggingface.co/papers/2606.02388

60. OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents

Keywords: OpenWebRL, visual web agents, online reinforcement learning, multi-turn RL, open-source

Category: Reinforcement Learning

Research Objective:

– Develop an open framework named OpenWebRL for training visual web agents using online reinforcement learning on real websites.

Research Methods:

– Utilize a comprehensive training pipeline including scalable live-browser infrastructure, supervised initialization, multimodal context management, trajectory-level success judging, and multi-turn policy optimization.

Research Conclusions:

– OpenWebRL-4B, utilizing minimal supervised initialization, achieves competitive success rates on challenging benchmarks, offering a new state-of-the-art for open-source visual web agents while remaining competitive with proprietary systems.

Paper link: https://huggingface.co/papers/2606.02031

61. StreamChar: Long-Horizon Streaming Character Audio-Video Generation with Decoupled Orchestration

Keywords: StreamChar, LLM-based orchestrator, joint audio-video DiT, two-stage distillation pipeline, audio-visual synchronization

Category: Generative Models

Research Objective:

– The goal of the research is to enable real-time streaming audio-video generation for character animation, ensuring transcript fidelity, visual identity maintenance, and efficient deployment.

Research Methods:

– The methods employed include separating orchestration from denoising using an LLM-based orchestrator and a joint audio-video DiT. A two-stage distillation pipeline is used for efficient deployment, while progress-aware pointers and sink-chunk memory ensure alignment and consistency.

Research Conclusions:

– Results demonstrate that StreamChar achieves real-time operation on a single H100 GPU, balancing transcript fidelity, audio-visual synchronization, and visual quality, with a superior performance in streaming stability compared to recent baselines.

Paper link: https://huggingface.co/papers/2605.25659

62. LongLive-RAG: A General Retrieval-Augmented Framework for Long Video Generation

Keywords: LongLive-RAG, retrieval-augmented generation, sliding-window attention, error accumulation, temporal coherence

Category: Generative Models

Research Objective:

– The paper aims to address the challenges in long-video generation by improving temporal coherence and quality through retrieval-augmented generation, specifically overcoming error accumulation prevalent in sliding-window attention methods.

Research Methods:

– The authors introduce LongLive-RAG, a framework that uses autoregressive video generation with a retrieval mechanism to treat previously generated latents as a dynamic, searchable history. This approach includes query embedding for retrieving relevant historical latents, thus allowing the generator to condition on non-local context.

Research Conclusions:

– LongLive-RAG effectively reduces error accumulation and improves the quality of long-video generation across various AR backbones, as demonstrated by its superior performance in VBench-Long rankings. It is noted as the first open-ended AR long video generation method to utilize self-generated latent history for content-addressable retrieval memory.

Paper link: https://huggingface.co/papers/2606.02553

63. LVSA: Training-Free Sparse Attention for Long Video Diffusion

Keywords: Sparse Attention, Video Diffusion, Structured Window Pattern, FlashInfer Kernel, VQeval

Category: Generative Models

Research Objective:

– To address computational bottlenecks in long-video diffusion models by introducing a model-agnostic block-sparse attention method that reduces computational costs while maintaining high video quality beyond training horizons.

Research Methods:

– Introduced Long Video Sparse Attention (LVSA) that combines structured window patterns with rotating global anchors to prevent fixed-grid bias and reduce compute costs up to 3.33x compared to dense attention.

– Implementation of a FlashInfer kernel and testing LVSA on various NPUs, achieving significant speedups.

Research Conclusions:

– LVSA effectively reduces computing resources while maintaining video generation quality. It also enables generation tasks that are otherwise impossible due to memory limitations. VQeval is introduced to fairly assess quality across different models.

Paper link: https://huggingface.co/papers/2605.31057

64. When Does Multi-Agent RL Improve LLM Workflows? Workflow, Scale, and Policy-Sharing Tradeoffs

Keywords: Multi-agent, Reinforcement Learning, Shared-Policy, Isolated-Policy, Gradient Dynamics

Category: Reinforcement Learning

Research Objective:

– The study investigates when multi-agent large language model workflows trained with reinforcement learning surpass their base models in accuracy.

Research Methods:

– Comparing Shared-Policy training and Isolated-Policy training across different workflows, tasks, and model scales, focusing on gradient dynamics and policy routing.

Research Conclusions:

– Multi-agent reinforcement learning typically offers improved accuracy over base models. However, the effectiveness depends on the interplay between workflow, task type, and model scale. Isolated-Policy training can achieve higher peak accuracy but is prone to terminal degradation, while Shared-Policy training presents different failure patterns tied to gradient dynamics and workflow topology.

Paper link: https://huggingface.co/papers/2605.24202

65. Masking Stale Observations Helps Search Agents — Until It Doesn’t: A Regime Map and Its Mechanism

Keywords: Observation Masking, Agentic Search, Context Management, Token-for-Turn Trade-Off, Retriever Recall

Category: Reinforcement Learning

Research Objective:

– To analyze the impact of Observation Masking in improving the accuracy of long-horizon search agents through effective Context Management.

Research Methods:

– Conducted a systematic evaluation of various agent backbones (ranging from 4B to 284B parameters) and three retrievers on offline and live-web benchmark tests.

Research Conclusions:

– The effectiveness of Observation Masking follows an asymmetric inverted-U pattern based on retriever capability and model capacity. Masking is beneficial particularly when a strong retriever is coupled with a mid-capacity model, although less effective when the model is overwhelmed or underutilized.

– The study reframes Context Management as a regime-dependent intervention. The researchers have provided a framework and released resources for future research in this area.

Paper link: https://huggingface.co/papers/2606.00408

66. VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion

Keywords: VideoMLA, video diffusion, Multi-Head Latent Attention, low-rank content, throughput

Category: Generative Models

Research Objective:

– To reduce memory usage in video diffusion models while maintaining quality and improving throughput, using a novel attention mechanism named VideoMLA.

Research Methods:

– Implemented VideoMLA with shared low-rank content and decoupled 3D-RoPE positional keys, replacing per-head keys and values to decrease memory usage by 92.7%.

Research Conclusions:

– VideoMLA effectively matches short-horizon streaming video diffusion performance and surpasses baselines at long horizons, achieving a 1.23x throughput improvement on a single B200.

Paper link: https://huggingface.co/papers/2605.30351

67. SkillAdaptor: Self-Adapting Skills for LLM Agents from Trajectories

Keywords: LLM agents, training-free skill adaptation, SkillAdaptor, failure attribution, reusable external skills

Category: Natural Language Processing

Research Objective:

– To enhance the stability and auditability of training-free skill maintenance in LLM agents by introducing a step-level skill adaptation framework with explicit failure attribution.

Research Methods:

– Developed SkillAdaptor, a framework that identifies first actionable fault steps and performs targeted updates with explicit acceptance checks on failed trajectories. Evaluated on benchmark environments such as WebShop, PinchBench, and Claw-Eval with models including Kimi-K2.5, GLM-5, and GPT-5.2.

Research Conclusions:

– SkillAdaptor showed improved performance over baseline approaches with notable single-metric gains across different evaluation suites, supporting step-level attribution as a means to achieve more stable skill maintenance.

Paper link: https://huggingface.co/papers/2606.01311

68. VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization

Keywords: Video Generation Models, Vision-Language Models, Differentiable Rewards, Test-Time Optimization, Video Reasoning

Category: Knowledge Representation and Reasoning

Research Objective:

– The paper aims to improve video reasoning performance by integrating Vision-Language Models (VLMs) as “teachers” during test-time to guide Video Generation Models (VGMs).

Research Methods:

– The approach involves using VLMs to extract task-specific rules to create differentiable rewards, which guide VGM reasoning through online test-time optimization of a LoRA module.

Research Conclusions:

– The proposed method achieves a 16.7-point average performance gain on video reasoning benchmarks, significantly outperforming both the VLM-as-Solver paradigm and Best-of-N scaling, showcasing the effectiveness of VLMs as a promising tool for generalizable video reasoning.

Paper link: https://huggingface.co/papers/2606.02564

69. Linear Ensembles Wash Away Watermarks: On the Fragility of Distributional Perturbations in LLMs

Keywords: Watermarking, AI-generated text, Statistical Hybridisation, Model Ensemble, Detection

Category: Generative Models

Research Objective:

– Investigate the vulnerability of watermarking mechanisms in AI-generated text when multiple models are employed by users.

Research Methods:

– Theoretical proof and empirical experiments demonstrating the effect of averaging outputs from multiple models on watermark detection and text quality.

Research Conclusions:

– Demonstrates that averaging 3-5 models cancels watermark perturbations, suppressing detection z-scores and reducing true positive rates, while notably improving quality and processing speed.

– Highlights a fundamental vulnerability in AI-text detection, suggesting the need for unprecedented coordination among AI model providers for robust watermarking.

Paper link: https://huggingface.co/papers/2605.30501

70. Draft-OPD: On-Policy Distillation for Speculative Draft Models

Keywords: Speculative decoding, Draft model, Supervised fine-tuning, On-policy distillation, Lossless acceleration

Category: Generative Models

Research Objective:

– The study aims to improve the acceleration of large language model inference by addressing the limitations of current speculative decoding methods.

Research Methods:

– The research investigates the on-policy distillation (OPD) method with target-assisted rollouts and error replay to enhance the effectiveness of draft models in speculative decoding. It introduces Draft-OPD for more stable continuations.

Research Conclusions:

– The proposed Draft-OPD method achieves over five times acceleration without loss across various tasks, outperforming existing models like EAGLE-3 and DFlash with improvements of 23% and 13% respectively.

Paper link: https://huggingface.co/papers/2605.29343

71. K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts

Keywords: K-BrowseComp, LLMs, Korean AI, Synthetic Split, Web-Browsing Agent

Category: Natural Language Processing

Research Objective:

– The study aims to evaluate the capabilities of frontier large language models (LLMs) in the context of Korean web-browsing tasks using a newly introduced benchmark called K-BrowseComp.

Research Methods:

– The research involves creating a benchmark consisting of 400 problems, with a verified subset constructed and validated by native Korean speakers. Additionally, a synthetic split is generated to act as a targeted stress test for the models.

Research Conclusions:

– Frontier LLMs, such as GPT-5.5, demonstrate significant performance gaps on the K-BrowseComp benchmark compared to existing English benchmarks, indicating a need for enhanced Korean AI development. Korean LLMs show particularly low performance scores.

Paper link: https://huggingface.co/papers/2606.02404

72. On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters

Keywords: Parameter-efficient fine-tuning, trainable adapters, shared foundation models, instance-specific behavior, persistent personal models

Category: Machine Learning

Research Objective:

– The study examines the role of parameter-efficient fine-tuning (PEFT) using small trainable adapters to enable persistent personal models on top of strong foundation models, moving beyond the typical view of PEFT as a cost-effective substitute for full fine-tuning.

Research Methods:

– The research explores three scaling axes: Scale Up to amplify the utility of small local updates; Scale Down to determine the minimal reliable size of adapters; and Scale Out to manage coexistence of multiple adapted instances using an infrastructure example, MinT, for handling adapter identity, revision, provenance, evaluation, and serving residency.

Research Conclusions:

– The findings suggest that PEFT serves as a compact substrate for creating and maintaining persistent personal models, demonstrating its utility beyond merely being a budget alternative to full fine-tuning.

Paper link: https://huggingface.co/papers/2606.02437

Keywords: agent safety alignment, taxonomy-guided training, lightweight agents, real-time safety moderation, AI models

Category: Robotics and Autonomous Systems

Research Objective:

– The main objective is to develop a lightweight and scalable agent safety alignment framework to address safety risks introduced by advanced AI models like OpenClaw and Codex.

Research Methods:

– The framework uses a taxonomy-guided data engine and influence-function purification to train AgentDoG 1.5 variants with minimal samples, enabling efficient deployment and real-time safety moderation.

Research Conclusions:

– AgentDoG 1.5 demonstrates state-of-the-art performance in complex interactive scenarios and significantly reduces deployment overhead, facilitating its broader real-world application.

Paper link: https://huggingface.co/papers/2605.29801

2. OmniRetrieval: Unified Retrieval across Heterogeneous Knowledge Sources

Keywords: OmniRetrieval, Knowledge Sources, Natural-Language Query, Source-Native Queries, Heterogeneous Sources

Category: Knowledge Representation and Reasoning

Research Objective:

– The paper introduces OmniRetrieval, a framework designed to handle diverse knowledge sources by identifying appropriate repositories and dispatching native queries, enhancing information retrieval across multiple dataset types.

Research Methods:

– The framework processes any natural-language query and selectively employs source-native queries dispatched to their respective execution engines. This approach contrasts with single-source retrievers and accounts for structural distinctions, tested across 13 datasets and 309 distinct knowledge bases.

Research Conclusions:

– OmniRetrieval outperforms existing single-source approaches by serving as a general-purpose interface to heterogeneous data sources, preserving the structural distinctions that make each source valuable.

Paper link: https://huggingface.co/papers/2605.29250

3. minWM: A Full-Stack Open-Source Framework for Real-Time Interactive Video World Models

Keywords: real-time interactive video world models, causality, low-latency, autoregressive training, camera control

Category: Generative Models

Research Objective:

– This paper presents minWM, a comprehensive framework aiming to transform bidirectional video diffusion models into real-time interactive world models with capabilities for control, causality, and low latency.

Research Methods:

– The framework uses a full-stack open-source approach, involving steps such as fine-tuning, autoregressive training, few-step distillation, and streaming inference to convert video models.

– Causal Forcing is employed alongside AR diffusion training, causal ODE, causal consistency distillation, and asymmetric DMD for model adaptation.

Research Conclusions:

– minWM provides an adaptable and reproducible method for developing real-time interactive video models, which is validated by integrating with different architectures and assessing features like camera trajectory quality and controllability.

Paper link: https://huggingface.co/papers/2605.30263

4. GenClaw: Code-Driven Agentic Image Generation

Keywords: agentic image generation, visual comprehension, LLMs, code-driven, generative models

Category: Generative Models

Research Objective:

– To develop GenClaw, a code-driven image generation framework that mimics human artistic processes including conceptualization, sketching, and coloring.

Research Methods:

– The integration of code (e.g., SVG, HTML, Three.js) to create executable visual sketches as an intermediate step, bridging linguistic reasoning and pixel synthesis with generative models.

Research Conclusions:

– GenClaw transforms image generation from a black-box process into a human-like staged creation system, enhancing control and interpretability in visual generation systems.

Paper link: https://huggingface.co/papers/2605.30248

5. How LoRA Remembers? A Parametric Memory Law for LLM Finetuning

Keywords: parametric memory, LoRA, power law, phase transition, MemFT

Category: Natural Language Processing

Research Objective:

– The research aims to investigate the quantitative limits of parametric memory in large language models and establish a power law relationship through the use of LoRA as a probe.

Research Methods:

– The study employs LoRA as a controlled memory capacity probe within the latent space to systematically quantify exact parametric memory and introduces a threshold-guided optimization strategy (MemFT) to improve memory performance.

Research Conclusions:

– The study introduces the Parametric Memory Law, demonstrating a robust power law linking loss reduction to effective parameters and sequence length. It reveals a deterministic phase transition at the token level and proves that MemFT can dynamically enhance memory fidelity and efficiency.

Paper link: https://huggingface.co/papers/2605.30260

6. Native Audio-Visual Alignment for Generation

Keywords: AI Native, audio-video generation, synchronization, controllability, Timbre-in-Context

Category: Multi-Modal Learning

Research Objective:

– Introduce NAVA, a Native Audio-Visual Alignment framework, to enhance synchronization and controllability in joint audio-video generation.

Research Methods:

– Utilize an Align-then-Fuse MMDiT architecture for modality-aware audio-video alignment followed by joint denoising.

– Implement Timbre-in-Context Conditioning to associate reference timbre cues with certain speech spans for controlled speech timbre.

Research Conclusions:

– Demonstrated NAVA’s superior video quality and precise audio-visual synchronization with competitive audio quality and effective timbre controllability using 6.3B parameters.

Paper link: https://huggingface.co/papers/2605.30073

7. LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training

Keywords: LaRA, data contamination, reinforcement learning, large language models, layer-wise representation analysis

Category: Reinforcement Learning

Research Objective:

– To detect data contamination in reinforcement learning post-trained large language models using LaRA, which examines geometric deviations across model layers.

Research Methods:

– Proposed a layer-wise representation analysis framework with three metrics: perturbation sensitivity, directional collapse, and local representation rigidity, under controlled perturbations.

Research Conclusions:

– LaRA’s contamination detection protocol outperforms existing output-level baselines, demonstrating that contamination leads to progressive geometric deviations across layers in RL-trained reasoning models.

Paper link: https://huggingface.co/papers/2605.29888

8. Xetrieval: Mechanistically Explaining Dense Retrieval

Keywords: dense retrieval, high-dimensional embeddings, Xetrieval, human-interpretable features, reasoning internalizer

Category: Knowledge Representation and Reasoning

Research Objective:

– The primary goal is to enhance sentence embeddings used in dense retrieval with reasoning information, decomposing them into human-interpretable sparse features for better explanation of retrieval decisions.

Research Methods:

– Introduction of the Xetrieval framework, which uses a reasoning internalizer to incorporate Chain-of-Thought reasoning directly within the embedding space, followed by decomposition into sparse features that are human-interpretable.

Research Conclusions:

– Xetrieval successfully uncovers interpretable features and provides feature-level explanations of retrieval decisions, showing coherence and strong intervention effects across various benchmarks and retrievers.

Paper link: https://huggingface.co/papers/2605.29507

9. Why Far Looks Up: Probing Spatial Representation in Vision-Language Models

Keywords: Vision-language models, spatial reasoning, representation-level analysis, perspective bias, robustness

Category: Multi-Modal Learning

Research Objective:

– The study aims to understand if vision-language models (VLMs) have a true 3D spatial understanding or if they rely on statistical shortcuts by analyzing the entangled spatial representations within these models.

Research Methods:

– A representation-level analysis framework is implemented using minimalist contrastive pairs to evaluate how spatial axes are organized and disentangled in VLM embeddings.

Research Conclusions:

– The analysis shows a consistent vertical-distance entanglement, reflecting the perspective bias of natural photos, causing accuracy gaps across benchmarks.

– Introducing a synthetic benchmark, SpatialTunnel, exposed model-intrinsic spatial shortcut biases, asserting that well-structured spatial representations enhance robustness and reliability in spatial reasoning tasks.

Paper link: https://huggingface.co/papers/2605.30161

10. Is Position Bias in Dense Retrievers Built In-or Learned from Data?

Keywords: Dense retrievers, positional bias, query-relevant information, retrieval performance, position-balanced training

Category: Natural Language Processing

Research Objective:

– To investigate how the positional distribution of evidence in training data affects retrieval-level bias in dense retrievers.

Research Methods:

– Construct synthetic position-targeted training sets and fine-tune eight architecturally diverse pretrained models under position-skewed and balanced training distributions.

Research Conclusions:

– Balanced training significantly reduces positional sensitivity by up to 87% while maintaining competitive retrieval performance, identifying training-position distribution as a major controllable factor in retrieval-level position bias.

Paper link: https://huggingface.co/papers/2605.26578

11. AsyncTool: Evaluating the Asynchronous Function Calling Capability under Multi-Task Scenarios

Keywords: LLM-based agents, asynchronous tool calling, AsyncTool, task coordination, temporal reasoning

Category: AI Systems and Tools

Research Objective:

– The study aims to evaluate LLM-based agents’ capability in asynchronous tool calling, focusing on their task coordination and temporal reasoning in environments with delayed tool feedback.

Research Methods:

– The researchers introduced a benchmark named AsyncTool, which assesses LLM-based agents in interactive multi-task tool-use scenarios with simulated tool response latency and evaluates models at different levels using efficiency-oriented metrics.

Research Conclusions:

– The findings reveal substantial challenges due to delayed tool feedback, leading to performance degradation. Efficient task coordination, dependency tracking, and state maintenance are critical for improved performance, suggesting directions for future enhancements in temporal reasoning and coordination capabilities.

Paper link: https://huggingface.co/papers/2605.27995

12. UI-KOBE: Knowledge-Oriented Behavior Exploration for Lightweight Graph-Guided GUI Agents

Keywords: GUI agents, app-specific graph knowledge, lightweight models, runtime decisions, task planning

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to enhance lightweight mobile GUI agents for improved task planning and execution efficiency using reusable app-specific graph knowledge.

Research Methods:

– Development of the UI-KOBE framework, which involves autonomously exploring a mobile app to construct an app knowledge graph representing UI states and executable transitions.

Research Conclusions:

– UI-KOBE supports runtime decision-making with graph guidance, reducing the burden on end-to-end GUI planning and enabling more efficient, interpretable, and privacy-conscious performance by lightweight models.

Paper link: https://huggingface.co/papers/2605.29534

13. PRISM: A Multi-Dimensional Benchmark for Evaluating LLM Peer Reviewers

Keywords: Automated peer review, LLMs, Argument mining, Retrieval-augmented verification, PRISM

Category: AI Systems and Tools

Research Objective:

– To evaluate the performance of LLM-based automated peer review systems against human reviewers across various dimensions of review quality.

Research Methods:

– Introduction of PRISM, a benchmarking framework that assesses review quality on dimensions such as depth of analysis, novelty assessment, flaw identification, and constructiveness using argument mining and retrieval-augmented verification.

Research Conclusions:

– LLMs can match or exceed human reviewers in individual aspects like novelty verification and priority of critique, but no single system matches human reviewers across all dimensions consistently. LLM-based systems are best used as supplements rather than standalone replacements for human reviews.

Paper link: https://huggingface.co/papers/2605.26730

14. PhyGenHOI: Physically-Aware 4D Generation of Dynamic Human-Object Interactions

Keywords: 4D Human-Object Interaction, Motion Diffusion Model, Material Point Method, AI-generated summary, 3D Gaussian representations

Category: Generative Models

Research Objective:

– To generate physically accurate and visually faithful 4D human-object interactions using AI techniques.

Research Methods:

– Combination of motion diffusion models with material point method simulations using 3D Gaussian representations.

– Introduction of PhyGenHOI framework to integrate generative human motion and physical object simulation.

Research Conclusions:

– PhyGenHOI effectively generates consistent 4D human-object interactions across varied actions, outperforming existing baselines.

Paper link: https://huggingface.co/papers/2605.30268

15. Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering

Keywords: CorVer, Wikipedia co-occurrence statistics, factual accuracy, reinforcement learning, sentence-level feedback

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance factual accuracy in knowledge-intensive question answering systems by developing a reward mechanism that improves upon traditional neural verifiers.

Research Methods:

– The CorVer framework employs a corpus-grounded reward signal based on Wikipedia co-occurrence statistics to provide precise sentence-level feedback, distinguishing correct from incorrect statements.

Research Conclusions:

– CorVer demonstrates improved performance over baselines, including four neural-verifier setups, and achieves faster training times. It consistently improves results across a diverse set of benchmark scenarios.

Paper link: https://huggingface.co/papers/2605.29648

16. REPOT: Recoverable Program-of-Thought via Checkpoint Repair

Keywords: RePoT, Program-of-Thought, deterministic verified replay, LLM call, checkpoint information

Category: AI Systems and Tools

Research Objective:

– The study aims to enhance the one-shot Program-of-Thought model by introducing RePoT, which allows for deterministic verified replay and recovery through interaction with the environment.

Research Methods:

– RePoT employs a deterministic verified replay that advances the action plan until an invalid transition occurs, then resumes with a single Large Language Model (LLM) call for continuation.

– An Adaptive RePoT approach is tested, utilizing a rule-based dispatcher for routing between suffix repair and a fresh PoT retry.

Research Conclusions:

– RePoT significantly improves success rates over the original PoT across various models and benchmarks, such as PuzzleZoo-775 and PlanBench Blocksworld.

– Performance gains are notable in certain models, with RePoT achieving peak rates and outperforming matched-budget PoT-retry baselines.

– Utilizing checkpoint information greatly enhances recovery success over error-only feedback, establishing it as a critical recovery signal.

Paper link: https://huggingface.co/papers/2605.30052

17. WorldMemArena: Evaluating Multimodal Agent Memory Through Action-World Interaction

Keywords: Multimodal large language models, long-horizon agents, Action-World Interaction Loop, WorldMemArena, harness-based memory agents

Category: Multi-Modal Learning

Research Objective:

– To formulate multimodal agent memory as an Action-World Interaction Loop and instantiate it in WorldMemArena for comprehensive analysis.

Research Methods:

– Development of WorldMemArena with 400 multi-session multimodal tasks for detailed diagnostics of memory systems through various stages.

– Comparison of long-context, manually designed memory systems and harness-based memory agents.

Research Conclusions:

– Better memory writing and storage do not necessarily lead to improved performance.

– Multimodal memory systems struggle to effectively utilize visual evidence.

– Systems show instability across different domains and degrade in agentic trajectories.

– Harness-based memory, while more flexible, is costly and less reliable.

Paper link: https://huggingface.co/papers/2605.29341

18. SmartDirector: Keyframe-Conditioned Cinematic Video Generation with Narrative Pacing Control

Keywords: SmartDirector, video generation, narrative structure, keyframes, temporal pacing

Category: Generative Models

Research Objective:

– The primary aim is to enhance the narrative capacity of video generation models using multiple keyframes to improve both narrative structure and temporal pacing.

Research Methods:

– SmartDirector employs a two-stage process comprising low-resolution generation (Director-Gen) and high-resolution refinement (Director-SR), using keyframes as semantic anchors. A data pipeline is constructed to support robust multi-keyframe training.

Research Conclusions:

– Extensive experiments indicate that SmartDirector significantly outperforms existing state-of-the-art video generation methodologies. Plans to release the code aim to facilitate further research in this area.

Paper link: https://huggingface.co/papers/2605.27891

19. AdaState: Self-Evolving Anchors for Streaming Video Generation

Keywords: Video diffusion models, AI-generated summary, Autoregressive video diffusion, Adaptive state

Category: Generative Models

Research Objective:

– The study aims to enhance video dynamics by utilizing adaptive state replacement in video diffusion models, moving beyond fixed initial frame references.

Research Methods:

– Implementing a recurrent denoising function as the transition mechanism, the model dynamically generates scene anchors at each step by referencing both previous states and current content.

Research Conclusions:

– The introduction of an adaptive state significantly improves video dynamics, allowing for more fluid motion and natural scene progression in generated videos.

Paper link: https://huggingface.co/papers/2605.30349

20. Learning A Unified Risk Map for Autonomous Driving in Partially Observable Environments

Keywords: Occlusion-aware prediction, unified risk map modeling, spatiotemporal modeling, diffusion-based scenario generation, risk-aware planning

Category: Robotics and Autonomous Systems

Research Objective:

– To address occlusion challenges in autonomous driving by integrating traffic flow and collision risks through spatiotemporal modeling and scenario generation.

Research Methods:

– Development of a unified risk map modeling and learning framework that leverages spatiotemporal modeling.

– Introduction of a diffusion-based scenario generation framework for producing realistic yet adversarial scenarios.

Research Conclusions:

– The proposed framework significantly outperforms existing baselines, offering improved minimum and average time-to-collision in tests using the Waymo Open Motion Dataset.

– Provides a comprehensive solution for risk-aware planning in partially observable environments.

Paper link: https://huggingface.co/papers/2605.22189

21. Thinking Before Constraining: A Unified Decoding Framework for Large Language Models

Keywords: In-Writing, Hybrid Approach, Free-form Reasoning, Structured Generation, Trigger Token, Classification, Reasoning, Constrained Decoding

Category: Natural Language Processing

Research Objective:

– To propose a hybrid approach called In-Writing that combines free-form reasoning with structured generation to enhance accuracy in classification and reasoning tasks.

Research Methods:

– Utilizing a newly proposed mechanism where structured decoding is only applied after a trigger token is generated, allowing for decoupling of reasoning from formatting.

Research Conclusions:

– The In-Writing approach outperforms state-of-the-art methods, achieving up to 27% accuracy gains over natural generation, effectively addressing issues with premature triggering in constrained decoding.

Paper link: https://huggingface.co/papers/2601.07525

22. NeuROK: Generative 4D Neural Object Kinematics

Keywords: 4D dynamics, latent space, Neural Object Kinematics, transformer-based encoder-decoder model, neural simulation framework

Category: Generative Models

Research Objective:

– Develop a data-driven kinematic state parameterization for dynamic object simulation, called Neural Object Kinematics (NeuROK), that utilizes a latent space and transformer-based encoding-decoding.

Research Methods:

– Leverage a large-scale 4D dataset to train a transformer-based encoder-decoder model, focusing on learning a latent space for object states and a decoder for shape deformation.

Research Conclusions:

– The proposed method significantly simplifies generating 4D dynamics by reducing the problem to a low-dimensional latent space, demonstrating effectiveness and generality across diverse dynamic objects.

Paper link: https://huggingface.co/papers/2605.30347

23. Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention

Keywords: Larger models, Model scaling, Gradient interference, Task features, Resource allocation

Category: Foundations of AI

Research Objective:

– To investigate why larger models outperform smaller ones on complex and rare tasks, even with infinite training data.

Research Methods:

– Study of model scaling effects using synthetic setups with a mixture of tasks.

– Pretraining OLMo models ranging from 4M to 4B parameters on tasks of varying frequency and complexity.

Research Conclusions:

– Larger models reduce gradient interference allowing for better task feature learning.

– Small models allocate resources poorly to rare and complex tasks, while larger models embed more task features.

– These findings provide insights into model sizing and training data mixtures for practical applications.

Paper link: https://huggingface.co/papers/2605.29548

24. ChildVox: A Speech, Audio, and Large Audio-Language Model Benchmark in Understanding and Characterizing Sound across Childhood

Keywords: ChildVox, acoustic signals, audio and speech foundation models, development stages, cross-domain comparison

Category: AI in Education

Research Objective:

– To develop ChildVox, a benchmark for analyzing children’s acoustic communication across developmental stages using diverse audio and speech models.

Research Methods:

– Integration of more than 20 sub-tasks from 17 child-centered datasets to enable systematic cross-corpus and cross-domain comparison.

– Evaluation of audio and speech models including self-supervised, ASR-oriented, and large audio-language models on various tasks.

Research Conclusions:

– ChildVox offers high-performance models for recognizing a wide range of children’s acoustic signals, aiding in characterizing language levels and tracking speech production with age.

Paper link: https://huggingface.co/papers/2605.29257

25. RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains

Keywords: RUBRIC-ARROW, reward modeling, rubric-based methods, pairwise preference data, LLM post-training

Category: Reinforcement Learning

Research Objective:

– The primary aim of RUBRIC-ARROW is to improve reward modeling by overcoming limitations of rubric-based methods, particularly focusing on reducing ties and utilizing pairwise preference data effectively.

Research Methods:

– The study employs an alternating framework that includes a rubric generator and a rubric-conditioned judge. The RL stage leverages pairwise preference data, using a probability-based scoring rule and phase-specific preference-based rewards within an alternating GRPO scheme.

Research Conclusions:

– RUBRIC-ARROW enhances reward-modeling accuracy and ensures consistent improvement for downstream policy post-training through its innovative approach.

Paper link: https://huggingface.co/papers/2605.29156

26. Towards Verifiable Multimodal Deep Research: A Multi-Agent Harness for Interleaved Report Generation

Keywords: Multi-agent system, multimodal reports, autonomous agents, Visual Working Memory, verifier agent

Category: Multi-Modal Learning

Research Objective:

– To develop a multi-agent system, Ptah, for generating reliable and visually informative multimodal reports by interleaving textual and visual evidence.

Research Methods:

– Utilization of specialized agents for constructing visual-aware plans and collecting claim-grounded evidence.

– Use of a verifier agent to ensure factual grounding and cross-modal consistency.

– Introduction of PtahEval, an evaluation protocol with image-level and presentation-level assessments.

Research Conclusions:

– Ptah produces more reliable and visually informative multimodal reports than existing strong baselines on deep research benchmarks.

Paper link: https://huggingface.co/papers/2605.29861

27. LiteCoder-Terminal: Scaling Long-Horizon Terminal Environments for Learning Language Agents

Keywords: LiteCoder-Terminal-Gen, language agents, multi-step planning, executable environments, AI-generated summary

Category: Reinforcement Learning

Research Objective:

– The study introduces LiteCoder-Terminal-Gen, aiming to enable scalable training of language agents in terminal environments through the use of synthetic and executable environments.

Research Methods:

– A zero-dependency synthesis pipeline is deployed to autonomously generate verifiable terminal training environments from domain specifications, creating resources like LiteCoder-Terminal-SFT and LiteCoder-Terminal-RL.

Research Conclusions:

– The synthetic environments created provide a scalable supervision signal, significantly improving performance in command-line workflows, with notable success in supervised fine-tuning and Direct Multi-turn Preference Optimization (DMPO).

Paper link: https://huggingface.co/papers/2605.29559

28. When Cloud Agents Meet Device Agents: Lessons from Hybrid Multi-Agent Systems

Keywords: Hybrid multi-agent systems, Large language models, Small language models, On-device inference, Task accuracy

Category: AI Systems and Tools

Research Objective:

– The study aims to systematically examine the design space of Hybrid multi-agent systems (MASs) that balance large and small language models to optimize task accuracy, cost efficiency, and energy consumption.

Research Methods:

– Two representative MAS architectures were adapted to support hybrid inference, analyzing how design choices affect the balance of power, cost, and performance along the Pareto frontier.

Research Conclusions:

– The research reveals that optimal architecture for MASs is highly task-dependent, with small language models benefiting from large model assistance. However, higher compute does not always lead to better performance.

Paper link: https://huggingface.co/papers/2605.30102

29. CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists

Keywords: CausaLab, LLM Agents, causal discovery, structural causal model, intervention

Category: Knowledge Representation and Reasoning

Research Objective:

– The paper introduces CausaLab, a scalable environment designed to evaluate interactive causal discovery by Large Language Model (LLM) agents, focusing on both accurate predictions and the faithful recovery of underlying causal mechanisms.

Research Methods:

– Agents are placed in a synthetic laboratory where they receive prior measurements, conduct interventions, and predict outcomes, with the hidden process being a randomly sampled structural causal model.

Research Conclusions:

– The study reveals a gap between prediction accuracy and causal mechanism recovery in LLM agents, suggesting intervention strategies improve results but remain challenging; consistency verification helps address premature stopping issues.

Paper link: https://huggingface.co/papers/2605.26029

30. Colored Noise Diffusion Sampling

Keywords: Diffusion models, spectral bias, Colored Noise Sampling, stochastic differential equation, FID

Category: Generative Models

Research Objective:

– Address spectral bias in image synthesis by developing a new sampling method called Colored Noise Sampling (CNS).

Research Methods:

– Introduce a novel mathematical framework reinterpreting SDE inference as frequency-decoupled energy transfer, utilizing dynamic, frequency-dependent schedules to allocate injected energy during image generation.

Research Conclusions:

– CNS significantly outperforms standard ODE and SDE baselines by reducing FID scores across various architectures and maintaining consistent improvements with Classifier-Free Guidance.

Paper link: https://huggingface.co/papers/2605.30332

31. When Should Models Change Their Minds? Contextual Belief Management in Large Language Models

Keywords: Contextual Belief Management, reinforcement learning, BeliefTrack, representation-level steering, Natural Language Processing

Category: Natural Language Processing

Research Objective:

– To enhance language models’ capabilities in managing long-term information, specifically focusing on updating, preserving, and filtering relevant information through Contextual Belief Management (CBM).

Research Methods:

– Implementation of a closed-world benchmark called BeliefTrack to measure CBM effectiveness in Rule Discovery and Circuit Diagnosis, using reinforcement learning with belief-state rewards and representation-level steering techniques.

Research Conclusions:

– Vanilla language models often fail in CBM tasks, but using reinforcement learning significantly lowers failure rates by 70.9%, with additional benefits from representation-level steering reducing failures by 46.1%.

Paper link: https://huggingface.co/papers/2605.30219

32. Skill0.5: Joint Skill Internalization and Utilization for Out-of-Distribution Generalization in Agentic Reinforcement Learning

Keywords: Skill0.5, Agentic Reinforcement Learning, Cognitive Foundation, Diagnostic Probing, Dynamic Router

Category: Reinforcement Learning

Research Objective:

– To propose Skill0.5, an innovative reinforcement learning framework for improving task performance by balancing general skill internalization and task-specific skill utilization.

Research Methods:

– Utilizes a dynamic, difficulty-aware router to stream tasks into mastery tiers, combining privileged distillation for complex tasks with diagnostic probing for simpler tasks.

Research Conclusions:

– Skill0.5 exhibits superior performance compared to traditional memory-based and skill-based RL methods, demonstrating effectiveness across both in-distribution and out-of-distribution scenarios.

Paper link: https://huggingface.co/papers/2605.28424

33. LoMo: Local Modality Substitution for Deeper Vision-Language Fusion

Keywords: Modality Substitution, Vision-Language Models, Data Curation, Multimodal Reasoning, Cross-Modal Representation

Category: Multi-Modal Learning

Research Objective:

– The research aims to address the performance degradation in vision-language models caused by modality sensitivity due to training data bias.

Research Methods:

– The introduction of Local Modality Substitution (LoMo), a novel data curation approach, which reformulates prompts into interleaved multimodal sequences to ensure cross-modal representational invariance.

Research Conclusions:

– LoMo significantly improves multimodal reasoning and cross-modal fusion, delivering consistent performance gains across various foundational models and benchmarks.

Paper link: https://huggingface.co/papers/2605.30265

34. UniSteer: Text-Guided Flow Matching in Activation Space for Versatile LLM Steering

Keywords: Activation-based control, Large Language Models, Text-guided activation flow matching, Conditional distribution, Universal conditional velocity field

Category: Natural Language Processing

Research Objective:

– Introduce UniSteer to control LLM behaviors and classification tasks through text-guided activation flow matching and universal conditional velocity field.

Research Methods:

– Develop a model that learns a conditional distribution over residual-stream activations, performs flow inversion, and supports activation-space classification.

Research Conclusions:

– UniSteer offers a unified interface for behavioral control in LLMs, achieving effectiveness in truthfulness steering, fine-grained concept steering, and multi-constraint instruction following.

Paper link: https://huggingface.co/papers/2605.30076

35. EarlyTom: Early Token Compression Completes Fast Video Understanding

Keywords: EarlyTom, Visual Tokens, Vision Encoder, Time-to-First-Token, Token Compression

Category: Computer Vision

Research Objective:

– The study aims to develop EarlyTom, a training-free framework to enhance the efficiency of vision encoders by compressing visual tokens early, reducing computational costs and maintaining model accuracy.

Research Methods:

– Introduced a strategy for early-stage visual token compression within the vision encoder, utilizing a decoupled spatial token selection to improve compression effectiveness.

Research Conclusions:

– EarlyTom reduces time-to-first-token (TTFT) by up to 2.65x and decreases FLOPs by up to 61%, while maintaining accuracy on par with full-token baselines, thus improving the practicality of deploying Video-LLMs in production environments.

Paper link: https://huggingface.co/papers/2605.30010

36. YoCausal: How Far is Video Generation from World Model? A Causality Perspective