1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.

Paper link: https://huggingface.co/papers/2607.27958

14. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Keywords: Explorative Modeling, Generative Models, FLOP Efficiency, Sample Efficiency, Parameter Efficiency

Category: Generative Models

Research Objective:

– Introduce Explorative Modeling as a new paradigm in generative modeling to enable end-to-end training by exploring candidate matches and training on the best predictions.

Research Methods:

– The Explorative Models are applied in generative domains, increasing exploration as a pretraining axis to improve performance across diverse datasets including images, video, and language.

Research Conclusions:

– Explorative Modeling enhances efficiency exponentially, achieving significant improvement in FLOP, sample, and parameter efficiency, and enables end-to-end reconstructive generative modeling with reduced inference steps, establishing itself as a new training axis for existing generative models.

Paper link: https://huggingface.co/papers/2607.27372

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Keywords: AI Tour Meeting, Large Language Model, agents, simulation tool

Category: AI Systems and Tools

Research Objective:

– To propose a travel planning framework called AI Tour Meeting using agents powered by Large Language Models (LLMs) to collaboratively create itineraries through natural language discussion.

Research Methods:

– Implementation of a framework for configuring agent personas and orchestrating discussion workflows, enabling monitoring and LLM deployment in the simulation of tour planning discussions.

Research Conclusions:

– The framework is validated by system demonstration and provides analytical insights into the behavior of multiple LLM agents during group travel planning.

Paper link: https://huggingface.co/papers/2607.18806

16. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Keywords: Speculative Decoding, Large Language Model, Truncation-based Verification, Collaborative Verification

Category: Natural Language Processing

Research Objective:

– To analyze the distributions induced by lossy verification schemes in speculative decoding and identify their potential pitfalls in large language model inference.

Research Methods:

– Classified lossy verification methods into truncation-based and collaborative verification, and developed a diagnostic evaluation framework across curated benchmarks.

Research Conclusions:

– Truncation-based methods are susceptible to significant performance degradation due to distributional distortion.

– For collaborative verification, controlling the overshoot of draft probabilities relative to target probabilities is crucial to maintaining output quality.

Paper link: https://huggingface.co/papers/2607.26627

17. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Keywords: Sparse mixture-of-experts, Expert Subspace Separation Index, coherent overlap, next-token prediction

Category: Natural Language Processing

Research Objective:

– The study aims to distinguish the benefits and interactions of experts within sparse mixture-of-experts (MoE) language models by analyzing their geometric arrangement.

Research Methods:

– Utilizes an Expert Subspace Separation Index (ESSI) and various controlled interventions, including matched-route residuals and prefix-controlled factorial setups, to analyze expert overlaps and their impact on representation.

Research Conclusions:

– Coherent overlap indicates that while routes overlap, useful multi-expert computation persists, clarifying why geometric similarity doesn’t alone determine redundancy or pruning value.

– Adding later experts improves next-token predictions significantly in most cases, emphasizing the importance of non-redundant expert selections.

Paper link: https://huggingface.co/papers/2607.28308

18. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Keywords: Fairness Pruning, Large Language Models, Demographic Bias, Causal Bias Localization, AI Ethics and Fairness

Category: AI Ethics and Fairness

Research Objective:

– To introduce and empirically validate Fairness Pruning, a method aimed at managing and mitigating demographic biases in large language models.

Research Methods:

– Utilization of minimally contrastive prompt pairs and inference-time activation capture to identify neurons that react differentially to demographic attributes.

– Empirical evaluation on large models such as Llama-3.2 and Salamandra-2B with structural intervention by zeroing neurons.

Research Conclusions:

– Demonstrated that zeroing specific neurons alters model responses to demographic variables, with a bidirectional bias destabilization effect.

– Confirmed that demographic bias processing and model capabilities are dissociable, setting foundations for transitioning to directional behavior modulation.

Paper link: https://huggingface.co/papers/2607.28319

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Keywords: Speech Emotion Recognition, Edge Devices, Knowledge Distillation, AI Native

Category: Machine Learning

Research Objective:

– This research aims to optimize on-device speech emotion recognition (SER) for real-time applications by developing a more efficient model that can operate effectively on edge devices.

Research Methods:

– Adaptive Multi-teacher Relational Distillation (AMRD) is proposed, using a one-class SVM to assign weights to teacher models and a relational distillation loss to improve inter-sample relational structure alignment between teacher and student models.

Research Conclusions:

– The AMRD approach outperforms single-teacher distillation baselines on the IEMOCAP and CREMA-D datasets, demonstrating its effectiveness across multiple student architectures and confirming that both components of the method yield complementary benefits.

Paper link: https://huggingface.co/papers/2607.25289

20.

Paper link: 

21. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Keywords: Pedestrian Archetypes, Real-world Traffic Scenarios, Behavior Patterns

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces new pedestrian archetypes that exhibit significant behavioral differences, enhancing the understanding of pedestrian behaviors beyond the original taxonomy.

Research Methods:

– Analysis of YouTube dash-cam videos for identifying additional pedestrian archetypes not covered in previous work.

Research Conclusions:

– Seven new pedestrian archetypes are presented, each defined by unique behaviors that are documented through video-frame evidence, thus expanding the existing framework for assessing pedestrian behavior in traffic.

Paper link: https://huggingface.co/papers/2607.16922

22. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Keywords: Vision-Language Models, GPU Memory Constraints, Image-QA Dataset

Category: Multi-Modal Learning

Research Objective:

– Address the challenge of long visual context in vision-language models by introducing ReToken, an explicit retrieval target for selecting query-relevant visual tokens.

Research Methods:

– ReToken is trained on a small image-QA dataset and designed to select a sparse set of visual tokens from a pre-filled visual KV cache.

Research Conclusions:

– ReToken shows consistent performance gains across image and video benchmarks, with significant improvements in Visual Haystacks and LVBench, while maintaining efficiency on a single H100 for both training and long-video inference.

Paper link: https://huggingface.co/papers/2607.28627

23. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Keywords: OmniScope, token compression, query, salience mismatch, AI Native

Category: Multi-Modal Learning

Research Objective:

– The objective is to address the cross-modal salience mismatch in token compression methods for omnimodal large language models by developing a framework named OmniScope that estimates relevance separately for audio and video.

Research Methods:

– Introduced a training-free token compression framework using the query as a shared semantic anchor, allocating modality-specific token budgets, pruning visual tokens with an anchor-delta strategy, and merging audio tokens to reduce redundancy.

Research Conclusions:

– OmniScope outperforms existing methods in accuracy across audio-video benchmarks and model scales, offering improvements in speed and memory efficiency while maintaining high accuracy with minimal drop.

Paper link: https://huggingface.co/papers/2607.23193

24. Harness-G: A Graph-Structured Harness for Search Agents

Keywords: Reinforcement Learning, retrieval aliasing, Harness-G, Structured Non-myopic Credit

Category: Reinforcement Learning

Research Objective:

– The study addresses the problem of retrieval aliasing in reinforcement learning search agents by proposing a new retrieval framework to optimize query generation and interaction.

Research Methods:

– The authors introduce Harness-G, a graph-structured retrieval framework that transforms query generation into finite action selection. They also develop Structured Non-myopic Credit (SNC) for better rewarding actions.

Research Conclusions:

– Harness-G achieves the highest average F1 across six QA benchmarks and outperforms the best baseline by significant margins, demonstrating its effectiveness in reducing linguistic aliasing and improving retrieval accuracy.

Paper link: https://huggingface.co/papers/2607.27652

25. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Keywords: Deep Research agents, Misleading Knowledge, Evidence Synthesis, Workflow Reliability

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to explore the reliability vulnerabilities in Deep Research agents when exposed to misleading knowledge in open information environments.

Research Methods:

– Introduces MisKnow-Agent, a framework for generating misleading knowledge instances to test these vulnerabilities, utilizing 5,933 quality-controlled instances derived from DeepResearch Benchmark tasks.

Research Conclusions:

– The experiments demonstrate that even minimal misleading knowledge can lead to false conclusions in final reports, highlighting a significant reliability issue.

– Despite verifier models identifying misleading instances, they can still be adopted in long-term workflows, indicating a gap in workflow-level evidence verification.

– Evaluated defenses show mitigation effects but fail to fully prevent the adoption of false conclusions, pointing to the need for improved verification and correction capabilities.

Paper link: https://huggingface.co/papers/2607.20891

26. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Keywords: LLM agents, filesystem-based memory, memory organization, search economy, management agent

Category: AI Systems and Tools

Research Objective:

– To systematically explore filesystem-based memory for Large Language Model (LLM) agents, assessing the structure and organization of such memory.

Research Methods:

– The study formalizes the concept of a memory filesystem with a management agent for organization, a search agent for querying, and an execution agent for task trajectory integration.

– Investigated variables include memory shape, stream scale, tool harness, and agent strength, with focus on memory growth, retrieval costs, and store health.

Research Conclusions:

– Organization in filesystem-based memory leads to improved search economy, significantly reducing retrieval costs.

– However, as memory grows, management agents struggle to maintain organization quality, and improvements in organization do not necessarily translate to better answer quality.

– The study identifies the tool set as a significant factor in shaping the memory store, highlighting a design space for agent memory rather than just an assumption.

Paper link: https://huggingface.co/papers/2607.26637

27. Multi-Head Attention Residuals

Keywords: Transformers, Multi-Head Attention Residuals, Nemotron, validation loss

Category: Machine Learning

Research Objective:

– Introduction of Multi-Head Attention Residuals (MHAR) to improve information propagation across model depth and width.

Research Methods:

– MHAR utilizes H per-subspace heads with distinct softmax distributions, trained on a Nemotron-based corpus to optimize validation loss across various model sizes.

Research Conclusions:

– MHAR outperforms standard Transformers with validation loss improvements across 100M, 350M, and 1B scales. Optimal head count H = 8 increases efficiency and maintains near-baseline memory consumption.

Paper link: https://huggingface.co/papers/2607.27230

28. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Keywords: Synthetic Environments, Echoverse, Reinforcement Learning, State Representation

Category: Reinforcement Learning

Research Objective:

– To improve computer-use agents’ training environments by focusing on behavioral depth, targeted interactions, and model improvement.

Research Methods:

– Developed Echoverse to compile specifications into stateful applications, incorporating a co-evolution loop for iterative environment and model enhancement.

Research Conclusions:

– Demonstrated significant model improvement across various environments, with increases in evaluation scores and live-site accuracy via deep environments and iterative environment repairs.

Paper link: https://huggingface.co/papers/2607.28074

29. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Keywords: INTACT, Intent-to-Action, Latent World Models, Action Law Semantics

Category: Reinforcement Learning

Research Objective:

– To introduce INTACT, an end-to-end JEPA, aimed at transforming action-labeled, reward-free trajectories into a deployable intent-to-action interface without expensive test-time search.

Research Methods:

– Utilizes a unique architecture with isomorphic intent representation for local and goal motion-intent models, and provides intact transfer from RGB evidence to action-effective latent intent coordinates.

Research Conclusions:

– The INTACT model achieved up to 100% success rates on LeWM tasks using a search-free policy, significantly reducing the need for candidate sequence sampling while improving performance through a shared four-task encoder. Direct inference is highly efficient, taking only 2.9-5.5 milliseconds.

Paper link: https://huggingface.co/papers/2607.26056

30. MemHarness: Memory Is Reconstructed, Not Replayed

Keywords: MemHarness, Negative Transfer, Context-grounded Guidance, Large Language Model Agents, GRPO

Category: Reinforcement Learning

Research Objective:

– The objective is to develop a framework named MemHarness that enables large language model agents to actively reconstruct past experiences based on the current context, thereby providing more accurate and context-relevant guidance.

Research Methods:

– MemHarness employs a unified policy model that critiques and reconstructs retrieved experiences for each decision step. This reconstructive ability is achieved through end-to-end training using the GRPO technique.

Research Conclusions:

– MemHarness significantly outperforms existing memory-augmented baselines and pure reinforcement learning models, especially in out-of-distribution scenarios, by preventing negative transfer and enhancing intrinsic reasoning capabilities.

Paper link: https://huggingface.co/papers/2607.28272

31. ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

Keywords: ShadowDancer, interactive video, dynamics, action transfer, cross-shadow prediction

Category: Computer Vision

Research Objective:

– To enable precise control of any-action, frame-level dynamics in interactive video world models through novel techniques.

Research Methods:

– Utilizes shadow pairs to construct scalable video pairs replaying the same dynamics with varied appearances.

– Implements cross-shadow prediction to learn actions by predicting dynamics from one shadow to another.

Research Conclusions:

– Demonstrated improved action transfer and extended action rollout in diverse dynamics families, achieving an average blinded win rate of 86% over existing models.

Paper link: https://huggingface.co/papers/2607.28362

32. See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Keywords: Multimodal large language models, Visual reasoning, See2Think, Visual Action-of-Thought

Category: Multi-Modal Learning

Research Objective:

– To introduce See2Think, an evaluation framework that assesses visual reasoning in multimodal large language models through benchmarks like See2ThinkBench and Visual Action-of-Thought.

Research Methods:

– The creation of See2ThinkBench with 1,200 visually dependent tasks across 12 categories and the use of Visual Action-of-Thought to record and analyze visual and textual operations under controlled settings.

Research Conclusions:

– Visual reasoning capabilities in models are strongly dependent on both the model and environment, showing varied performance across tasks. Additionally, even though models often select the correct visual operations, the accuracy improvements are hindered by challenges in rendering fidelity and reliance on feedback which does not always improve accuracy.

Paper link: https://huggingface.co/papers/2607.26769

33. RefCaptioner: Multi-Reference Image-Grounded Video Captioning

Keywords: multi-reference image-grounded video captioning, RefCaptioner, factual video descriptions, Hierarchical Coverage-Discounted GRPO, AI-generated videos

Category: Multi-Modal Learning

Research Objective:

– To introduce and address the task of multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level reference grounding.

Research Methods:

– Development of RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO to enhance reference selection and cross-reference consistency.

Research Conclusions:

– RefCaptioner outperforms other open-source models in overall performance, yielding captions preferred by human annotators and allowing for more source-faithful video reconstruction.

Paper link: https://huggingface.co/papers/2607.28509

34. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Keywords: Embodied intelligence, Ambient Capture Engine, AI Native, Multi-modal datasets, Human-centric data

Category: Robotics and Autonomous Systems

Research Objective:

– The study focuses on addressing the data bottleneck in embodied intelligence by introducing the Ambient Capture Engine (ACE), which captures comprehensive sensory data in real home environments.

Research Methods:

– The authors developed ACE to record multisensory data through a combination of table-scale and room-scale configurations, creating a detailed dataset, ACE-Data-0, encompassing 150 hours of interactions across 200 task categories.

Research Conclusions:

– Evaluations reveal significant limitations in current state-of-the-art methods, particularly under challenging conditions like occlusion and long temporal horizons, highlighting ACE’s potential in supporting scalable advancements in imitation learning and embodied AI systems.

Paper link: https://huggingface.co/papers/2607.28625

35. Flux-OPD: On-Policy Distillation with Evolving Contexts

Keywords: Large language model, evolving contexts, Flux-OPD, open-ended domains, teacher supervision

Category: Natural Language Processing

Research Objective:

– The study aims to develop a mechanism for more effective supervision in large language model training across open-ended domains, addressing the issue of lacking verifiable rewards and formalized task preferences.

Research Methods:

– The research introduces Flux-OPD, a novel OPD paradigm that leverages evolving contexts as a form of in-training supervision. It analyzes the effect of context through the reverse KL objective and uses context-conditioned teachers to provide contextual difference signals.

Research Conclusions:

– Flux-OPD demonstrates superiority over existing OPD paradigms, effectively capturing task preferences and enhancing performance in open-ended tasks by combining evolving contexts with teacher supervision.

Paper link: https://huggingface.co/papers/2607.28022

36. Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Keywords: Agentic Visual Reasoning, Multimodal Large Language Models, Mode Adaptiveness, Tool Effect, Beacon

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance the success rate of multimodal large language models (MLLMs) in complex tasks by rethinking agentic visual reasoning, focusing on Mode Adaptiveness and Tool Effect.

Research Methods:

– Comprehensive analysis to quantify Mode Adaptiveness and Tool Effect in existing models, followed by the proposal of a new model named Beacon, featuring Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion.

Research Conclusions:

– The research concludes that existing models have limited Mode Adaptiveness, and tool benefits are often neutralized by their drawbacks. Beacon, however, shows improved overall performance, notably enhancing Mode Adaptiveness and achieving genuine gains from tool use.

Paper link: https://huggingface.co/papers/2607.28595

37. VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Keywords: Text-to-video, VideoCoCo, Blender program, spatiotemporal process, generative video engine

Category: Generative Models

Research Objective:

– The study aims to address the limitations of text-to-video models in generating physically consistent dynamics by introducing a new framework called VideoCoCo.

Research Methods:

– A dual-engine framework is employed, where a coding agent synthesizes executable Blender code to specify scene evolution and produce deterministic drafts, followed by a generative video engine for photorealistic video transformation.

Research Conclusions:

– VideoCoCo effectively improves performance on benchmarks PhyGenBench and VBench-2.0 by providing an executable, controllable, and inspectable representation for video generation, achieving the best average scores on both benchmarks.

Paper link: https://huggingface.co/papers/2607.27380

38. PhiZero: A World Model Built Around Physical Language

Keywords: PhiZero, physical language, world model, zero-shot motion transfer

Category: Generative Models

Research Objective:

– The primary aim is to develop PhiZero, a model that abstracts predictive structure from visual experience to explicitly reason about physical world evolution using physical language.

Research Methods:

– PhiZero employs a reason-then-render paradigm where it infers future world evolution sequences in physical language and renders transitions into video through self-supervision from in-the-wild videos.

Research Conclusions:

– The experiments demonstrate PhiZero’s capability to model physically coherent world evolution and highlight its potential for realistic interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

Paper link: https://huggingface.co/papers/2607.28624

39. Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Keywords: GUI agents, Qwen-UI-Agent, Reinforcement Learning, MobileWorld, Online RL

Category: AI Systems and Tools

Research Objective:

– The paper aims to advance GUI agents for real-world applications, enabling reliable operation across various devices and platforms, proactive service initiation, and autonomous capability improvements.

Research Methods:

– Introduces Qwen-UI-Agent, utilizing a unified action space that combines GUI operations with CLI execution, and employs an AutoResearch-style data flywheel and Online Reinforcement Learning (RL) for training and advancing agent performance.

Research Conclusions:

– Qwen-UI-Agent demonstrates state-of-the-art performance in mobile-use benchmarks and competitive results in computer- and browser-use tasks, achieving high scores on MobileWorld, AndroidDaily, and WebArena among others.

Paper link: https://huggingface.co/papers/2607.28227

The post AI Native Daily Paper Digest – 20261002 – Metis | SpatialCLI | Chimera appeared first on AI Native Foundation.

1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.

Paper link: https://huggingface.co/papers/2607.27958

14. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Keywords: Explorative Modeling, Generative Models, FLOP Efficiency, Sample Efficiency, Parameter Efficiency

Category: Generative Models

Research Objective:

– Introduce Explorative Modeling as a new paradigm in generative modeling to enable end-to-end training by exploring candidate matches and training on the best predictions.

Research Methods:

– The Explorative Models are applied in generative domains, increasing exploration as a pretraining axis to improve performance across diverse datasets including images, video, and language.

Research Conclusions:

– Explorative Modeling enhances efficiency exponentially, achieving significant improvement in FLOP, sample, and parameter efficiency, and enables end-to-end reconstructive generative modeling with reduced inference steps, establishing itself as a new training axis for existing generative models.

Paper link: https://huggingface.co/papers/2607.27372

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Keywords: AI Tour Meeting, Large Language Model, agents, simulation tool

Category: AI Systems and Tools

Research Objective:

– To propose a travel planning framework called AI Tour Meeting using agents powered by Large Language Models (LLMs) to collaboratively create itineraries through natural language discussion.

Research Methods:

– Implementation of a framework for configuring agent personas and orchestrating discussion workflows, enabling monitoring and LLM deployment in the simulation of tour planning discussions.

Research Conclusions:

– The framework is validated by system demonstration and provides analytical insights into the behavior of multiple LLM agents during group travel planning.

Paper link: https://huggingface.co/papers/2607.18806

16. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Keywords: Speculative Decoding, Large Language Model, Truncation-based Verification, Collaborative Verification

Category: Natural Language Processing

Research Objective:

– To analyze the distributions induced by lossy verification schemes in speculative decoding and identify their potential pitfalls in large language model inference.

Research Methods:

– Classified lossy verification methods into truncation-based and collaborative verification, and developed a diagnostic evaluation framework across curated benchmarks.

Research Conclusions:

– Truncation-based methods are susceptible to significant performance degradation due to distributional distortion.

– For collaborative verification, controlling the overshoot of draft probabilities relative to target probabilities is crucial to maintaining output quality.

Paper link: https://huggingface.co/papers/2607.26627

17. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Keywords: Sparse mixture-of-experts, Expert Subspace Separation Index, coherent overlap, next-token prediction

Category: Natural Language Processing

Research Objective:

– The study aims to distinguish the benefits and interactions of experts within sparse mixture-of-experts (MoE) language models by analyzing their geometric arrangement.

Research Methods:

– Utilizes an Expert Subspace Separation Index (ESSI) and various controlled interventions, including matched-route residuals and prefix-controlled factorial setups, to analyze expert overlaps and their impact on representation.

Research Conclusions:

– Coherent overlap indicates that while routes overlap, useful multi-expert computation persists, clarifying why geometric similarity doesn’t alone determine redundancy or pruning value.

– Adding later experts improves next-token predictions significantly in most cases, emphasizing the importance of non-redundant expert selections.

Paper link: https://huggingface.co/papers/2607.28308

18. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Keywords: Fairness Pruning, Large Language Models, Demographic Bias, Causal Bias Localization, AI Ethics and Fairness

Category: AI Ethics and Fairness

Research Objective:

– To introduce and empirically validate Fairness Pruning, a method aimed at managing and mitigating demographic biases in large language models.

Research Methods:

– Utilization of minimally contrastive prompt pairs and inference-time activation capture to identify neurons that react differentially to demographic attributes.

– Empirical evaluation on large models such as Llama-3.2 and Salamandra-2B with structural intervention by zeroing neurons.

Research Conclusions:

– Demonstrated that zeroing specific neurons alters model responses to demographic variables, with a bidirectional bias destabilization effect.

– Confirmed that demographic bias processing and model capabilities are dissociable, setting foundations for transitioning to directional behavior modulation.

Paper link: https://huggingface.co/papers/2607.28319

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Keywords: Speech Emotion Recognition, Edge Devices, Knowledge Distillation, AI Native

Category: Machine Learning

Research Objective:

– This research aims to optimize on-device speech emotion recognition (SER) for real-time applications by developing a more efficient model that can operate effectively on edge devices.

Research Methods:

– Adaptive Multi-teacher Relational Distillation (AMRD) is proposed, using a one-class SVM to assign weights to teacher models and a relational distillation loss to improve inter-sample relational structure alignment between teacher and student models.

Research Conclusions:

– The AMRD approach outperforms single-teacher distillation baselines on the IEMOCAP and CREMA-D datasets, demonstrating its effectiveness across multiple student architectures and confirming that both components of the method yield complementary benefits.

Paper link: https://huggingface.co/papers/2607.25289

20.

Paper link: 

21. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Keywords: Pedestrian Archetypes, Real-world Traffic Scenarios, Behavior Patterns

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces new pedestrian archetypes that exhibit significant behavioral differences, enhancing the understanding of pedestrian behaviors beyond the original taxonomy.

Research Methods:

– Analysis of YouTube dash-cam videos for identifying additional pedestrian archetypes not covered in previous work.

Research Conclusions:

– Seven new pedestrian archetypes are presented, each defined by unique behaviors that are documented through video-frame evidence, thus expanding the existing framework for assessing pedestrian behavior in traffic.

Paper link: https://huggingface.co/papers/2607.16922

22. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Keywords: Vision-Language Models, GPU Memory Constraints, Image-QA Dataset

Category: Multi-Modal Learning

Research Objective:

– Address the challenge of long visual context in vision-language models by introducing ReToken, an explicit retrieval target for selecting query-relevant visual tokens.

Research Methods:

– ReToken is trained on a small image-QA dataset and designed to select a sparse set of visual tokens from a pre-filled visual KV cache.

Research Conclusions:

– ReToken shows consistent performance gains across image and video benchmarks, with significant improvements in Visual Haystacks and LVBench, while maintaining efficiency on a single H100 for both training and long-video inference.

Paper link: https://huggingface.co/papers/2607.28627

23. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Keywords: OmniScope, token compression, query, salience mismatch, AI Native

Category: Multi-Modal Learning

Research Objective:

– The objective is to address the cross-modal salience mismatch in token compression methods for omnimodal large language models by developing a framework named OmniScope that estimates relevance separately for audio and video.

Research Methods:

– Introduced a training-free token compression framework using the query as a shared semantic anchor, allocating modality-specific token budgets, pruning visual tokens with an anchor-delta strategy, and merging audio tokens to reduce redundancy.

Research Conclusions:

– OmniScope outperforms existing methods in accuracy across audio-video benchmarks and model scales, offering improvements in speed and memory efficiency while maintaining high accuracy with minimal drop.

Paper link: https://huggingface.co/papers/2607.23193

24. Harness-G: A Graph-Structured Harness for Search Agents

Keywords: Reinforcement Learning, retrieval aliasing, Harness-G, Structured Non-myopic Credit

Category: Reinforcement Learning

Research Objective:

– The study addresses the problem of retrieval aliasing in reinforcement learning search agents by proposing a new retrieval framework to optimize query generation and interaction.

Research Methods:

– The authors introduce Harness-G, a graph-structured retrieval framework that transforms query generation into finite action selection. They also develop Structured Non-myopic Credit (SNC) for better rewarding actions.

Research Conclusions:

– Harness-G achieves the highest average F1 across six QA benchmarks and outperforms the best baseline by significant margins, demonstrating its effectiveness in reducing linguistic aliasing and improving retrieval accuracy.

Paper link: https://huggingface.co/papers/2607.27652

25. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Keywords: Deep Research agents, Misleading Knowledge, Evidence Synthesis, Workflow Reliability

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to explore the reliability vulnerabilities in Deep Research agents when exposed to misleading knowledge in open information environments.

Research Methods:

– Introduces MisKnow-Agent, a framework for generating misleading knowledge instances to test these vulnerabilities, utilizing 5,933 quality-controlled instances derived from DeepResearch Benchmark tasks.

Research Conclusions:

– The experiments demonstrate that even minimal misleading knowledge can lead to false conclusions in final reports, highlighting a significant reliability issue.

– Despite verifier models identifying misleading instances, they can still be adopted in long-term workflows, indicating a gap in workflow-level evidence verification.

– Evaluated defenses show mitigation effects but fail to fully prevent the adoption of false conclusions, pointing to the need for improved verification and correction capabilities.

Paper link: https://huggingface.co/papers/2607.20891

26. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Keywords: LLM agents, filesystem-based memory, memory organization, search economy, management agent

Category: AI Systems and Tools

Research Objective:

– To systematically explore filesystem-based memory for Large Language Model (LLM) agents, assessing the structure and organization of such memory.

Research Methods:

– The study formalizes the concept of a memory filesystem with a management agent for organization, a search agent for querying, and an execution agent for task trajectory integration.

– Investigated variables include memory shape, stream scale, tool harness, and agent strength, with focus on memory growth, retrieval costs, and store health.

Research Conclusions:

– Organization in filesystem-based memory leads to improved search economy, significantly reducing retrieval costs.

– However, as memory grows, management agents struggle to maintain organization quality, and improvements in organization do not necessarily translate to better answer quality.

– The study identifies the tool set as a significant factor in shaping the memory store, highlighting a design space for agent memory rather than just an assumption.

Paper link: https://huggingface.co/papers/2607.26637

27. Multi-Head Attention Residuals

Keywords: Transformers, Multi-Head Attention Residuals, Nemotron, validation loss

Category: Machine Learning

Research Objective:

– Introduction of Multi-Head Attention Residuals (MHAR) to improve information propagation across model depth and width.

Research Methods:

– MHAR utilizes H per-subspace heads with distinct softmax distributions, trained on a Nemotron-based corpus to optimize validation loss across various model sizes.

Research Conclusions:

– MHAR outperforms standard Transformers with validation loss improvements across 100M, 350M, and 1B scales. Optimal head count H = 8 increases efficiency and maintains near-baseline memory consumption.

Paper link: https://huggingface.co/papers/2607.27230

28. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Keywords: Synthetic Environments, Echoverse, Reinforcement Learning, State Representation

Category: Reinforcement Learning

Research Objective:

– To improve computer-use agents’ training environments by focusing on behavioral depth, targeted interactions, and model improvement.

Research Methods:

– Developed Echoverse to compile specifications into stateful applications, incorporating a co-evolution loop for iterative environment and model enhancement.

Research Conclusions:

– Demonstrated significant model improvement across various environments, with increases in evaluation scores and live-site accuracy via deep environments and iterative environment repairs.

Paper link: https://huggingface.co/papers/2607.28074

29. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Keywords: INTACT, Intent-to-Action, Latent World Models, Action Law Semantics

Category: Reinforcement Learning

Research Objective:

– To introduce INTACT, an end-to-end JEPA, aimed at transforming action-labeled, reward-free trajectories into a deployable intent-to-action interface without expensive test-time search.

Research Methods:

– Utilizes a unique architecture with isomorphic intent representation for local and goal motion-intent models, and provides intact transfer from RGB evidence to action-effective latent intent coordinates.

Research Conclusions:

– The INTACT model achieved up to 100% success rates on LeWM tasks using a search-free policy, significantly reducing the need for candidate sequence sampling while improving performance through a shared four-task encoder. Direct inference is highly efficient, taking only 2.9-5.5 milliseconds.

Paper link: https://huggingface.co/papers/2607.26056

30. MemHarness: Memory Is Reconstructed, Not Replayed

Keywords: MemHarness, Negative Transfer, Context-grounded Guidance, Large Language Model Agents, GRPO

Category: Reinforcement Learning

Research Objective:

– The objective is to develop a framework named MemHarness that enables large language model agents to actively reconstruct past experiences based on the current context, thereby providing more accurate and context-relevant guidance.

Research Methods:

– MemHarness employs a unified policy model that critiques and reconstructs retrieved experiences for each decision step. This reconstructive ability is achieved through end-to-end training using the GRPO technique.

Research Conclusions:

– MemHarness significantly outperforms existing memory-augmented baselines and pure reinforcement learning models, especially in out-of-distribution scenarios, by preventing negative transfer and enhancing intrinsic reasoning capabilities.

Paper link: https://huggingface.co/papers/2607.28272

31. ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

Keywords: ShadowDancer, interactive video, dynamics, action transfer, cross-shadow prediction

Category: Computer Vision

Research Objective:

– To enable precise control of any-action, frame-level dynamics in interactive video world models through novel techniques.

Research Methods:

– Utilizes shadow pairs to construct scalable video pairs replaying the same dynamics with varied appearances.

– Implements cross-shadow prediction to learn actions by predicting dynamics from one shadow to another.

Research Conclusions:

– Demonstrated improved action transfer and extended action rollout in diverse dynamics families, achieving an average blinded win rate of 86% over existing models.

Paper link: https://huggingface.co/papers/2607.28362

32. See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Keywords: Multimodal large language models, Visual reasoning, See2Think, Visual Action-of-Thought

Category: Multi-Modal Learning

Research Objective:

– To introduce See2Think, an evaluation framework that assesses visual reasoning in multimodal large language models through benchmarks like See2ThinkBench and Visual Action-of-Thought.

Research Methods:

– The creation of See2ThinkBench with 1,200 visually dependent tasks across 12 categories and the use of Visual Action-of-Thought to record and analyze visual and textual operations under controlled settings.

Research Conclusions:

– Visual reasoning capabilities in models are strongly dependent on both the model and environment, showing varied performance across tasks. Additionally, even though models often select the correct visual operations, the accuracy improvements are hindered by challenges in rendering fidelity and reliance on feedback which does not always improve accuracy.

Paper link: https://huggingface.co/papers/2607.26769

33. RefCaptioner: Multi-Reference Image-Grounded Video Captioning

Keywords: multi-reference image-grounded video captioning, RefCaptioner, factual video descriptions, Hierarchical Coverage-Discounted GRPO, AI-generated videos

Category: Multi-Modal Learning

Research Objective:

– To introduce and address the task of multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level reference grounding.

Research Methods:

– Development of RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO to enhance reference selection and cross-reference consistency.

Research Conclusions:

– RefCaptioner outperforms other open-source models in overall performance, yielding captions preferred by human annotators and allowing for more source-faithful video reconstruction.

Paper link: https://huggingface.co/papers/2607.28509

34. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Keywords: Embodied intelligence, Ambient Capture Engine, AI Native, Multi-modal datasets, Human-centric data

Category: Robotics and Autonomous Systems

Research Objective:

– The study focuses on addressing the data bottleneck in embodied intelligence by introducing the Ambient Capture Engine (ACE), which captures comprehensive sensory data in real home environments.

Research Methods:

– The authors developed ACE to record multisensory data through a combination of table-scale and room-scale configurations, creating a detailed dataset, ACE-Data-0, encompassing 150 hours of interactions across 200 task categories.

Research Conclusions:

– Evaluations reveal significant limitations in current state-of-the-art methods, particularly under challenging conditions like occlusion and long temporal horizons, highlighting ACE’s potential in supporting scalable advancements in imitation learning and embodied AI systems.

Paper link: https://huggingface.co/papers/2607.28625

35. Flux-OPD: On-Policy Distillation with Evolving Contexts

Keywords: Large language model, evolving contexts, Flux-OPD, open-ended domains, teacher supervision

Category: Natural Language Processing

Research Objective:

– The study aims to develop a mechanism for more effective supervision in large language model training across open-ended domains, addressing the issue of lacking verifiable rewards and formalized task preferences.

Research Methods:

– The research introduces Flux-OPD, a novel OPD paradigm that leverages evolving contexts as a form of in-training supervision. It analyzes the effect of context through the reverse KL objective and uses context-conditioned teachers to provide contextual difference signals.

Research Conclusions:

– Flux-OPD demonstrates superiority over existing OPD paradigms, effectively capturing task preferences and enhancing performance in open-ended tasks by combining evolving contexts with teacher supervision.

Paper link: https://huggingface.co/papers/2607.28022

36. Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Keywords: Agentic Visual Reasoning, Multimodal Large Language Models, Mode Adaptiveness, Tool Effect, Beacon

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance the success rate of multimodal large language models (MLLMs) in complex tasks by rethinking agentic visual reasoning, focusing on Mode Adaptiveness and Tool Effect.

Research Methods:

– Comprehensive analysis to quantify Mode Adaptiveness and Tool Effect in existing models, followed by the proposal of a new model named Beacon, featuring Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion.

Research Conclusions:

– The research concludes that existing models have limited Mode Adaptiveness, and tool benefits are often neutralized by their drawbacks. Beacon, however, shows improved overall performance, notably enhancing Mode Adaptiveness and achieving genuine gains from tool use.

Paper link: https://huggingface.co/papers/2607.28595

37. VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Keywords: Text-to-video, VideoCoCo, Blender program, spatiotemporal process, generative video engine

Category: Generative Models

Research Objective:

– The study aims to address the limitations of text-to-video models in generating physically consistent dynamics by introducing a new framework called VideoCoCo.

Research Methods:

– A dual-engine framework is employed, where a coding agent synthesizes executable Blender code to specify scene evolution and produce deterministic drafts, followed by a generative video engine for photorealistic video transformation.

Research Conclusions:

– VideoCoCo effectively improves performance on benchmarks PhyGenBench and VBench-2.0 by providing an executable, controllable, and inspectable representation for video generation, achieving the best average scores on both benchmarks.

Paper link: https://huggingface.co/papers/2607.27380

38. PhiZero: A World Model Built Around Physical Language

Keywords: PhiZero, physical language, world model, zero-shot motion transfer

Category: Generative Models

Research Objective:

– The primary aim is to develop PhiZero, a model that abstracts predictive structure from visual experience to explicitly reason about physical world evolution using physical language.

Research Methods:

– PhiZero employs a reason-then-render paradigm where it infers future world evolution sequences in physical language and renders transitions into video through self-supervision from in-the-wild videos.

Research Conclusions:

– The experiments demonstrate PhiZero’s capability to model physically coherent world evolution and highlight its potential for realistic interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

Paper link: https://huggingface.co/papers/2607.28624

39. Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Keywords: GUI agents, Qwen-UI-Agent, Reinforcement Learning, MobileWorld, Online RL

Category: AI Systems and Tools

Research Objective:

– The paper aims to advance GUI agents for real-world applications, enabling reliable operation across various devices and platforms, proactive service initiation, and autonomous capability improvements.

Research Methods:

– Introduces Qwen-UI-Agent, utilizing a unified action space that combines GUI operations with CLI execution, and employs an AutoResearch-style data flywheel and Online Reinforcement Learning (RL) for training and advancing agent performance.

Research Conclusions:

– Qwen-UI-Agent demonstrates state-of-the-art performance in mobile-use benchmarks and competitive results in computer- and browser-use tasks, achieving high scores on MobileWorld, AndroidDaily, and WebArena among others.

Paper link: https://huggingface.co/papers/2607.28227

The post AI Native Daily Paper Digest – 20261001 – Metis | AskChem | Chimera appeared first on AI Native Foundation.

1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.

Paper link: https://huggingface.co/papers/2607.27958

14. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Keywords: Explorative Modeling, Generative Models, FLOP Efficiency, Sample Efficiency, Parameter Efficiency

Category: Generative Models

Research Objective:

– Introduce Explorative Modeling as a new paradigm in generative modeling to enable end-to-end training by exploring candidate matches and training on the best predictions.

Research Methods:

– The Explorative Models are applied in generative domains, increasing exploration as a pretraining axis to improve performance across diverse datasets including images, video, and language.

Research Conclusions:

– Explorative Modeling enhances efficiency exponentially, achieving significant improvement in FLOP, sample, and parameter efficiency, and enables end-to-end reconstructive generative modeling with reduced inference steps, establishing itself as a new training axis for existing generative models.

Paper link: https://huggingface.co/papers/2607.27372

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Keywords: AI Tour Meeting, Large Language Model, agents, simulation tool

Category: AI Systems and Tools

Research Objective:

– To propose a travel planning framework called AI Tour Meeting using agents powered by Large Language Models (LLMs) to collaboratively create itineraries through natural language discussion.

Research Methods:

– Implementation of a framework for configuring agent personas and orchestrating discussion workflows, enabling monitoring and LLM deployment in the simulation of tour planning discussions.

Research Conclusions:

– The framework is validated by system demonstration and provides analytical insights into the behavior of multiple LLM agents during group travel planning.

Paper link: https://huggingface.co/papers/2607.18806

16. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Keywords: Speculative Decoding, Large Language Model, Truncation-based Verification, Collaborative Verification

Category: Natural Language Processing

Research Objective:

– To analyze the distributions induced by lossy verification schemes in speculative decoding and identify their potential pitfalls in large language model inference.

Research Methods:

– Classified lossy verification methods into truncation-based and collaborative verification, and developed a diagnostic evaluation framework across curated benchmarks.

Research Conclusions:

– Truncation-based methods are susceptible to significant performance degradation due to distributional distortion.

– For collaborative verification, controlling the overshoot of draft probabilities relative to target probabilities is crucial to maintaining output quality.

Paper link: https://huggingface.co/papers/2607.26627

17. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Keywords: Sparse mixture-of-experts, Expert Subspace Separation Index, coherent overlap, next-token prediction

Category: Natural Language Processing

Research Objective:

– The study aims to distinguish the benefits and interactions of experts within sparse mixture-of-experts (MoE) language models by analyzing their geometric arrangement.

Research Methods:

– Utilizes an Expert Subspace Separation Index (ESSI) and various controlled interventions, including matched-route residuals and prefix-controlled factorial setups, to analyze expert overlaps and their impact on representation.

Research Conclusions:

– Coherent overlap indicates that while routes overlap, useful multi-expert computation persists, clarifying why geometric similarity doesn’t alone determine redundancy or pruning value.

– Adding later experts improves next-token predictions significantly in most cases, emphasizing the importance of non-redundant expert selections.

Paper link: https://huggingface.co/papers/2607.28308

18. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Keywords: Fairness Pruning, Large Language Models, Demographic Bias, Causal Bias Localization, AI Ethics and Fairness

Category: AI Ethics and Fairness

Research Objective:

– To introduce and empirically validate Fairness Pruning, a method aimed at managing and mitigating demographic biases in large language models.

Research Methods:

– Utilization of minimally contrastive prompt pairs and inference-time activation capture to identify neurons that react differentially to demographic attributes.

– Empirical evaluation on large models such as Llama-3.2 and Salamandra-2B with structural intervention by zeroing neurons.

Research Conclusions:

– Demonstrated that zeroing specific neurons alters model responses to demographic variables, with a bidirectional bias destabilization effect.

– Confirmed that demographic bias processing and model capabilities are dissociable, setting foundations for transitioning to directional behavior modulation.

Paper link: https://huggingface.co/papers/2607.28319

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Keywords: Speech Emotion Recognition, Edge Devices, Knowledge Distillation, AI Native

Category: Machine Learning

Research Objective:

– This research aims to optimize on-device speech emotion recognition (SER) for real-time applications by developing a more efficient model that can operate effectively on edge devices.

Research Methods:

– Adaptive Multi-teacher Relational Distillation (AMRD) is proposed, using a one-class SVM to assign weights to teacher models and a relational distillation loss to improve inter-sample relational structure alignment between teacher and student models.

Research Conclusions:

– The AMRD approach outperforms single-teacher distillation baselines on the IEMOCAP and CREMA-D datasets, demonstrating its effectiveness across multiple student architectures and confirming that both components of the method yield complementary benefits.

Paper link: https://huggingface.co/papers/2607.25289

20.

Paper link: 

21. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Keywords: Pedestrian Archetypes, Real-world Traffic Scenarios, Behavior Patterns

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces new pedestrian archetypes that exhibit significant behavioral differences, enhancing the understanding of pedestrian behaviors beyond the original taxonomy.

Research Methods:

– Analysis of YouTube dash-cam videos for identifying additional pedestrian archetypes not covered in previous work.

Research Conclusions:

– Seven new pedestrian archetypes are presented, each defined by unique behaviors that are documented through video-frame evidence, thus expanding the existing framework for assessing pedestrian behavior in traffic.

Paper link: https://huggingface.co/papers/2607.16922

22. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Keywords: Vision-Language Models, GPU Memory Constraints, Image-QA Dataset

Category: Multi-Modal Learning

Research Objective:

– Address the challenge of long visual context in vision-language models by introducing ReToken, an explicit retrieval target for selecting query-relevant visual tokens.

Research Methods:

– ReToken is trained on a small image-QA dataset and designed to select a sparse set of visual tokens from a pre-filled visual KV cache.

Research Conclusions:

– ReToken shows consistent performance gains across image and video benchmarks, with significant improvements in Visual Haystacks and LVBench, while maintaining efficiency on a single H100 for both training and long-video inference.

Paper link: https://huggingface.co/papers/2607.28627

23. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Keywords: OmniScope, token compression, query, salience mismatch, AI Native

Category: Multi-Modal Learning

Research Objective:

– The objective is to address the cross-modal salience mismatch in token compression methods for omnimodal large language models by developing a framework named OmniScope that estimates relevance separately for audio and video.

Research Methods:

– Introduced a training-free token compression framework using the query as a shared semantic anchor, allocating modality-specific token budgets, pruning visual tokens with an anchor-delta strategy, and merging audio tokens to reduce redundancy.

Research Conclusions:

– OmniScope outperforms existing methods in accuracy across audio-video benchmarks and model scales, offering improvements in speed and memory efficiency while maintaining high accuracy with minimal drop.

Paper link: https://huggingface.co/papers/2607.23193

24. Harness-G: A Graph-Structured Harness for Search Agents

Keywords: Reinforcement Learning, retrieval aliasing, Harness-G, Structured Non-myopic Credit

Category: Reinforcement Learning

Research Objective:

– The study addresses the problem of retrieval aliasing in reinforcement learning search agents by proposing a new retrieval framework to optimize query generation and interaction.

Research Methods:

– The authors introduce Harness-G, a graph-structured retrieval framework that transforms query generation into finite action selection. They also develop Structured Non-myopic Credit (SNC) for better rewarding actions.

Research Conclusions:

– Harness-G achieves the highest average F1 across six QA benchmarks and outperforms the best baseline by significant margins, demonstrating its effectiveness in reducing linguistic aliasing and improving retrieval accuracy.

Paper link: https://huggingface.co/papers/2607.27652

25. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Keywords: Deep Research agents, Misleading Knowledge, Evidence Synthesis, Workflow Reliability

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to explore the reliability vulnerabilities in Deep Research agents when exposed to misleading knowledge in open information environments.

Research Methods:

– Introduces MisKnow-Agent, a framework for generating misleading knowledge instances to test these vulnerabilities, utilizing 5,933 quality-controlled instances derived from DeepResearch Benchmark tasks.

Research Conclusions:

– The experiments demonstrate that even minimal misleading knowledge can lead to false conclusions in final reports, highlighting a significant reliability issue.

– Despite verifier models identifying misleading instances, they can still be adopted in long-term workflows, indicating a gap in workflow-level evidence verification.

– Evaluated defenses show mitigation effects but fail to fully prevent the adoption of false conclusions, pointing to the need for improved verification and correction capabilities.

Paper link: https://huggingface.co/papers/2607.20891

26. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Keywords: LLM agents, filesystem-based memory, memory organization, search economy, management agent

Category: AI Systems and Tools

Research Objective:

– To systematically explore filesystem-based memory for Large Language Model (LLM) agents, assessing the structure and organization of such memory.

Research Methods:

– The study formalizes the concept of a memory filesystem with a management agent for organization, a search agent for querying, and an execution agent for task trajectory integration.

– Investigated variables include memory shape, stream scale, tool harness, and agent strength, with focus on memory growth, retrieval costs, and store health.

Research Conclusions:

– Organization in filesystem-based memory leads to improved search economy, significantly reducing retrieval costs.

– However, as memory grows, management agents struggle to maintain organization quality, and improvements in organization do not necessarily translate to better answer quality.

– The study identifies the tool set as a significant factor in shaping the memory store, highlighting a design space for agent memory rather than just an assumption.

Paper link: https://huggingface.co/papers/2607.26637

27. Multi-Head Attention Residuals

Keywords: Transformers, Multi-Head Attention Residuals, Nemotron, validation loss

Category: Machine Learning

Research Objective:

– Introduction of Multi-Head Attention Residuals (MHAR) to improve information propagation across model depth and width.

Research Methods:

– MHAR utilizes H per-subspace heads with distinct softmax distributions, trained on a Nemotron-based corpus to optimize validation loss across various model sizes.

Research Conclusions:

– MHAR outperforms standard Transformers with validation loss improvements across 100M, 350M, and 1B scales. Optimal head count H = 8 increases efficiency and maintains near-baseline memory consumption.

Paper link: https://huggingface.co/papers/2607.27230

28. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Keywords: Synthetic Environments, Echoverse, Reinforcement Learning, State Representation

Category: Reinforcement Learning

Research Objective:

– To improve computer-use agents’ training environments by focusing on behavioral depth, targeted interactions, and model improvement.

Research Methods:

– Developed Echoverse to compile specifications into stateful applications, incorporating a co-evolution loop for iterative environment and model enhancement.

Research Conclusions:

– Demonstrated significant model improvement across various environments, with increases in evaluation scores and live-site accuracy via deep environments and iterative environment repairs.

Paper link: https://huggingface.co/papers/2607.28074

29. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Keywords: INTACT, Intent-to-Action, Latent World Models, Action Law Semantics

Category: Reinforcement Learning

Research Objective:

– To introduce INTACT, an end-to-end JEPA, aimed at transforming action-labeled, reward-free trajectories into a deployable intent-to-action interface without expensive test-time search.

Research Methods:

– Utilizes a unique architecture with isomorphic intent representation for local and goal motion-intent models, and provides intact transfer from RGB evidence to action-effective latent intent coordinates.

Research Conclusions:

– The INTACT model achieved up to 100% success rates on LeWM tasks using a search-free policy, significantly reducing the need for candidate sequence sampling while improving performance through a shared four-task encoder. Direct inference is highly efficient, taking only 2.9-5.5 milliseconds.

Paper link: https://huggingface.co/papers/2607.26056

30. MemHarness: Memory Is Reconstructed, Not Replayed

Keywords: MemHarness, Negative Transfer, Context-grounded Guidance, Large Language Model Agents, GRPO

Category: Reinforcement Learning

Research Objective:

– The objective is to develop a framework named MemHarness that enables large language model agents to actively reconstruct past experiences based on the current context, thereby providing more accurate and context-relevant guidance.

Research Methods:

– MemHarness employs a unified policy model that critiques and reconstructs retrieved experiences for each decision step. This reconstructive ability is achieved through end-to-end training using the GRPO technique.

Research Conclusions:

– MemHarness significantly outperforms existing memory-augmented baselines and pure reinforcement learning models, especially in out-of-distribution scenarios, by preventing negative transfer and enhancing intrinsic reasoning capabilities.

Paper link: https://huggingface.co/papers/2607.28272

31. ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

Keywords: ShadowDancer, interactive video, dynamics, action transfer, cross-shadow prediction

Category: Computer Vision

Research Objective:

– To enable precise control of any-action, frame-level dynamics in interactive video world models through novel techniques.

Research Methods:

– Utilizes shadow pairs to construct scalable video pairs replaying the same dynamics with varied appearances.

– Implements cross-shadow prediction to learn actions by predicting dynamics from one shadow to another.

Research Conclusions:

– Demonstrated improved action transfer and extended action rollout in diverse dynamics families, achieving an average blinded win rate of 86% over existing models.

Paper link: https://huggingface.co/papers/2607.28362

32. See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Keywords: Multimodal large language models, Visual reasoning, See2Think, Visual Action-of-Thought

Category: Multi-Modal Learning

Research Objective:

– To introduce See2Think, an evaluation framework that assesses visual reasoning in multimodal large language models through benchmarks like See2ThinkBench and Visual Action-of-Thought.

Research Methods:

– The creation of See2ThinkBench with 1,200 visually dependent tasks across 12 categories and the use of Visual Action-of-Thought to record and analyze visual and textual operations under controlled settings.

Research Conclusions:

– Visual reasoning capabilities in models are strongly dependent on both the model and environment, showing varied performance across tasks. Additionally, even though models often select the correct visual operations, the accuracy improvements are hindered by challenges in rendering fidelity and reliance on feedback which does not always improve accuracy.

Paper link: https://huggingface.co/papers/2607.26769

33. RefCaptioner: Multi-Reference Image-Grounded Video Captioning

Keywords: multi-reference image-grounded video captioning, RefCaptioner, factual video descriptions, Hierarchical Coverage-Discounted GRPO, AI-generated videos

Category: Multi-Modal Learning

Research Objective:

– To introduce and address the task of multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level reference grounding.

Research Methods:

– Development of RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO to enhance reference selection and cross-reference consistency.

Research Conclusions:

– RefCaptioner outperforms other open-source models in overall performance, yielding captions preferred by human annotators and allowing for more source-faithful video reconstruction.

Paper link: https://huggingface.co/papers/2607.28509

34. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Keywords: Embodied intelligence, Ambient Capture Engine, AI Native, Multi-modal datasets, Human-centric data

Category: Robotics and Autonomous Systems

Research Objective:

– The study focuses on addressing the data bottleneck in embodied intelligence by introducing the Ambient Capture Engine (ACE), which captures comprehensive sensory data in real home environments.

Research Methods:

– The authors developed ACE to record multisensory data through a combination of table-scale and room-scale configurations, creating a detailed dataset, ACE-Data-0, encompassing 150 hours of interactions across 200 task categories.

Research Conclusions:

– Evaluations reveal significant limitations in current state-of-the-art methods, particularly under challenging conditions like occlusion and long temporal horizons, highlighting ACE’s potential in supporting scalable advancements in imitation learning and embodied AI systems.

Paper link: https://huggingface.co/papers/2607.28625

35. Flux-OPD: On-Policy Distillation with Evolving Contexts

Keywords: Large language model, evolving contexts, Flux-OPD, open-ended domains, teacher supervision

Category: Natural Language Processing

Research Objective:

– The study aims to develop a mechanism for more effective supervision in large language model training across open-ended domains, addressing the issue of lacking verifiable rewards and formalized task preferences.

Research Methods:

– The research introduces Flux-OPD, a novel OPD paradigm that leverages evolving contexts as a form of in-training supervision. It analyzes the effect of context through the reverse KL objective and uses context-conditioned teachers to provide contextual difference signals.

Research Conclusions:

– Flux-OPD demonstrates superiority over existing OPD paradigms, effectively capturing task preferences and enhancing performance in open-ended tasks by combining evolving contexts with teacher supervision.

Paper link: https://huggingface.co/papers/2607.28022

36. Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Keywords: Agentic Visual Reasoning, Multimodal Large Language Models, Mode Adaptiveness, Tool Effect, Beacon

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance the success rate of multimodal large language models (MLLMs) in complex tasks by rethinking agentic visual reasoning, focusing on Mode Adaptiveness and Tool Effect.

Research Methods:

– Comprehensive analysis to quantify Mode Adaptiveness and Tool Effect in existing models, followed by the proposal of a new model named Beacon, featuring Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion.

Research Conclusions:

– The research concludes that existing models have limited Mode Adaptiveness, and tool benefits are often neutralized by their drawbacks. Beacon, however, shows improved overall performance, notably enhancing Mode Adaptiveness and achieving genuine gains from tool use.

Paper link: https://huggingface.co/papers/2607.28595

37. VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Keywords: Text-to-video, VideoCoCo, Blender program, spatiotemporal process, generative video engine

Category: Generative Models

Research Objective:

– The study aims to address the limitations of text-to-video models in generating physically consistent dynamics by introducing a new framework called VideoCoCo.

Research Methods:

– A dual-engine framework is employed, where a coding agent synthesizes executable Blender code to specify scene evolution and produce deterministic drafts, followed by a generative video engine for photorealistic video transformation.

Research Conclusions:

– VideoCoCo effectively improves performance on benchmarks PhyGenBench and VBench-2.0 by providing an executable, controllable, and inspectable representation for video generation, achieving the best average scores on both benchmarks.

Paper link: https://huggingface.co/papers/2607.27380

38. PhiZero: A World Model Built Around Physical Language

Keywords: PhiZero, physical language, world model, zero-shot motion transfer

Category: Generative Models

Research Objective:

– The primary aim is to develop PhiZero, a model that abstracts predictive structure from visual experience to explicitly reason about physical world evolution using physical language.

Research Methods:

– PhiZero employs a reason-then-render paradigm where it infers future world evolution sequences in physical language and renders transitions into video through self-supervision from in-the-wild videos.

Research Conclusions:

– The experiments demonstrate PhiZero’s capability to model physically coherent world evolution and highlight its potential for realistic interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

Paper link: https://huggingface.co/papers/2607.28624

39. Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Keywords: GUI agents, Qwen-UI-Agent, Reinforcement Learning, MobileWorld, Online RL

Category: AI Systems and Tools

Research Objective:

– The paper aims to advance GUI agents for real-world applications, enabling reliable operation across various devices and platforms, proactive service initiation, and autonomous capability improvements.

Research Methods:

– Introduces Qwen-UI-Agent, utilizing a unified action space that combines GUI operations with CLI execution, and employs an AutoResearch-style data flywheel and Online Reinforcement Learning (RL) for training and advancing agent performance.

Research Conclusions:

– Qwen-UI-Agent demonstrates state-of-the-art performance in mobile-use benchmarks and competitive results in computer- and browser-use tasks, achieving high scores on MobileWorld, AndroidDaily, and WebArena among others.

Paper link: https://huggingface.co/papers/2607.28227

The post AI Native Daily Paper Digest – 20260930 – Metis | Frontis-MA1 | Chimera appeared first on AI Native Foundation.

1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.

Paper link: https://huggingface.co/papers/2607.27958

14. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Keywords: Explorative Modeling, Generative Models, FLOP Efficiency, Sample Efficiency, Parameter Efficiency

Category: Generative Models

Research Objective:

– Introduce Explorative Modeling as a new paradigm in generative modeling to enable end-to-end training by exploring candidate matches and training on the best predictions.

Research Methods:

– The Explorative Models are applied in generative domains, increasing exploration as a pretraining axis to improve performance across diverse datasets including images, video, and language.

Research Conclusions:

– Explorative Modeling enhances efficiency exponentially, achieving significant improvement in FLOP, sample, and parameter efficiency, and enables end-to-end reconstructive generative modeling with reduced inference steps, establishing itself as a new training axis for existing generative models.

Paper link: https://huggingface.co/papers/2607.27372

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Keywords: AI Tour Meeting, Large Language Model, agents, simulation tool

Category: AI Systems and Tools

Research Objective:

– To propose a travel planning framework called AI Tour Meeting using agents powered by Large Language Models (LLMs) to collaboratively create itineraries through natural language discussion.

Research Methods:

– Implementation of a framework for configuring agent personas and orchestrating discussion workflows, enabling monitoring and LLM deployment in the simulation of tour planning discussions.

Research Conclusions:

– The framework is validated by system demonstration and provides analytical insights into the behavior of multiple LLM agents during group travel planning.

Paper link: https://huggingface.co/papers/2607.18806

16. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Keywords: Speculative Decoding, Large Language Model, Truncation-based Verification, Collaborative Verification

Category: Natural Language Processing

Research Objective:

– To analyze the distributions induced by lossy verification schemes in speculative decoding and identify their potential pitfalls in large language model inference.

Research Methods:

– Classified lossy verification methods into truncation-based and collaborative verification, and developed a diagnostic evaluation framework across curated benchmarks.

Research Conclusions:

– Truncation-based methods are susceptible to significant performance degradation due to distributional distortion.

– For collaborative verification, controlling the overshoot of draft probabilities relative to target probabilities is crucial to maintaining output quality.

Paper link: https://huggingface.co/papers/2607.26627

17. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Keywords: Sparse mixture-of-experts, Expert Subspace Separation Index, coherent overlap, next-token prediction

Category: Natural Language Processing

Research Objective:

– The study aims to distinguish the benefits and interactions of experts within sparse mixture-of-experts (MoE) language models by analyzing their geometric arrangement.

Research Methods:

– Utilizes an Expert Subspace Separation Index (ESSI) and various controlled interventions, including matched-route residuals and prefix-controlled factorial setups, to analyze expert overlaps and their impact on representation.

Research Conclusions:

– Coherent overlap indicates that while routes overlap, useful multi-expert computation persists, clarifying why geometric similarity doesn’t alone determine redundancy or pruning value.

– Adding later experts improves next-token predictions significantly in most cases, emphasizing the importance of non-redundant expert selections.

Paper link: https://huggingface.co/papers/2607.28308

18. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Keywords: Fairness Pruning, Large Language Models, Demographic Bias, Causal Bias Localization, AI Ethics and Fairness

Category: AI Ethics and Fairness

Research Objective:

– To introduce and empirically validate Fairness Pruning, a method aimed at managing and mitigating demographic biases in large language models.

Research Methods:

– Utilization of minimally contrastive prompt pairs and inference-time activation capture to identify neurons that react differentially to demographic attributes.

– Empirical evaluation on large models such as Llama-3.2 and Salamandra-2B with structural intervention by zeroing neurons.

Research Conclusions:

– Demonstrated that zeroing specific neurons alters model responses to demographic variables, with a bidirectional bias destabilization effect.

– Confirmed that demographic bias processing and model capabilities are dissociable, setting foundations for transitioning to directional behavior modulation.

Paper link: https://huggingface.co/papers/2607.28319

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Keywords: Speech Emotion Recognition, Edge Devices, Knowledge Distillation, AI Native

Category: Machine Learning

Research Objective:

– This research aims to optimize on-device speech emotion recognition (SER) for real-time applications by developing a more efficient model that can operate effectively on edge devices.

Research Methods:

– Adaptive Multi-teacher Relational Distillation (AMRD) is proposed, using a one-class SVM to assign weights to teacher models and a relational distillation loss to improve inter-sample relational structure alignment between teacher and student models.

Research Conclusions:

– The AMRD approach outperforms single-teacher distillation baselines on the IEMOCAP and CREMA-D datasets, demonstrating its effectiveness across multiple student architectures and confirming that both components of the method yield complementary benefits.

Paper link: https://huggingface.co/papers/2607.25289

20.

Paper link: 

21. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Keywords: Pedestrian Archetypes, Real-world Traffic Scenarios, Behavior Patterns

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces new pedestrian archetypes that exhibit significant behavioral differences, enhancing the understanding of pedestrian behaviors beyond the original taxonomy.

Research Methods:

– Analysis of YouTube dash-cam videos for identifying additional pedestrian archetypes not covered in previous work.

Research Conclusions:

– Seven new pedestrian archetypes are presented, each defined by unique behaviors that are documented through video-frame evidence, thus expanding the existing framework for assessing pedestrian behavior in traffic.

Paper link: https://huggingface.co/papers/2607.16922

22. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Keywords: Vision-Language Models, GPU Memory Constraints, Image-QA Dataset

Category: Multi-Modal Learning

Research Objective:

– Address the challenge of long visual context in vision-language models by introducing ReToken, an explicit retrieval target for selecting query-relevant visual tokens.

Research Methods:

– ReToken is trained on a small image-QA dataset and designed to select a sparse set of visual tokens from a pre-filled visual KV cache.

Research Conclusions:

– ReToken shows consistent performance gains across image and video benchmarks, with significant improvements in Visual Haystacks and LVBench, while maintaining efficiency on a single H100 for both training and long-video inference.

Paper link: https://huggingface.co/papers/2607.28627

23. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Keywords: OmniScope, token compression, query, salience mismatch, AI Native

Category: Multi-Modal Learning

Research Objective:

– The objective is to address the cross-modal salience mismatch in token compression methods for omnimodal large language models by developing a framework named OmniScope that estimates relevance separately for audio and video.

Research Methods:

– Introduced a training-free token compression framework using the query as a shared semantic anchor, allocating modality-specific token budgets, pruning visual tokens with an anchor-delta strategy, and merging audio tokens to reduce redundancy.

Research Conclusions:

– OmniScope outperforms existing methods in accuracy across audio-video benchmarks and model scales, offering improvements in speed and memory efficiency while maintaining high accuracy with minimal drop.

Paper link: https://huggingface.co/papers/2607.23193

24. Harness-G: A Graph-Structured Harness for Search Agents

Keywords: Reinforcement Learning, retrieval aliasing, Harness-G, Structured Non-myopic Credit

Category: Reinforcement Learning

Research Objective:

– The study addresses the problem of retrieval aliasing in reinforcement learning search agents by proposing a new retrieval framework to optimize query generation and interaction.

Research Methods:

– The authors introduce Harness-G, a graph-structured retrieval framework that transforms query generation into finite action selection. They also develop Structured Non-myopic Credit (SNC) for better rewarding actions.

Research Conclusions:

– Harness-G achieves the highest average F1 across six QA benchmarks and outperforms the best baseline by significant margins, demonstrating its effectiveness in reducing linguistic aliasing and improving retrieval accuracy.

Paper link: https://huggingface.co/papers/2607.27652

25. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Keywords: Deep Research agents, Misleading Knowledge, Evidence Synthesis, Workflow Reliability

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to explore the reliability vulnerabilities in Deep Research agents when exposed to misleading knowledge in open information environments.

Research Methods:

– Introduces MisKnow-Agent, a framework for generating misleading knowledge instances to test these vulnerabilities, utilizing 5,933 quality-controlled instances derived from DeepResearch Benchmark tasks.

Research Conclusions:

– The experiments demonstrate that even minimal misleading knowledge can lead to false conclusions in final reports, highlighting a significant reliability issue.

– Despite verifier models identifying misleading instances, they can still be adopted in long-term workflows, indicating a gap in workflow-level evidence verification.

– Evaluated defenses show mitigation effects but fail to fully prevent the adoption of false conclusions, pointing to the need for improved verification and correction capabilities.

Paper link: https://huggingface.co/papers/2607.20891

26. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Keywords: LLM agents, filesystem-based memory, memory organization, search economy, management agent

Category: AI Systems and Tools

Research Objective:

– To systematically explore filesystem-based memory for Large Language Model (LLM) agents, assessing the structure and organization of such memory.

Research Methods:

– The study formalizes the concept of a memory filesystem with a management agent for organization, a search agent for querying, and an execution agent for task trajectory integration.

– Investigated variables include memory shape, stream scale, tool harness, and agent strength, with focus on memory growth, retrieval costs, and store health.

Research Conclusions:

– Organization in filesystem-based memory leads to improved search economy, significantly reducing retrieval costs.

– However, as memory grows, management agents struggle to maintain organization quality, and improvements in organization do not necessarily translate to better answer quality.

– The study identifies the tool set as a significant factor in shaping the memory store, highlighting a design space for agent memory rather than just an assumption.

Paper link: https://huggingface.co/papers/2607.26637

27. Multi-Head Attention Residuals

Keywords: Transformers, Multi-Head Attention Residuals, Nemotron, validation loss

Category: Machine Learning

Research Objective:

– Introduction of Multi-Head Attention Residuals (MHAR) to improve information propagation across model depth and width.

Research Methods:

– MHAR utilizes H per-subspace heads with distinct softmax distributions, trained on a Nemotron-based corpus to optimize validation loss across various model sizes.

Research Conclusions:

– MHAR outperforms standard Transformers with validation loss improvements across 100M, 350M, and 1B scales. Optimal head count H = 8 increases efficiency and maintains near-baseline memory consumption.

Paper link: https://huggingface.co/papers/2607.27230

28. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Keywords: Synthetic Environments, Echoverse, Reinforcement Learning, State Representation

Category: Reinforcement Learning

Research Objective:

– To improve computer-use agents’ training environments by focusing on behavioral depth, targeted interactions, and model improvement.

Research Methods:

– Developed Echoverse to compile specifications into stateful applications, incorporating a co-evolution loop for iterative environment and model enhancement.

Research Conclusions:

– Demonstrated significant model improvement across various environments, with increases in evaluation scores and live-site accuracy via deep environments and iterative environment repairs.

Paper link: https://huggingface.co/papers/2607.28074

29. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Keywords: INTACT, Intent-to-Action, Latent World Models, Action Law Semantics

Category: Reinforcement Learning

Research Objective:

– To introduce INTACT, an end-to-end JEPA, aimed at transforming action-labeled, reward-free trajectories into a deployable intent-to-action interface without expensive test-time search.

Research Methods:

– Utilizes a unique architecture with isomorphic intent representation for local and goal motion-intent models, and provides intact transfer from RGB evidence to action-effective latent intent coordinates.

Research Conclusions:

– The INTACT model achieved up to 100% success rates on LeWM tasks using a search-free policy, significantly reducing the need for candidate sequence sampling while improving performance through a shared four-task encoder. Direct inference is highly efficient, taking only 2.9-5.5 milliseconds.

Paper link: https://huggingface.co/papers/2607.26056

30. MemHarness: Memory Is Reconstructed, Not Replayed

Keywords: MemHarness, Negative Transfer, Context-grounded Guidance, Large Language Model Agents, GRPO

Category: Reinforcement Learning

Research Objective:

– The objective is to develop a framework named MemHarness that enables large language model agents to actively reconstruct past experiences based on the current context, thereby providing more accurate and context-relevant guidance.

Research Methods:

– MemHarness employs a unified policy model that critiques and reconstructs retrieved experiences for each decision step. This reconstructive ability is achieved through end-to-end training using the GRPO technique.

Research Conclusions:

– MemHarness significantly outperforms existing memory-augmented baselines and pure reinforcement learning models, especially in out-of-distribution scenarios, by preventing negative transfer and enhancing intrinsic reasoning capabilities.

Paper link: https://huggingface.co/papers/2607.28272

31. ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

Keywords: ShadowDancer, interactive video, dynamics, action transfer, cross-shadow prediction

Category: Computer Vision

Research Objective:

– To enable precise control of any-action, frame-level dynamics in interactive video world models through novel techniques.

Research Methods:

– Utilizes shadow pairs to construct scalable video pairs replaying the same dynamics with varied appearances.

– Implements cross-shadow prediction to learn actions by predicting dynamics from one shadow to another.

Research Conclusions:

– Demonstrated improved action transfer and extended action rollout in diverse dynamics families, achieving an average blinded win rate of 86% over existing models.

Paper link: https://huggingface.co/papers/2607.28362

32. See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Keywords: Multimodal large language models, Visual reasoning, See2Think, Visual Action-of-Thought

Category: Multi-Modal Learning

Research Objective:

– To introduce See2Think, an evaluation framework that assesses visual reasoning in multimodal large language models through benchmarks like See2ThinkBench and Visual Action-of-Thought.

Research Methods:

– The creation of See2ThinkBench with 1,200 visually dependent tasks across 12 categories and the use of Visual Action-of-Thought to record and analyze visual and textual operations under controlled settings.

Research Conclusions:

– Visual reasoning capabilities in models are strongly dependent on both the model and environment, showing varied performance across tasks. Additionally, even though models often select the correct visual operations, the accuracy improvements are hindered by challenges in rendering fidelity and reliance on feedback which does not always improve accuracy.

Paper link: https://huggingface.co/papers/2607.26769

33. RefCaptioner: Multi-Reference Image-Grounded Video Captioning

Keywords: multi-reference image-grounded video captioning, RefCaptioner, factual video descriptions, Hierarchical Coverage-Discounted GRPO, AI-generated videos

Category: Multi-Modal Learning

Research Objective:

– To introduce and address the task of multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level reference grounding.

Research Methods:

– Development of RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO to enhance reference selection and cross-reference consistency.

Research Conclusions:

– RefCaptioner outperforms other open-source models in overall performance, yielding captions preferred by human annotators and allowing for more source-faithful video reconstruction.

Paper link: https://huggingface.co/papers/2607.28509

34. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Keywords: Embodied intelligence, Ambient Capture Engine, AI Native, Multi-modal datasets, Human-centric data

Category: Robotics and Autonomous Systems

Research Objective:

– The study focuses on addressing the data bottleneck in embodied intelligence by introducing the Ambient Capture Engine (ACE), which captures comprehensive sensory data in real home environments.

Research Methods:

– The authors developed ACE to record multisensory data through a combination of table-scale and room-scale configurations, creating a detailed dataset, ACE-Data-0, encompassing 150 hours of interactions across 200 task categories.

Research Conclusions:

– Evaluations reveal significant limitations in current state-of-the-art methods, particularly under challenging conditions like occlusion and long temporal horizons, highlighting ACE’s potential in supporting scalable advancements in imitation learning and embodied AI systems.

Paper link: https://huggingface.co/papers/2607.28625

35. Flux-OPD: On-Policy Distillation with Evolving Contexts

Keywords: Large language model, evolving contexts, Flux-OPD, open-ended domains, teacher supervision

Category: Natural Language Processing

Research Objective:

– The study aims to develop a mechanism for more effective supervision in large language model training across open-ended domains, addressing the issue of lacking verifiable rewards and formalized task preferences.

Research Methods:

– The research introduces Flux-OPD, a novel OPD paradigm that leverages evolving contexts as a form of in-training supervision. It analyzes the effect of context through the reverse KL objective and uses context-conditioned teachers to provide contextual difference signals.

Research Conclusions:

– Flux-OPD demonstrates superiority over existing OPD paradigms, effectively capturing task preferences and enhancing performance in open-ended tasks by combining evolving contexts with teacher supervision.

Paper link: https://huggingface.co/papers/2607.28022

36. Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Keywords: Agentic Visual Reasoning, Multimodal Large Language Models, Mode Adaptiveness, Tool Effect, Beacon

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance the success rate of multimodal large language models (MLLMs) in complex tasks by rethinking agentic visual reasoning, focusing on Mode Adaptiveness and Tool Effect.

Research Methods:

– Comprehensive analysis to quantify Mode Adaptiveness and Tool Effect in existing models, followed by the proposal of a new model named Beacon, featuring Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion.

Research Conclusions:

– The research concludes that existing models have limited Mode Adaptiveness, and tool benefits are often neutralized by their drawbacks. Beacon, however, shows improved overall performance, notably enhancing Mode Adaptiveness and achieving genuine gains from tool use.

Paper link: https://huggingface.co/papers/2607.28595

37. VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Keywords: Text-to-video, VideoCoCo, Blender program, spatiotemporal process, generative video engine

Category: Generative Models

Research Objective:

– The study aims to address the limitations of text-to-video models in generating physically consistent dynamics by introducing a new framework called VideoCoCo.

Research Methods:

– A dual-engine framework is employed, where a coding agent synthesizes executable Blender code to specify scene evolution and produce deterministic drafts, followed by a generative video engine for photorealistic video transformation.

Research Conclusions:

– VideoCoCo effectively improves performance on benchmarks PhyGenBench and VBench-2.0 by providing an executable, controllable, and inspectable representation for video generation, achieving the best average scores on both benchmarks.

Paper link: https://huggingface.co/papers/2607.27380

38. PhiZero: A World Model Built Around Physical Language

Keywords: PhiZero, physical language, world model, zero-shot motion transfer

Category: Generative Models

Research Objective:

– The primary aim is to develop PhiZero, a model that abstracts predictive structure from visual experience to explicitly reason about physical world evolution using physical language.

Research Methods:

– PhiZero employs a reason-then-render paradigm where it infers future world evolution sequences in physical language and renders transitions into video through self-supervision from in-the-wild videos.

Research Conclusions:

– The experiments demonstrate PhiZero’s capability to model physically coherent world evolution and highlight its potential for realistic interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

Paper link: https://huggingface.co/papers/2607.28624

39. Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Keywords: GUI agents, Qwen-UI-Agent, Reinforcement Learning, MobileWorld, Online RL

Category: AI Systems and Tools

Research Objective:

– The paper aims to advance GUI agents for real-world applications, enabling reliable operation across various devices and platforms, proactive service initiation, and autonomous capability improvements.

Research Methods:

– Introduces Qwen-UI-Agent, utilizing a unified action space that combines GUI operations with CLI execution, and employs an AutoResearch-style data flywheel and Online Reinforcement Learning (RL) for training and advancing agent performance.

Research Conclusions:

– Qwen-UI-Agent demonstrates state-of-the-art performance in mobile-use benchmarks and competitive results in computer- and browser-use tasks, achieving high scores on MobileWorld, AndroidDaily, and WebArena among others.

Paper link: https://huggingface.co/papers/2607.28227

The post AI Native Daily Paper Digest – 20260929 – AskChem | Metis | β-OPSD appeared first on AI Native Foundation.

1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.

Paper link: https://huggingface.co/papers/2607.27958

14. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Keywords: Explorative Modeling, Generative Models, FLOP Efficiency, Sample Efficiency, Parameter Efficiency

Category: Generative Models

Research Objective:

– Introduce Explorative Modeling as a new paradigm in generative modeling to enable end-to-end training by exploring candidate matches and training on the best predictions.

Research Methods:

– The Explorative Models are applied in generative domains, increasing exploration as a pretraining axis to improve performance across diverse datasets including images, video, and language.

Research Conclusions:

– Explorative Modeling enhances efficiency exponentially, achieving significant improvement in FLOP, sample, and parameter efficiency, and enables end-to-end reconstructive generative modeling with reduced inference steps, establishing itself as a new training axis for existing generative models.

Paper link: https://huggingface.co/papers/2607.27372

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Keywords: AI Tour Meeting, Large Language Model, agents, simulation tool

Category: AI Systems and Tools

Research Objective:

– To propose a travel planning framework called AI Tour Meeting using agents powered by Large Language Models (LLMs) to collaboratively create itineraries through natural language discussion.

Research Methods:

– Implementation of a framework for configuring agent personas and orchestrating discussion workflows, enabling monitoring and LLM deployment in the simulation of tour planning discussions.

Research Conclusions:

– The framework is validated by system demonstration and provides analytical insights into the behavior of multiple LLM agents during group travel planning.

Paper link: https://huggingface.co/papers/2607.18806

16. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Keywords: Speculative Decoding, Large Language Model, Truncation-based Verification, Collaborative Verification

Category: Natural Language Processing

Research Objective:

– To analyze the distributions induced by lossy verification schemes in speculative decoding and identify their potential pitfalls in large language model inference.

Research Methods:

– Classified lossy verification methods into truncation-based and collaborative verification, and developed a diagnostic evaluation framework across curated benchmarks.

Research Conclusions:

– Truncation-based methods are susceptible to significant performance degradation due to distributional distortion.

– For collaborative verification, controlling the overshoot of draft probabilities relative to target probabilities is crucial to maintaining output quality.

Paper link: https://huggingface.co/papers/2607.26627

17. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Keywords: Sparse mixture-of-experts, Expert Subspace Separation Index, coherent overlap, next-token prediction

Category: Natural Language Processing

Research Objective:

– The study aims to distinguish the benefits and interactions of experts within sparse mixture-of-experts (MoE) language models by analyzing their geometric arrangement.

Research Methods:

– Utilizes an Expert Subspace Separation Index (ESSI) and various controlled interventions, including matched-route residuals and prefix-controlled factorial setups, to analyze expert overlaps and their impact on representation.

Research Conclusions:

– Coherent overlap indicates that while routes overlap, useful multi-expert computation persists, clarifying why geometric similarity doesn’t alone determine redundancy or pruning value.

– Adding later experts improves next-token predictions significantly in most cases, emphasizing the importance of non-redundant expert selections.

Paper link: https://huggingface.co/papers/2607.28308

18. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Keywords: Fairness Pruning, Large Language Models, Demographic Bias, Causal Bias Localization, AI Ethics and Fairness

Category: AI Ethics and Fairness

Research Objective:

– To introduce and empirically validate Fairness Pruning, a method aimed at managing and mitigating demographic biases in large language models.

Research Methods:

– Utilization of minimally contrastive prompt pairs and inference-time activation capture to identify neurons that react differentially to demographic attributes.

– Empirical evaluation on large models such as Llama-3.2 and Salamandra-2B with structural intervention by zeroing neurons.

Research Conclusions:

– Demonstrated that zeroing specific neurons alters model responses to demographic variables, with a bidirectional bias destabilization effect.

– Confirmed that demographic bias processing and model capabilities are dissociable, setting foundations for transitioning to directional behavior modulation.

Paper link: https://huggingface.co/papers/2607.28319

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Keywords: Speech Emotion Recognition, Edge Devices, Knowledge Distillation, AI Native

Category: Machine Learning

Research Objective:

– This research aims to optimize on-device speech emotion recognition (SER) for real-time applications by developing a more efficient model that can operate effectively on edge devices.

Research Methods:

– Adaptive Multi-teacher Relational Distillation (AMRD) is proposed, using a one-class SVM to assign weights to teacher models and a relational distillation loss to improve inter-sample relational structure alignment between teacher and student models.

Research Conclusions:

– The AMRD approach outperforms single-teacher distillation baselines on the IEMOCAP and CREMA-D datasets, demonstrating its effectiveness across multiple student architectures and confirming that both components of the method yield complementary benefits.

Paper link: https://huggingface.co/papers/2607.25289

20.

Paper link: 

21. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Keywords: Pedestrian Archetypes, Real-world Traffic Scenarios, Behavior Patterns

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces new pedestrian archetypes that exhibit significant behavioral differences, enhancing the understanding of pedestrian behaviors beyond the original taxonomy.

Research Methods:

– Analysis of YouTube dash-cam videos for identifying additional pedestrian archetypes not covered in previous work.

Research Conclusions:

– Seven new pedestrian archetypes are presented, each defined by unique behaviors that are documented through video-frame evidence, thus expanding the existing framework for assessing pedestrian behavior in traffic.

Paper link: https://huggingface.co/papers/2607.16922

22. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Keywords: Vision-Language Models, GPU Memory Constraints, Image-QA Dataset

Category: Multi-Modal Learning

Research Objective:

– Address the challenge of long visual context in vision-language models by introducing ReToken, an explicit retrieval target for selecting query-relevant visual tokens.

Research Methods:

– ReToken is trained on a small image-QA dataset and designed to select a sparse set of visual tokens from a pre-filled visual KV cache.

Research Conclusions:

– ReToken shows consistent performance gains across image and video benchmarks, with significant improvements in Visual Haystacks and LVBench, while maintaining efficiency on a single H100 for both training and long-video inference.

Paper link: https://huggingface.co/papers/2607.28627

23. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Keywords: OmniScope, token compression, query, salience mismatch, AI Native

Category: Multi-Modal Learning

Research Objective:

– The objective is to address the cross-modal salience mismatch in token compression methods for omnimodal large language models by developing a framework named OmniScope that estimates relevance separately for audio and video.

Research Methods:

– Introduced a training-free token compression framework using the query as a shared semantic anchor, allocating modality-specific token budgets, pruning visual tokens with an anchor-delta strategy, and merging audio tokens to reduce redundancy.

Research Conclusions:

– OmniScope outperforms existing methods in accuracy across audio-video benchmarks and model scales, offering improvements in speed and memory efficiency while maintaining high accuracy with minimal drop.

Paper link: https://huggingface.co/papers/2607.23193

24. Harness-G: A Graph-Structured Harness for Search Agents

Keywords: Reinforcement Learning, retrieval aliasing, Harness-G, Structured Non-myopic Credit

Category: Reinforcement Learning

Research Objective:

– The study addresses the problem of retrieval aliasing in reinforcement learning search agents by proposing a new retrieval framework to optimize query generation and interaction.

Research Methods:

– The authors introduce Harness-G, a graph-structured retrieval framework that transforms query generation into finite action selection. They also develop Structured Non-myopic Credit (SNC) for better rewarding actions.

Research Conclusions:

– Harness-G achieves the highest average F1 across six QA benchmarks and outperforms the best baseline by significant margins, demonstrating its effectiveness in reducing linguistic aliasing and improving retrieval accuracy.

Paper link: https://huggingface.co/papers/2607.27652

25. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Keywords: Deep Research agents, Misleading Knowledge, Evidence Synthesis, Workflow Reliability

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to explore the reliability vulnerabilities in Deep Research agents when exposed to misleading knowledge in open information environments.

Research Methods:

– Introduces MisKnow-Agent, a framework for generating misleading knowledge instances to test these vulnerabilities, utilizing 5,933 quality-controlled instances derived from DeepResearch Benchmark tasks.

Research Conclusions:

– The experiments demonstrate that even minimal misleading knowledge can lead to false conclusions in final reports, highlighting a significant reliability issue.

– Despite verifier models identifying misleading instances, they can still be adopted in long-term workflows, indicating a gap in workflow-level evidence verification.

– Evaluated defenses show mitigation effects but fail to fully prevent the adoption of false conclusions, pointing to the need for improved verification and correction capabilities.

Paper link: https://huggingface.co/papers/2607.20891

26. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Keywords: LLM agents, filesystem-based memory, memory organization, search economy, management agent

Category: AI Systems and Tools

Research Objective:

– To systematically explore filesystem-based memory for Large Language Model (LLM) agents, assessing the structure and organization of such memory.

Research Methods:

– The study formalizes the concept of a memory filesystem with a management agent for organization, a search agent for querying, and an execution agent for task trajectory integration.

– Investigated variables include memory shape, stream scale, tool harness, and agent strength, with focus on memory growth, retrieval costs, and store health.

Research Conclusions:

– Organization in filesystem-based memory leads to improved search economy, significantly reducing retrieval costs.

– However, as memory grows, management agents struggle to maintain organization quality, and improvements in organization do not necessarily translate to better answer quality.

– The study identifies the tool set as a significant factor in shaping the memory store, highlighting a design space for agent memory rather than just an assumption.

Paper link: https://huggingface.co/papers/2607.26637

27. Multi-Head Attention Residuals

Keywords: Transformers, Multi-Head Attention Residuals, Nemotron, validation loss

Category: Machine Learning

Research Objective:

– Introduction of Multi-Head Attention Residuals (MHAR) to improve information propagation across model depth and width.

Research Methods:

– MHAR utilizes H per-subspace heads with distinct softmax distributions, trained on a Nemotron-based corpus to optimize validation loss across various model sizes.

Research Conclusions:

– MHAR outperforms standard Transformers with validation loss improvements across 100M, 350M, and 1B scales. Optimal head count H = 8 increases efficiency and maintains near-baseline memory consumption.

Paper link: https://huggingface.co/papers/2607.27230

28. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Keywords: Synthetic Environments, Echoverse, Reinforcement Learning, State Representation

Category: Reinforcement Learning

Research Objective:

– To improve computer-use agents’ training environments by focusing on behavioral depth, targeted interactions, and model improvement.

Research Methods:

– Developed Echoverse to compile specifications into stateful applications, incorporating a co-evolution loop for iterative environment and model enhancement.

Research Conclusions:

– Demonstrated significant model improvement across various environments, with increases in evaluation scores and live-site accuracy via deep environments and iterative environment repairs.

Paper link: https://huggingface.co/papers/2607.28074

29. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Keywords: INTACT, Intent-to-Action, Latent World Models, Action Law Semantics

Category: Reinforcement Learning

Research Objective:

– To introduce INTACT, an end-to-end JEPA, aimed at transforming action-labeled, reward-free trajectories into a deployable intent-to-action interface without expensive test-time search.

Research Methods:

– Utilizes a unique architecture with isomorphic intent representation for local and goal motion-intent models, and provides intact transfer from RGB evidence to action-effective latent intent coordinates.

Research Conclusions:

– The INTACT model achieved up to 100% success rates on LeWM tasks using a search-free policy, significantly reducing the need for candidate sequence sampling while improving performance through a shared four-task encoder. Direct inference is highly efficient, taking only 2.9-5.5 milliseconds.

Paper link: https://huggingface.co/papers/2607.26056

30. MemHarness: Memory Is Reconstructed, Not Replayed

Keywords: MemHarness, Negative Transfer, Context-grounded Guidance, Large Language Model Agents, GRPO

Category: Reinforcement Learning

Research Objective:

– The objective is to develop a framework named MemHarness that enables large language model agents to actively reconstruct past experiences based on the current context, thereby providing more accurate and context-relevant guidance.

Research Methods:

– MemHarness employs a unified policy model that critiques and reconstructs retrieved experiences for each decision step. This reconstructive ability is achieved through end-to-end training using the GRPO technique.

Research Conclusions:

– MemHarness significantly outperforms existing memory-augmented baselines and pure reinforcement learning models, especially in out-of-distribution scenarios, by preventing negative transfer and enhancing intrinsic reasoning capabilities.

Paper link: https://huggingface.co/papers/2607.28272

31. ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

Keywords: ShadowDancer, interactive video, dynamics, action transfer, cross-shadow prediction

Category: Computer Vision

Research Objective:

– To enable precise control of any-action, frame-level dynamics in interactive video world models through novel techniques.

Research Methods:

– Utilizes shadow pairs to construct scalable video pairs replaying the same dynamics with varied appearances.

– Implements cross-shadow prediction to learn actions by predicting dynamics from one shadow to another.

Research Conclusions:

– Demonstrated improved action transfer and extended action rollout in diverse dynamics families, achieving an average blinded win rate of 86% over existing models.

Paper link: https://huggingface.co/papers/2607.28362

32. See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Keywords: Multimodal large language models, Visual reasoning, See2Think, Visual Action-of-Thought

Category: Multi-Modal Learning

Research Objective:

– To introduce See2Think, an evaluation framework that assesses visual reasoning in multimodal large language models through benchmarks like See2ThinkBench and Visual Action-of-Thought.

Research Methods:

– The creation of See2ThinkBench with 1,200 visually dependent tasks across 12 categories and the use of Visual Action-of-Thought to record and analyze visual and textual operations under controlled settings.

Research Conclusions:

– Visual reasoning capabilities in models are strongly dependent on both the model and environment, showing varied performance across tasks. Additionally, even though models often select the correct visual operations, the accuracy improvements are hindered by challenges in rendering fidelity and reliance on feedback which does not always improve accuracy.

Paper link: https://huggingface.co/papers/2607.26769

33. RefCaptioner: Multi-Reference Image-Grounded Video Captioning

Keywords: multi-reference image-grounded video captioning, RefCaptioner, factual video descriptions, Hierarchical Coverage-Discounted GRPO, AI-generated videos

Category: Multi-Modal Learning

Research Objective:

– To introduce and address the task of multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level reference grounding.

Research Methods:

– Development of RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO to enhance reference selection and cross-reference consistency.

Research Conclusions:

– RefCaptioner outperforms other open-source models in overall performance, yielding captions preferred by human annotators and allowing for more source-faithful video reconstruction.

Paper link: https://huggingface.co/papers/2607.28509

34. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Keywords: Embodied intelligence, Ambient Capture Engine, AI Native, Multi-modal datasets, Human-centric data

Category: Robotics and Autonomous Systems

Research Objective:

– The study focuses on addressing the data bottleneck in embodied intelligence by introducing the Ambient Capture Engine (ACE), which captures comprehensive sensory data in real home environments.

Research Methods:

– The authors developed ACE to record multisensory data through a combination of table-scale and room-scale configurations, creating a detailed dataset, ACE-Data-0, encompassing 150 hours of interactions across 200 task categories.

Research Conclusions:

– Evaluations reveal significant limitations in current state-of-the-art methods, particularly under challenging conditions like occlusion and long temporal horizons, highlighting ACE’s potential in supporting scalable advancements in imitation learning and embodied AI systems.

Paper link: https://huggingface.co/papers/2607.28625

35. Flux-OPD: On-Policy Distillation with Evolving Contexts

Keywords: Large language model, evolving contexts, Flux-OPD, open-ended domains, teacher supervision

Category: Natural Language Processing

Research Objective:

– The study aims to develop a mechanism for more effective supervision in large language model training across open-ended domains, addressing the issue of lacking verifiable rewards and formalized task preferences.

Research Methods:

– The research introduces Flux-OPD, a novel OPD paradigm that leverages evolving contexts as a form of in-training supervision. It analyzes the effect of context through the reverse KL objective and uses context-conditioned teachers to provide contextual difference signals.

Research Conclusions:

– Flux-OPD demonstrates superiority over existing OPD paradigms, effectively capturing task preferences and enhancing performance in open-ended tasks by combining evolving contexts with teacher supervision.

Paper link: https://huggingface.co/papers/2607.28022

36. Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Keywords: Agentic Visual Reasoning, Multimodal Large Language Models, Mode Adaptiveness, Tool Effect, Beacon

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance the success rate of multimodal large language models (MLLMs) in complex tasks by rethinking agentic visual reasoning, focusing on Mode Adaptiveness and Tool Effect.

Research Methods:

– Comprehensive analysis to quantify Mode Adaptiveness and Tool Effect in existing models, followed by the proposal of a new model named Beacon, featuring Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion.

Research Conclusions:

– The research concludes that existing models have limited Mode Adaptiveness, and tool benefits are often neutralized by their drawbacks. Beacon, however, shows improved overall performance, notably enhancing Mode Adaptiveness and achieving genuine gains from tool use.

Paper link: https://huggingface.co/papers/2607.28595

37. VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Keywords: Text-to-video, VideoCoCo, Blender program, spatiotemporal process, generative video engine

Category: Generative Models

Research Objective:

– The study aims to address the limitations of text-to-video models in generating physically consistent dynamics by introducing a new framework called VideoCoCo.

Research Methods:

– A dual-engine framework is employed, where a coding agent synthesizes executable Blender code to specify scene evolution and produce deterministic drafts, followed by a generative video engine for photorealistic video transformation.

Research Conclusions:

– VideoCoCo effectively improves performance on benchmarks PhyGenBench and VBench-2.0 by providing an executable, controllable, and inspectable representation for video generation, achieving the best average scores on both benchmarks.

Paper link: https://huggingface.co/papers/2607.27380

38. PhiZero: A World Model Built Around Physical Language

Keywords: PhiZero, physical language, world model, zero-shot motion transfer

Category: Generative Models

Research Objective:

– The primary aim is to develop PhiZero, a model that abstracts predictive structure from visual experience to explicitly reason about physical world evolution using physical language.

Research Methods:

– PhiZero employs a reason-then-render paradigm where it infers future world evolution sequences in physical language and renders transitions into video through self-supervision from in-the-wild videos.

Research Conclusions:

– The experiments demonstrate PhiZero’s capability to model physically coherent world evolution and highlight its potential for realistic interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

Paper link: https://huggingface.co/papers/2607.28624

39. Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Keywords: GUI agents, Qwen-UI-Agent, Reinforcement Learning, MobileWorld, Online RL

Category: AI Systems and Tools

Research Objective:

– The paper aims to advance GUI agents for real-world applications, enabling reliable operation across various devices and platforms, proactive service initiation, and autonomous capability improvements.

Research Methods:

– Introduces Qwen-UI-Agent, utilizing a unified action space that combines GUI operations with CLI execution, and employs an AutoResearch-style data flywheel and Online Reinforcement Learning (RL) for training and advancing agent performance.

Research Conclusions:

– Qwen-UI-Agent demonstrates state-of-the-art performance in mobile-use benchmarks and competitive results in computer- and browser-use tasks, achieving high scores on MobileWorld, AndroidDaily, and WebArena among others.

Paper link: https://huggingface.co/papers/2607.28227

The post AI Native Daily Paper Digest – 20260928 – AskChem | Metis | BM25 appeared first on AI Native Foundation.

1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.

Paper link: https://huggingface.co/papers/2607.27958

14. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Keywords: Explorative Modeling, Generative Models, FLOP Efficiency, Sample Efficiency, Parameter Efficiency

Category: Generative Models

Research Objective:

– Introduce Explorative Modeling as a new paradigm in generative modeling to enable end-to-end training by exploring candidate matches and training on the best predictions.

Research Methods:

– The Explorative Models are applied in generative domains, increasing exploration as a pretraining axis to improve performance across diverse datasets including images, video, and language.

Research Conclusions:

– Explorative Modeling enhances efficiency exponentially, achieving significant improvement in FLOP, sample, and parameter efficiency, and enables end-to-end reconstructive generative modeling with reduced inference steps, establishing itself as a new training axis for existing generative models.

Paper link: https://huggingface.co/papers/2607.27372

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Keywords: AI Tour Meeting, Large Language Model, agents, simulation tool

Category: AI Systems and Tools

Research Objective:

– To propose a travel planning framework called AI Tour Meeting using agents powered by Large Language Models (LLMs) to collaboratively create itineraries through natural language discussion.

Research Methods:

– Implementation of a framework for configuring agent personas and orchestrating discussion workflows, enabling monitoring and LLM deployment in the simulation of tour planning discussions.

Research Conclusions:

– The framework is validated by system demonstration and provides analytical insights into the behavior of multiple LLM agents during group travel planning.

Paper link: https://huggingface.co/papers/2607.18806

16. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Keywords: Speculative Decoding, Large Language Model, Truncation-based Verification, Collaborative Verification

Category: Natural Language Processing

Research Objective:

– To analyze the distributions induced by lossy verification schemes in speculative decoding and identify their potential pitfalls in large language model inference.

Research Methods:

– Classified lossy verification methods into truncation-based and collaborative verification, and developed a diagnostic evaluation framework across curated benchmarks.

Research Conclusions:

– Truncation-based methods are susceptible to significant performance degradation due to distributional distortion.

– For collaborative verification, controlling the overshoot of draft probabilities relative to target probabilities is crucial to maintaining output quality.

Paper link: https://huggingface.co/papers/2607.26627

17. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Keywords: Sparse mixture-of-experts, Expert Subspace Separation Index, coherent overlap, next-token prediction

Category: Natural Language Processing

Research Objective:

– The study aims to distinguish the benefits and interactions of experts within sparse mixture-of-experts (MoE) language models by analyzing their geometric arrangement.

Research Methods:

– Utilizes an Expert Subspace Separation Index (ESSI) and various controlled interventions, including matched-route residuals and prefix-controlled factorial setups, to analyze expert overlaps and their impact on representation.

Research Conclusions:

– Coherent overlap indicates that while routes overlap, useful multi-expert computation persists, clarifying why geometric similarity doesn’t alone determine redundancy or pruning value.

– Adding later experts improves next-token predictions significantly in most cases, emphasizing the importance of non-redundant expert selections.

Paper link: https://huggingface.co/papers/2607.28308

18. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Keywords: Fairness Pruning, Large Language Models, Demographic Bias, Causal Bias Localization, AI Ethics and Fairness

Category: AI Ethics and Fairness

Research Objective:

– To introduce and empirically validate Fairness Pruning, a method aimed at managing and mitigating demographic biases in large language models.

Research Methods:

– Utilization of minimally contrastive prompt pairs and inference-time activation capture to identify neurons that react differentially to demographic attributes.

– Empirical evaluation on large models such as Llama-3.2 and Salamandra-2B with structural intervention by zeroing neurons.

Research Conclusions:

– Demonstrated that zeroing specific neurons alters model responses to demographic variables, with a bidirectional bias destabilization effect.

– Confirmed that demographic bias processing and model capabilities are dissociable, setting foundations for transitioning to directional behavior modulation.

Paper link: https://huggingface.co/papers/2607.28319

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Keywords: Speech Emotion Recognition, Edge Devices, Knowledge Distillation, AI Native

Category: Machine Learning

Research Objective:

– This research aims to optimize on-device speech emotion recognition (SER) for real-time applications by developing a more efficient model that can operate effectively on edge devices.

Research Methods:

– Adaptive Multi-teacher Relational Distillation (AMRD) is proposed, using a one-class SVM to assign weights to teacher models and a relational distillation loss to improve inter-sample relational structure alignment between teacher and student models.

Research Conclusions:

– The AMRD approach outperforms single-teacher distillation baselines on the IEMOCAP and CREMA-D datasets, demonstrating its effectiveness across multiple student architectures and confirming that both components of the method yield complementary benefits.

Paper link: https://huggingface.co/papers/2607.25289

20.

Paper link: 

21. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Keywords: Pedestrian Archetypes, Real-world Traffic Scenarios, Behavior Patterns

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces new pedestrian archetypes that exhibit significant behavioral differences, enhancing the understanding of pedestrian behaviors beyond the original taxonomy.

Research Methods:

– Analysis of YouTube dash-cam videos for identifying additional pedestrian archetypes not covered in previous work.

Research Conclusions:

– Seven new pedestrian archetypes are presented, each defined by unique behaviors that are documented through video-frame evidence, thus expanding the existing framework for assessing pedestrian behavior in traffic.

Paper link: https://huggingface.co/papers/2607.16922

22. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Keywords: Vision-Language Models, GPU Memory Constraints, Image-QA Dataset

Category: Multi-Modal Learning

Research Objective:

– Address the challenge of long visual context in vision-language models by introducing ReToken, an explicit retrieval target for selecting query-relevant visual tokens.

Research Methods:

– ReToken is trained on a small image-QA dataset and designed to select a sparse set of visual tokens from a pre-filled visual KV cache.

Research Conclusions:

– ReToken shows consistent performance gains across image and video benchmarks, with significant improvements in Visual Haystacks and LVBench, while maintaining efficiency on a single H100 for both training and long-video inference.

Paper link: https://huggingface.co/papers/2607.28627

23. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Keywords: OmniScope, token compression, query, salience mismatch, AI Native

Category: Multi-Modal Learning

Research Objective:

– The objective is to address the cross-modal salience mismatch in token compression methods for omnimodal large language models by developing a framework named OmniScope that estimates relevance separately for audio and video.

Research Methods:

– Introduced a training-free token compression framework using the query as a shared semantic anchor, allocating modality-specific token budgets, pruning visual tokens with an anchor-delta strategy, and merging audio tokens to reduce redundancy.

Research Conclusions:

– OmniScope outperforms existing methods in accuracy across audio-video benchmarks and model scales, offering improvements in speed and memory efficiency while maintaining high accuracy with minimal drop.

Paper link: https://huggingface.co/papers/2607.23193

24. Harness-G: A Graph-Structured Harness for Search Agents

Keywords: Reinforcement Learning, retrieval aliasing, Harness-G, Structured Non-myopic Credit

Category: Reinforcement Learning

Research Objective:

– The study addresses the problem of retrieval aliasing in reinforcement learning search agents by proposing a new retrieval framework to optimize query generation and interaction.

Research Methods:

– The authors introduce Harness-G, a graph-structured retrieval framework that transforms query generation into finite action selection. They also develop Structured Non-myopic Credit (SNC) for better rewarding actions.

Research Conclusions:

– Harness-G achieves the highest average F1 across six QA benchmarks and outperforms the best baseline by significant margins, demonstrating its effectiveness in reducing linguistic aliasing and improving retrieval accuracy.

Paper link: https://huggingface.co/papers/2607.27652

25. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Keywords: Deep Research agents, Misleading Knowledge, Evidence Synthesis, Workflow Reliability

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to explore the reliability vulnerabilities in Deep Research agents when exposed to misleading knowledge in open information environments.

Research Methods:

– Introduces MisKnow-Agent, a framework for generating misleading knowledge instances to test these vulnerabilities, utilizing 5,933 quality-controlled instances derived from DeepResearch Benchmark tasks.

Research Conclusions:

– The experiments demonstrate that even minimal misleading knowledge can lead to false conclusions in final reports, highlighting a significant reliability issue.

– Despite verifier models identifying misleading instances, they can still be adopted in long-term workflows, indicating a gap in workflow-level evidence verification.

– Evaluated defenses show mitigation effects but fail to fully prevent the adoption of false conclusions, pointing to the need for improved verification and correction capabilities.

Paper link: https://huggingface.co/papers/2607.20891

26. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Keywords: LLM agents, filesystem-based memory, memory organization, search economy, management agent

Category: AI Systems and Tools

Research Objective:

– To systematically explore filesystem-based memory for Large Language Model (LLM) agents, assessing the structure and organization of such memory.

Research Methods:

– The study formalizes the concept of a memory filesystem with a management agent for organization, a search agent for querying, and an execution agent for task trajectory integration.

– Investigated variables include memory shape, stream scale, tool harness, and agent strength, with focus on memory growth, retrieval costs, and store health.

Research Conclusions:

– Organization in filesystem-based memory leads to improved search economy, significantly reducing retrieval costs.

– However, as memory grows, management agents struggle to maintain organization quality, and improvements in organization do not necessarily translate to better answer quality.

– The study identifies the tool set as a significant factor in shaping the memory store, highlighting a design space for agent memory rather than just an assumption.

Paper link: https://huggingface.co/papers/2607.26637

27. Multi-Head Attention Residuals

Keywords: Transformers, Multi-Head Attention Residuals, Nemotron, validation loss

Category: Machine Learning

Research Objective:

– Introduction of Multi-Head Attention Residuals (MHAR) to improve information propagation across model depth and width.

Research Methods:

– MHAR utilizes H per-subspace heads with distinct softmax distributions, trained on a Nemotron-based corpus to optimize validation loss across various model sizes.

Research Conclusions:

– MHAR outperforms standard Transformers with validation loss improvements across 100M, 350M, and 1B scales. Optimal head count H = 8 increases efficiency and maintains near-baseline memory consumption.

Paper link: https://huggingface.co/papers/2607.27230

28. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Keywords: Synthetic Environments, Echoverse, Reinforcement Learning, State Representation

Category: Reinforcement Learning

Research Objective:

– To improve computer-use agents’ training environments by focusing on behavioral depth, targeted interactions, and model improvement.

Research Methods:

– Developed Echoverse to compile specifications into stateful applications, incorporating a co-evolution loop for iterative environment and model enhancement.

Research Conclusions:

– Demonstrated significant model improvement across various environments, with increases in evaluation scores and live-site accuracy via deep environments and iterative environment repairs.

Paper link: https://huggingface.co/papers/2607.28074

29. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Keywords: INTACT, Intent-to-Action, Latent World Models, Action Law Semantics

Category: Reinforcement Learning

Research Objective:

– To introduce INTACT, an end-to-end JEPA, aimed at transforming action-labeled, reward-free trajectories into a deployable intent-to-action interface without expensive test-time search.

Research Methods:

– Utilizes a unique architecture with isomorphic intent representation for local and goal motion-intent models, and provides intact transfer from RGB evidence to action-effective latent intent coordinates.

Research Conclusions:

– The INTACT model achieved up to 100% success rates on LeWM tasks using a search-free policy, significantly reducing the need for candidate sequence sampling while improving performance through a shared four-task encoder. Direct inference is highly efficient, taking only 2.9-5.5 milliseconds.

Paper link: https://huggingface.co/papers/2607.26056

30. MemHarness: Memory Is Reconstructed, Not Replayed

Keywords: MemHarness, Negative Transfer, Context-grounded Guidance, Large Language Model Agents, GRPO

Category: Reinforcement Learning

Research Objective:

– The objective is to develop a framework named MemHarness that enables large language model agents to actively reconstruct past experiences based on the current context, thereby providing more accurate and context-relevant guidance.

Research Methods:

– MemHarness employs a unified policy model that critiques and reconstructs retrieved experiences for each decision step. This reconstructive ability is achieved through end-to-end training using the GRPO technique.

Research Conclusions:

– MemHarness significantly outperforms existing memory-augmented baselines and pure reinforcement learning models, especially in out-of-distribution scenarios, by preventing negative transfer and enhancing intrinsic reasoning capabilities.

Paper link: https://huggingface.co/papers/2607.28272

31. ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

Keywords: ShadowDancer, interactive video, dynamics, action transfer, cross-shadow prediction

Category: Computer Vision

Research Objective:

– To enable precise control of any-action, frame-level dynamics in interactive video world models through novel techniques.

Research Methods:

– Utilizes shadow pairs to construct scalable video pairs replaying the same dynamics with varied appearances.

– Implements cross-shadow prediction to learn actions by predicting dynamics from one shadow to another.

Research Conclusions:

– Demonstrated improved action transfer and extended action rollout in diverse dynamics families, achieving an average blinded win rate of 86% over existing models.

Paper link: https://huggingface.co/papers/2607.28362

32. See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Keywords: Multimodal large language models, Visual reasoning, See2Think, Visual Action-of-Thought

Category: Multi-Modal Learning

Research Objective:

– To introduce See2Think, an evaluation framework that assesses visual reasoning in multimodal large language models through benchmarks like See2ThinkBench and Visual Action-of-Thought.

Research Methods:

– The creation of See2ThinkBench with 1,200 visually dependent tasks across 12 categories and the use of Visual Action-of-Thought to record and analyze visual and textual operations under controlled settings.

Research Conclusions:

– Visual reasoning capabilities in models are strongly dependent on both the model and environment, showing varied performance across tasks. Additionally, even though models often select the correct visual operations, the accuracy improvements are hindered by challenges in rendering fidelity and reliance on feedback which does not always improve accuracy.

Paper link: https://huggingface.co/papers/2607.26769

33. RefCaptioner: Multi-Reference Image-Grounded Video Captioning

Keywords: multi-reference image-grounded video captioning, RefCaptioner, factual video descriptions, Hierarchical Coverage-Discounted GRPO, AI-generated videos

Category: Multi-Modal Learning

Research Objective:

– To introduce and address the task of multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level reference grounding.

Research Methods:

– Development of RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO to enhance reference selection and cross-reference consistency.

Research Conclusions:

– RefCaptioner outperforms other open-source models in overall performance, yielding captions preferred by human annotators and allowing for more source-faithful video reconstruction.

Paper link: https://huggingface.co/papers/2607.28509

34. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Keywords: Embodied intelligence, Ambient Capture Engine, AI Native, Multi-modal datasets, Human-centric data

Category: Robotics and Autonomous Systems

Research Objective:

– The study focuses on addressing the data bottleneck in embodied intelligence by introducing the Ambient Capture Engine (ACE), which captures comprehensive sensory data in real home environments.

Research Methods:

– The authors developed ACE to record multisensory data through a combination of table-scale and room-scale configurations, creating a detailed dataset, ACE-Data-0, encompassing 150 hours of interactions across 200 task categories.

Research Conclusions:

– Evaluations reveal significant limitations in current state-of-the-art methods, particularly under challenging conditions like occlusion and long temporal horizons, highlighting ACE’s potential in supporting scalable advancements in imitation learning and embodied AI systems.

Paper link: https://huggingface.co/papers/2607.28625

35. Flux-OPD: On-Policy Distillation with Evolving Contexts

Keywords: Large language model, evolving contexts, Flux-OPD, open-ended domains, teacher supervision

Category: Natural Language Processing

Research Objective:

– The study aims to develop a mechanism for more effective supervision in large language model training across open-ended domains, addressing the issue of lacking verifiable rewards and formalized task preferences.

Research Methods:

– The research introduces Flux-OPD, a novel OPD paradigm that leverages evolving contexts as a form of in-training supervision. It analyzes the effect of context through the reverse KL objective and uses context-conditioned teachers to provide contextual difference signals.

Research Conclusions:

– Flux-OPD demonstrates superiority over existing OPD paradigms, effectively capturing task preferences and enhancing performance in open-ended tasks by combining evolving contexts with teacher supervision.

Paper link: https://huggingface.co/papers/2607.28022

36. Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Keywords: Agentic Visual Reasoning, Multimodal Large Language Models, Mode Adaptiveness, Tool Effect, Beacon

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance the success rate of multimodal large language models (MLLMs) in complex tasks by rethinking agentic visual reasoning, focusing on Mode Adaptiveness and Tool Effect.

Research Methods:

– Comprehensive analysis to quantify Mode Adaptiveness and Tool Effect in existing models, followed by the proposal of a new model named Beacon, featuring Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion.

Research Conclusions:

– The research concludes that existing models have limited Mode Adaptiveness, and tool benefits are often neutralized by their drawbacks. Beacon, however, shows improved overall performance, notably enhancing Mode Adaptiveness and achieving genuine gains from tool use.

Paper link: https://huggingface.co/papers/2607.28595

37. VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Keywords: Text-to-video, VideoCoCo, Blender program, spatiotemporal process, generative video engine

Category: Generative Models

Research Objective:

– The study aims to address the limitations of text-to-video models in generating physically consistent dynamics by introducing a new framework called VideoCoCo.

Research Methods:

– A dual-engine framework is employed, where a coding agent synthesizes executable Blender code to specify scene evolution and produce deterministic drafts, followed by a generative video engine for photorealistic video transformation.

Research Conclusions:

– VideoCoCo effectively improves performance on benchmarks PhyGenBench and VBench-2.0 by providing an executable, controllable, and inspectable representation for video generation, achieving the best average scores on both benchmarks.

Paper link: https://huggingface.co/papers/2607.27380

38. PhiZero: A World Model Built Around Physical Language

Keywords: PhiZero, physical language, world model, zero-shot motion transfer

Category: Generative Models

Research Objective:

– The primary aim is to develop PhiZero, a model that abstracts predictive structure from visual experience to explicitly reason about physical world evolution using physical language.

Research Methods:

– PhiZero employs a reason-then-render paradigm where it infers future world evolution sequences in physical language and renders transitions into video through self-supervision from in-the-wild videos.

Research Conclusions:

– The experiments demonstrate PhiZero’s capability to model physically coherent world evolution and highlight its potential for realistic interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

Paper link: https://huggingface.co/papers/2607.28624

39. Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Keywords: GUI agents, Qwen-UI-Agent, Reinforcement Learning, MobileWorld, Online RL

Category: AI Systems and Tools

Research Objective:

– The paper aims to advance GUI agents for real-world applications, enabling reliable operation across various devices and platforms, proactive service initiation, and autonomous capability improvements.

Research Methods:

– Introduces Qwen-UI-Agent, utilizing a unified action space that combines GUI operations with CLI execution, and employs an AutoResearch-style data flywheel and Online Reinforcement Learning (RL) for training and advancing agent performance.

Research Conclusions:

– Qwen-UI-Agent demonstrates state-of-the-art performance in mobile-use benchmarks and competitive results in computer- and browser-use tasks, achieving high scores on MobileWorld, AndroidDaily, and WebArena among others.

Paper link: https://huggingface.co/papers/2607.28227

The post AI Native Daily Paper Digest – 20260925 – AskChem | Metis | Frontis-MA1 appeared first on AI Native Foundation.

1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.

Paper link: https://huggingface.co/papers/2607.27958

14. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Keywords: Explorative Modeling, Generative Models, FLOP Efficiency, Sample Efficiency, Parameter Efficiency

Category: Generative Models

Research Objective:

– Introduce Explorative Modeling as a new paradigm in generative modeling to enable end-to-end training by exploring candidate matches and training on the best predictions.

Research Methods:

– The Explorative Models are applied in generative domains, increasing exploration as a pretraining axis to improve performance across diverse datasets including images, video, and language.

Research Conclusions:

– Explorative Modeling enhances efficiency exponentially, achieving significant improvement in FLOP, sample, and parameter efficiency, and enables end-to-end reconstructive generative modeling with reduced inference steps, establishing itself as a new training axis for existing generative models.

Paper link: https://huggingface.co/papers/2607.27372

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Keywords: AI Tour Meeting, Large Language Model, agents, simulation tool

Category: AI Systems and Tools

Research Objective:

– To propose a travel planning framework called AI Tour Meeting using agents powered by Large Language Models (LLMs) to collaboratively create itineraries through natural language discussion.

Research Methods:

– Implementation of a framework for configuring agent personas and orchestrating discussion workflows, enabling monitoring and LLM deployment in the simulation of tour planning discussions.

Research Conclusions:

– The framework is validated by system demonstration and provides analytical insights into the behavior of multiple LLM agents during group travel planning.

Paper link: https://huggingface.co/papers/2607.18806

16. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Keywords: Speculative Decoding, Large Language Model, Truncation-based Verification, Collaborative Verification

Category: Natural Language Processing

Research Objective:

– To analyze the distributions induced by lossy verification schemes in speculative decoding and identify their potential pitfalls in large language model inference.

Research Methods:

– Classified lossy verification methods into truncation-based and collaborative verification, and developed a diagnostic evaluation framework across curated benchmarks.

Research Conclusions:

– Truncation-based methods are susceptible to significant performance degradation due to distributional distortion.

– For collaborative verification, controlling the overshoot of draft probabilities relative to target probabilities is crucial to maintaining output quality.

Paper link: https://huggingface.co/papers/2607.26627

17. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Keywords: Sparse mixture-of-experts, Expert Subspace Separation Index, coherent overlap, next-token prediction

Category: Natural Language Processing

Research Objective:

– The study aims to distinguish the benefits and interactions of experts within sparse mixture-of-experts (MoE) language models by analyzing their geometric arrangement.

Research Methods:

– Utilizes an Expert Subspace Separation Index (ESSI) and various controlled interventions, including matched-route residuals and prefix-controlled factorial setups, to analyze expert overlaps and their impact on representation.

Research Conclusions:

– Coherent overlap indicates that while routes overlap, useful multi-expert computation persists, clarifying why geometric similarity doesn’t alone determine redundancy or pruning value.

– Adding later experts improves next-token predictions significantly in most cases, emphasizing the importance of non-redundant expert selections.

Paper link: https://huggingface.co/papers/2607.28308

18. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Keywords: Fairness Pruning, Large Language Models, Demographic Bias, Causal Bias Localization, AI Ethics and Fairness

Category: AI Ethics and Fairness

Research Objective:

– To introduce and empirically validate Fairness Pruning, a method aimed at managing and mitigating demographic biases in large language models.

Research Methods:

– Utilization of minimally contrastive prompt pairs and inference-time activation capture to identify neurons that react differentially to demographic attributes.

– Empirical evaluation on large models such as Llama-3.2 and Salamandra-2B with structural intervention by zeroing neurons.

Research Conclusions:

– Demonstrated that zeroing specific neurons alters model responses to demographic variables, with a bidirectional bias destabilization effect.

– Confirmed that demographic bias processing and model capabilities are dissociable, setting foundations for transitioning to directional behavior modulation.

Paper link: https://huggingface.co/papers/2607.28319

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Keywords: Speech Emotion Recognition, Edge Devices, Knowledge Distillation, AI Native

Category: Machine Learning

Research Objective:

– This research aims to optimize on-device speech emotion recognition (SER) for real-time applications by developing a more efficient model that can operate effectively on edge devices.

Research Methods:

– Adaptive Multi-teacher Relational Distillation (AMRD) is proposed, using a one-class SVM to assign weights to teacher models and a relational distillation loss to improve inter-sample relational structure alignment between teacher and student models.

Research Conclusions:

– The AMRD approach outperforms single-teacher distillation baselines on the IEMOCAP and CREMA-D datasets, demonstrating its effectiveness across multiple student architectures and confirming that both components of the method yield complementary benefits.

Paper link: https://huggingface.co/papers/2607.25289

20.

Paper link: 

21. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Keywords: Pedestrian Archetypes, Real-world Traffic Scenarios, Behavior Patterns

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces new pedestrian archetypes that exhibit significant behavioral differences, enhancing the understanding of pedestrian behaviors beyond the original taxonomy.

Research Methods:

– Analysis of YouTube dash-cam videos for identifying additional pedestrian archetypes not covered in previous work.

Research Conclusions:

– Seven new pedestrian archetypes are presented, each defined by unique behaviors that are documented through video-frame evidence, thus expanding the existing framework for assessing pedestrian behavior in traffic.

Paper link: https://huggingface.co/papers/2607.16922

22. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Keywords: Vision-Language Models, GPU Memory Constraints, Image-QA Dataset

Category: Multi-Modal Learning

Research Objective:

– Address the challenge of long visual context in vision-language models by introducing ReToken, an explicit retrieval target for selecting query-relevant visual tokens.

Research Methods:

– ReToken is trained on a small image-QA dataset and designed to select a sparse set of visual tokens from a pre-filled visual KV cache.

Research Conclusions:

– ReToken shows consistent performance gains across image and video benchmarks, with significant improvements in Visual Haystacks and LVBench, while maintaining efficiency on a single H100 for both training and long-video inference.

Paper link: https://huggingface.co/papers/2607.28627

23. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Keywords: OmniScope, token compression, query, salience mismatch, AI Native

Category: Multi-Modal Learning

Research Objective:

– The objective is to address the cross-modal salience mismatch in token compression methods for omnimodal large language models by developing a framework named OmniScope that estimates relevance separately for audio and video.

Research Methods:

– Introduced a training-free token compression framework using the query as a shared semantic anchor, allocating modality-specific token budgets, pruning visual tokens with an anchor-delta strategy, and merging audio tokens to reduce redundancy.

Research Conclusions:

– OmniScope outperforms existing methods in accuracy across audio-video benchmarks and model scales, offering improvements in speed and memory efficiency while maintaining high accuracy with minimal drop.

Paper link: https://huggingface.co/papers/2607.23193

24. Harness-G: A Graph-Structured Harness for Search Agents

Keywords: Reinforcement Learning, retrieval aliasing, Harness-G, Structured Non-myopic Credit

Category: Reinforcement Learning

Research Objective:

– The study addresses the problem of retrieval aliasing in reinforcement learning search agents by proposing a new retrieval framework to optimize query generation and interaction.

Research Methods:

– The authors introduce Harness-G, a graph-structured retrieval framework that transforms query generation into finite action selection. They also develop Structured Non-myopic Credit (SNC) for better rewarding actions.

Research Conclusions:

– Harness-G achieves the highest average F1 across six QA benchmarks and outperforms the best baseline by significant margins, demonstrating its effectiveness in reducing linguistic aliasing and improving retrieval accuracy.

Paper link: https://huggingface.co/papers/2607.27652

25. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Keywords: Deep Research agents, Misleading Knowledge, Evidence Synthesis, Workflow Reliability

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to explore the reliability vulnerabilities in Deep Research agents when exposed to misleading knowledge in open information environments.

Research Methods:

– Introduces MisKnow-Agent, a framework for generating misleading knowledge instances to test these vulnerabilities, utilizing 5,933 quality-controlled instances derived from DeepResearch Benchmark tasks.

Research Conclusions:

– The experiments demonstrate that even minimal misleading knowledge can lead to false conclusions in final reports, highlighting a significant reliability issue.

– Despite verifier models identifying misleading instances, they can still be adopted in long-term workflows, indicating a gap in workflow-level evidence verification.

– Evaluated defenses show mitigation effects but fail to fully prevent the adoption of false conclusions, pointing to the need for improved verification and correction capabilities.

Paper link: https://huggingface.co/papers/2607.20891

26. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Keywords: LLM agents, filesystem-based memory, memory organization, search economy, management agent

Category: AI Systems and Tools

Research Objective:

– To systematically explore filesystem-based memory for Large Language Model (LLM) agents, assessing the structure and organization of such memory.

Research Methods:

– The study formalizes the concept of a memory filesystem with a management agent for organization, a search agent for querying, and an execution agent for task trajectory integration.

– Investigated variables include memory shape, stream scale, tool harness, and agent strength, with focus on memory growth, retrieval costs, and store health.

Research Conclusions:

– Organization in filesystem-based memory leads to improved search economy, significantly reducing retrieval costs.

– However, as memory grows, management agents struggle to maintain organization quality, and improvements in organization do not necessarily translate to better answer quality.

– The study identifies the tool set as a significant factor in shaping the memory store, highlighting a design space for agent memory rather than just an assumption.

Paper link: https://huggingface.co/papers/2607.26637

27. Multi-Head Attention Residuals

Keywords: Transformers, Multi-Head Attention Residuals, Nemotron, validation loss

Category: Machine Learning

Research Objective:

– Introduction of Multi-Head Attention Residuals (MHAR) to improve information propagation across model depth and width.

Research Methods:

– MHAR utilizes H per-subspace heads with distinct softmax distributions, trained on a Nemotron-based corpus to optimize validation loss across various model sizes.

Research Conclusions:

– MHAR outperforms standard Transformers with validation loss improvements across 100M, 350M, and 1B scales. Optimal head count H = 8 increases efficiency and maintains near-baseline memory consumption.

Paper link: https://huggingface.co/papers/2607.27230

28. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Keywords: Synthetic Environments, Echoverse, Reinforcement Learning, State Representation

Category: Reinforcement Learning

Research Objective:

– To improve computer-use agents’ training environments by focusing on behavioral depth, targeted interactions, and model improvement.

Research Methods:

– Developed Echoverse to compile specifications into stateful applications, incorporating a co-evolution loop for iterative environment and model enhancement.

Research Conclusions:

– Demonstrated significant model improvement across various environments, with increases in evaluation scores and live-site accuracy via deep environments and iterative environment repairs.

Paper link: https://huggingface.co/papers/2607.28074

29. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Keywords: INTACT, Intent-to-Action, Latent World Models, Action Law Semantics

Category: Reinforcement Learning

Research Objective:

– To introduce INTACT, an end-to-end JEPA, aimed at transforming action-labeled, reward-free trajectories into a deployable intent-to-action interface without expensive test-time search.

Research Methods:

– Utilizes a unique architecture with isomorphic intent representation for local and goal motion-intent models, and provides intact transfer from RGB evidence to action-effective latent intent coordinates.

Research Conclusions:

– The INTACT model achieved up to 100% success rates on LeWM tasks using a search-free policy, significantly reducing the need for candidate sequence sampling while improving performance through a shared four-task encoder. Direct inference is highly efficient, taking only 2.9-5.5 milliseconds.

Paper link: https://huggingface.co/papers/2607.26056

30. MemHarness: Memory Is Reconstructed, Not Replayed

Keywords: MemHarness, Negative Transfer, Context-grounded Guidance, Large Language Model Agents, GRPO

Category: Reinforcement Learning

Research Objective:

– The objective is to develop a framework named MemHarness that enables large language model agents to actively reconstruct past experiences based on the current context, thereby providing more accurate and context-relevant guidance.

Research Methods:

– MemHarness employs a unified policy model that critiques and reconstructs retrieved experiences for each decision step. This reconstructive ability is achieved through end-to-end training using the GRPO technique.

Research Conclusions:

– MemHarness significantly outperforms existing memory-augmented baselines and pure reinforcement learning models, especially in out-of-distribution scenarios, by preventing negative transfer and enhancing intrinsic reasoning capabilities.

Paper link: https://huggingface.co/papers/2607.28272

31. ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

Keywords: ShadowDancer, interactive video, dynamics, action transfer, cross-shadow prediction

Category: Computer Vision

Research Objective:

– To enable precise control of any-action, frame-level dynamics in interactive video world models through novel techniques.

Research Methods:

– Utilizes shadow pairs to construct scalable video pairs replaying the same dynamics with varied appearances.

– Implements cross-shadow prediction to learn actions by predicting dynamics from one shadow to another.

Research Conclusions:

– Demonstrated improved action transfer and extended action rollout in diverse dynamics families, achieving an average blinded win rate of 86% over existing models.

Paper link: https://huggingface.co/papers/2607.28362

32. See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Keywords: Multimodal large language models, Visual reasoning, See2Think, Visual Action-of-Thought

Category: Multi-Modal Learning

Research Objective:

– To introduce See2Think, an evaluation framework that assesses visual reasoning in multimodal large language models through benchmarks like See2ThinkBench and Visual Action-of-Thought.

Research Methods:

– The creation of See2ThinkBench with 1,200 visually dependent tasks across 12 categories and the use of Visual Action-of-Thought to record and analyze visual and textual operations under controlled settings.

Research Conclusions:

– Visual reasoning capabilities in models are strongly dependent on both the model and environment, showing varied performance across tasks. Additionally, even though models often select the correct visual operations, the accuracy improvements are hindered by challenges in rendering fidelity and reliance on feedback which does not always improve accuracy.

Paper link: https://huggingface.co/papers/2607.26769

33. RefCaptioner: Multi-Reference Image-Grounded Video Captioning

Keywords: multi-reference image-grounded video captioning, RefCaptioner, factual video descriptions, Hierarchical Coverage-Discounted GRPO, AI-generated videos

Category: Multi-Modal Learning

Research Objective:

– To introduce and address the task of multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level reference grounding.

Research Methods:

– Development of RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO to enhance reference selection and cross-reference consistency.

Research Conclusions:

– RefCaptioner outperforms other open-source models in overall performance, yielding captions preferred by human annotators and allowing for more source-faithful video reconstruction.

Paper link: https://huggingface.co/papers/2607.28509

34. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Keywords: Embodied intelligence, Ambient Capture Engine, AI Native, Multi-modal datasets, Human-centric data

Category: Robotics and Autonomous Systems

Research Objective:

– The study focuses on addressing the data bottleneck in embodied intelligence by introducing the Ambient Capture Engine (ACE), which captures comprehensive sensory data in real home environments.

Research Methods:

– The authors developed ACE to record multisensory data through a combination of table-scale and room-scale configurations, creating a detailed dataset, ACE-Data-0, encompassing 150 hours of interactions across 200 task categories.

Research Conclusions:

– Evaluations reveal significant limitations in current state-of-the-art methods, particularly under challenging conditions like occlusion and long temporal horizons, highlighting ACE’s potential in supporting scalable advancements in imitation learning and embodied AI systems.

Paper link: https://huggingface.co/papers/2607.28625

35. Flux-OPD: On-Policy Distillation with Evolving Contexts

Keywords: Large language model, evolving contexts, Flux-OPD, open-ended domains, teacher supervision

Category: Natural Language Processing

Research Objective:

– The study aims to develop a mechanism for more effective supervision in large language model training across open-ended domains, addressing the issue of lacking verifiable rewards and formalized task preferences.

Research Methods:

– The research introduces Flux-OPD, a novel OPD paradigm that leverages evolving contexts as a form of in-training supervision. It analyzes the effect of context through the reverse KL objective and uses context-conditioned teachers to provide contextual difference signals.

Research Conclusions:

– Flux-OPD demonstrates superiority over existing OPD paradigms, effectively capturing task preferences and enhancing performance in open-ended tasks by combining evolving contexts with teacher supervision.

Paper link: https://huggingface.co/papers/2607.28022

36. Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Keywords: Agentic Visual Reasoning, Multimodal Large Language Models, Mode Adaptiveness, Tool Effect, Beacon

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance the success rate of multimodal large language models (MLLMs) in complex tasks by rethinking agentic visual reasoning, focusing on Mode Adaptiveness and Tool Effect.

Research Methods:

– Comprehensive analysis to quantify Mode Adaptiveness and Tool Effect in existing models, followed by the proposal of a new model named Beacon, featuring Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion.

Research Conclusions:

– The research concludes that existing models have limited Mode Adaptiveness, and tool benefits are often neutralized by their drawbacks. Beacon, however, shows improved overall performance, notably enhancing Mode Adaptiveness and achieving genuine gains from tool use.

Paper link: https://huggingface.co/papers/2607.28595

37. VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Keywords: Text-to-video, VideoCoCo, Blender program, spatiotemporal process, generative video engine

Category: Generative Models

Research Objective:

– The study aims to address the limitations of text-to-video models in generating physically consistent dynamics by introducing a new framework called VideoCoCo.

Research Methods:

– A dual-engine framework is employed, where a coding agent synthesizes executable Blender code to specify scene evolution and produce deterministic drafts, followed by a generative video engine for photorealistic video transformation.

Research Conclusions:

– VideoCoCo effectively improves performance on benchmarks PhyGenBench and VBench-2.0 by providing an executable, controllable, and inspectable representation for video generation, achieving the best average scores on both benchmarks.

Paper link: https://huggingface.co/papers/2607.27380

38. PhiZero: A World Model Built Around Physical Language

Keywords: PhiZero, physical language, world model, zero-shot motion transfer

Category: Generative Models

Research Objective:

– The primary aim is to develop PhiZero, a model that abstracts predictive structure from visual experience to explicitly reason about physical world evolution using physical language.

Research Methods:

– PhiZero employs a reason-then-render paradigm where it infers future world evolution sequences in physical language and renders transitions into video through self-supervision from in-the-wild videos.

Research Conclusions:

– The experiments demonstrate PhiZero’s capability to model physically coherent world evolution and highlight its potential for realistic interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

Paper link: https://huggingface.co/papers/2607.28624

39. Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Keywords: GUI agents, Qwen-UI-Agent, Reinforcement Learning, MobileWorld, Online RL

Category: AI Systems and Tools

Research Objective:

– The paper aims to advance GUI agents for real-world applications, enabling reliable operation across various devices and platforms, proactive service initiation, and autonomous capability improvements.

Research Methods:

– Introduces Qwen-UI-Agent, utilizing a unified action space that combines GUI operations with CLI execution, and employs an AutoResearch-style data flywheel and Online Reinforcement Learning (RL) for training and advancing agent performance.

Research Conclusions:

– Qwen-UI-Agent demonstrates state-of-the-art performance in mobile-use benchmarks and competitive results in computer- and browser-use tasks, achieving high scores on MobileWorld, AndroidDaily, and WebArena among others.

Paper link: https://huggingface.co/papers/2607.28227

The post AI Native Daily Paper Digest – 20260924 – AskChem | Metis | Chimera appeared first on AI Native Foundation.

1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.

Paper link: https://huggingface.co/papers/2607.27958

14. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Keywords: Explorative Modeling, Generative Models, FLOP Efficiency, Sample Efficiency, Parameter Efficiency

Category: Generative Models

Research Objective:

– Introduce Explorative Modeling as a new paradigm in generative modeling to enable end-to-end training by exploring candidate matches and training on the best predictions.

Research Methods:

– The Explorative Models are applied in generative domains, increasing exploration as a pretraining axis to improve performance across diverse datasets including images, video, and language.

Research Conclusions:

– Explorative Modeling enhances efficiency exponentially, achieving significant improvement in FLOP, sample, and parameter efficiency, and enables end-to-end reconstructive generative modeling with reduced inference steps, establishing itself as a new training axis for existing generative models.

Paper link: https://huggingface.co/papers/2607.27372

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Keywords: AI Tour Meeting, Large Language Model, agents, simulation tool

Category: AI Systems and Tools

Research Objective:

– To propose a travel planning framework called AI Tour Meeting using agents powered by Large Language Models (LLMs) to collaboratively create itineraries through natural language discussion.

Research Methods:

– Implementation of a framework for configuring agent personas and orchestrating discussion workflows, enabling monitoring and LLM deployment in the simulation of tour planning discussions.

Research Conclusions:

– The framework is validated by system demonstration and provides analytical insights into the behavior of multiple LLM agents during group travel planning.

Paper link: https://huggingface.co/papers/2607.18806

16. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Keywords: Speculative Decoding, Large Language Model, Truncation-based Verification, Collaborative Verification

Category: Natural Language Processing

Research Objective:

– To analyze the distributions induced by lossy verification schemes in speculative decoding and identify their potential pitfalls in large language model inference.

Research Methods:

– Classified lossy verification methods into truncation-based and collaborative verification, and developed a diagnostic evaluation framework across curated benchmarks.

Research Conclusions:

– Truncation-based methods are susceptible to significant performance degradation due to distributional distortion.

– For collaborative verification, controlling the overshoot of draft probabilities relative to target probabilities is crucial to maintaining output quality.

Paper link: https://huggingface.co/papers/2607.26627

17. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Keywords: Sparse mixture-of-experts, Expert Subspace Separation Index, coherent overlap, next-token prediction

Category: Natural Language Processing

Research Objective:

– The study aims to distinguish the benefits and interactions of experts within sparse mixture-of-experts (MoE) language models by analyzing their geometric arrangement.

Research Methods:

– Utilizes an Expert Subspace Separation Index (ESSI) and various controlled interventions, including matched-route residuals and prefix-controlled factorial setups, to analyze expert overlaps and their impact on representation.

Research Conclusions:

– Coherent overlap indicates that while routes overlap, useful multi-expert computation persists, clarifying why geometric similarity doesn’t alone determine redundancy or pruning value.

– Adding later experts improves next-token predictions significantly in most cases, emphasizing the importance of non-redundant expert selections.

Paper link: https://huggingface.co/papers/2607.28308

18. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Keywords: Fairness Pruning, Large Language Models, Demographic Bias, Causal Bias Localization, AI Ethics and Fairness

Category: AI Ethics and Fairness

Research Objective:

– To introduce and empirically validate Fairness Pruning, a method aimed at managing and mitigating demographic biases in large language models.

Research Methods:

– Utilization of minimally contrastive prompt pairs and inference-time activation capture to identify neurons that react differentially to demographic attributes.

– Empirical evaluation on large models such as Llama-3.2 and Salamandra-2B with structural intervention by zeroing neurons.

Research Conclusions:

– Demonstrated that zeroing specific neurons alters model responses to demographic variables, with a bidirectional bias destabilization effect.

– Confirmed that demographic bias processing and model capabilities are dissociable, setting foundations for transitioning to directional behavior modulation.

Paper link: https://huggingface.co/papers/2607.28319

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Keywords: Speech Emotion Recognition, Edge Devices, Knowledge Distillation, AI Native

Category: Machine Learning

Research Objective:

– This research aims to optimize on-device speech emotion recognition (SER) for real-time applications by developing a more efficient model that can operate effectively on edge devices.

Research Methods:

– Adaptive Multi-teacher Relational Distillation (AMRD) is proposed, using a one-class SVM to assign weights to teacher models and a relational distillation loss to improve inter-sample relational structure alignment between teacher and student models.

Research Conclusions:

– The AMRD approach outperforms single-teacher distillation baselines on the IEMOCAP and CREMA-D datasets, demonstrating its effectiveness across multiple student architectures and confirming that both components of the method yield complementary benefits.

Paper link: https://huggingface.co/papers/2607.25289

20.

Paper link: 

21. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Keywords: Pedestrian Archetypes, Real-world Traffic Scenarios, Behavior Patterns

Category: Robotics and Autonomous Systems

Research Objective:

– The research introduces new pedestrian archetypes that exhibit significant behavioral differences, enhancing the understanding of pedestrian behaviors beyond the original taxonomy.

Research Methods:

– Analysis of YouTube dash-cam videos for identifying additional pedestrian archetypes not covered in previous work.

Research Conclusions:

– Seven new pedestrian archetypes are presented, each defined by unique behaviors that are documented through video-frame evidence, thus expanding the existing framework for assessing pedestrian behavior in traffic.

Paper link: https://huggingface.co/papers/2607.16922

22. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Keywords: Vision-Language Models, GPU Memory Constraints, Image-QA Dataset

Category: Multi-Modal Learning

Research Objective:

– Address the challenge of long visual context in vision-language models by introducing ReToken, an explicit retrieval target for selecting query-relevant visual tokens.

Research Methods:

– ReToken is trained on a small image-QA dataset and designed to select a sparse set of visual tokens from a pre-filled visual KV cache.

Research Conclusions:

– ReToken shows consistent performance gains across image and video benchmarks, with significant improvements in Visual Haystacks and LVBench, while maintaining efficiency on a single H100 for both training and long-video inference.

Paper link: https://huggingface.co/papers/2607.28627

23. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Keywords: OmniScope, token compression, query, salience mismatch, AI Native

Category: Multi-Modal Learning

Research Objective:

– The objective is to address the cross-modal salience mismatch in token compression methods for omnimodal large language models by developing a framework named OmniScope that estimates relevance separately for audio and video.

Research Methods:

– Introduced a training-free token compression framework using the query as a shared semantic anchor, allocating modality-specific token budgets, pruning visual tokens with an anchor-delta strategy, and merging audio tokens to reduce redundancy.

Research Conclusions:

– OmniScope outperforms existing methods in accuracy across audio-video benchmarks and model scales, offering improvements in speed and memory efficiency while maintaining high accuracy with minimal drop.

Paper link: https://huggingface.co/papers/2607.23193

24. Harness-G: A Graph-Structured Harness for Search Agents

Keywords: Reinforcement Learning, retrieval aliasing, Harness-G, Structured Non-myopic Credit

Category: Reinforcement Learning

Research Objective:

– The study addresses the problem of retrieval aliasing in reinforcement learning search agents by proposing a new retrieval framework to optimize query generation and interaction.

Research Methods:

– The authors introduce Harness-G, a graph-structured retrieval framework that transforms query generation into finite action selection. They also develop Structured Non-myopic Credit (SNC) for better rewarding actions.

Research Conclusions:

– Harness-G achieves the highest average F1 across six QA benchmarks and outperforms the best baseline by significant margins, demonstrating its effectiveness in reducing linguistic aliasing and improving retrieval accuracy.

Paper link: https://huggingface.co/papers/2607.27652

25. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Keywords: Deep Research agents, Misleading Knowledge, Evidence Synthesis, Workflow Reliability

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to explore the reliability vulnerabilities in Deep Research agents when exposed to misleading knowledge in open information environments.

Research Methods:

– Introduces MisKnow-Agent, a framework for generating misleading knowledge instances to test these vulnerabilities, utilizing 5,933 quality-controlled instances derived from DeepResearch Benchmark tasks.

Research Conclusions:

– The experiments demonstrate that even minimal misleading knowledge can lead to false conclusions in final reports, highlighting a significant reliability issue.

– Despite verifier models identifying misleading instances, they can still be adopted in long-term workflows, indicating a gap in workflow-level evidence verification.

– Evaluated defenses show mitigation effects but fail to fully prevent the adoption of false conclusions, pointing to the need for improved verification and correction capabilities.

Paper link: https://huggingface.co/papers/2607.20891

26. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Keywords: LLM agents, filesystem-based memory, memory organization, search economy, management agent

Category: AI Systems and Tools

Research Objective:

– To systematically explore filesystem-based memory for Large Language Model (LLM) agents, assessing the structure and organization of such memory.

Research Methods:

– The study formalizes the concept of a memory filesystem with a management agent for organization, a search agent for querying, and an execution agent for task trajectory integration.

– Investigated variables include memory shape, stream scale, tool harness, and agent strength, with focus on memory growth, retrieval costs, and store health.

Research Conclusions:

– Organization in filesystem-based memory leads to improved search economy, significantly reducing retrieval costs.

– However, as memory grows, management agents struggle to maintain organization quality, and improvements in organization do not necessarily translate to better answer quality.

– The study identifies the tool set as a significant factor in shaping the memory store, highlighting a design space for agent memory rather than just an assumption.

Paper link: https://huggingface.co/papers/2607.26637

27. Multi-Head Attention Residuals

Keywords: Transformers, Multi-Head Attention Residuals, Nemotron, validation loss

Category: Machine Learning

Research Objective:

– Introduction of Multi-Head Attention Residuals (MHAR) to improve information propagation across model depth and width.

Research Methods:

– MHAR utilizes H per-subspace heads with distinct softmax distributions, trained on a Nemotron-based corpus to optimize validation loss across various model sizes.

Research Conclusions:

– MHAR outperforms standard Transformers with validation loss improvements across 100M, 350M, and 1B scales. Optimal head count H = 8 increases efficiency and maintains near-baseline memory consumption.

Paper link: https://huggingface.co/papers/2607.27230

28. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Keywords: Synthetic Environments, Echoverse, Reinforcement Learning, State Representation

Category: Reinforcement Learning

Research Objective:

– To improve computer-use agents’ training environments by focusing on behavioral depth, targeted interactions, and model improvement.

Research Methods:

– Developed Echoverse to compile specifications into stateful applications, incorporating a co-evolution loop for iterative environment and model enhancement.

Research Conclusions:

– Demonstrated significant model improvement across various environments, with increases in evaluation scores and live-site accuracy via deep environments and iterative environment repairs.

Paper link: https://huggingface.co/papers/2607.28074

29. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Keywords: INTACT, Intent-to-Action, Latent World Models, Action Law Semantics

Category: Reinforcement Learning

Research Objective:

– To introduce INTACT, an end-to-end JEPA, aimed at transforming action-labeled, reward-free trajectories into a deployable intent-to-action interface without expensive test-time search.

Research Methods:

– Utilizes a unique architecture with isomorphic intent representation for local and goal motion-intent models, and provides intact transfer from RGB evidence to action-effective latent intent coordinates.

Research Conclusions:

– The INTACT model achieved up to 100% success rates on LeWM tasks using a search-free policy, significantly reducing the need for candidate sequence sampling while improving performance through a shared four-task encoder. Direct inference is highly efficient, taking only 2.9-5.5 milliseconds.

Paper link: https://huggingface.co/papers/2607.26056

30. MemHarness: Memory Is Reconstructed, Not Replayed

Keywords: MemHarness, Negative Transfer, Context-grounded Guidance, Large Language Model Agents, GRPO

Category: Reinforcement Learning

Research Objective:

– The objective is to develop a framework named MemHarness that enables large language model agents to actively reconstruct past experiences based on the current context, thereby providing more accurate and context-relevant guidance.

Research Methods:

– MemHarness employs a unified policy model that critiques and reconstructs retrieved experiences for each decision step. This reconstructive ability is achieved through end-to-end training using the GRPO technique.

Research Conclusions:

– MemHarness significantly outperforms existing memory-augmented baselines and pure reinforcement learning models, especially in out-of-distribution scenarios, by preventing negative transfer and enhancing intrinsic reasoning capabilities.

Paper link: https://huggingface.co/papers/2607.28272

31. ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

Keywords: ShadowDancer, interactive video, dynamics, action transfer, cross-shadow prediction

Category: Computer Vision

Research Objective:

– To enable precise control of any-action, frame-level dynamics in interactive video world models through novel techniques.

Research Methods:

– Utilizes shadow pairs to construct scalable video pairs replaying the same dynamics with varied appearances.

– Implements cross-shadow prediction to learn actions by predicting dynamics from one shadow to another.

Research Conclusions:

– Demonstrated improved action transfer and extended action rollout in diverse dynamics families, achieving an average blinded win rate of 86% over existing models.

Paper link: https://huggingface.co/papers/2607.28362

32. See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Keywords: Multimodal large language models, Visual reasoning, See2Think, Visual Action-of-Thought

Category: Multi-Modal Learning

Research Objective:

– To introduce See2Think, an evaluation framework that assesses visual reasoning in multimodal large language models through benchmarks like See2ThinkBench and Visual Action-of-Thought.

Research Methods:

– The creation of See2ThinkBench with 1,200 visually dependent tasks across 12 categories and the use of Visual Action-of-Thought to record and analyze visual and textual operations under controlled settings.

Research Conclusions:

– Visual reasoning capabilities in models are strongly dependent on both the model and environment, showing varied performance across tasks. Additionally, even though models often select the correct visual operations, the accuracy improvements are hindered by challenges in rendering fidelity and reliance on feedback which does not always improve accuracy.

Paper link: https://huggingface.co/papers/2607.26769

33. RefCaptioner: Multi-Reference Image-Grounded Video Captioning

Keywords: multi-reference image-grounded video captioning, RefCaptioner, factual video descriptions, Hierarchical Coverage-Discounted GRPO, AI-generated videos

Category: Multi-Modal Learning

Research Objective:

– To introduce and address the task of multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level reference grounding.

Research Methods:

– Development of RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO to enhance reference selection and cross-reference consistency.

Research Conclusions:

– RefCaptioner outperforms other open-source models in overall performance, yielding captions preferred by human annotators and allowing for more source-faithful video reconstruction.

Paper link: https://huggingface.co/papers/2607.28509

34. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Keywords: Embodied intelligence, Ambient Capture Engine, AI Native, Multi-modal datasets, Human-centric data

Category: Robotics and Autonomous Systems

Research Objective:

– The study focuses on addressing the data bottleneck in embodied intelligence by introducing the Ambient Capture Engine (ACE), which captures comprehensive sensory data in real home environments.

Research Methods:

– The authors developed ACE to record multisensory data through a combination of table-scale and room-scale configurations, creating a detailed dataset, ACE-Data-0, encompassing 150 hours of interactions across 200 task categories.

Research Conclusions:

– Evaluations reveal significant limitations in current state-of-the-art methods, particularly under challenging conditions like occlusion and long temporal horizons, highlighting ACE’s potential in supporting scalable advancements in imitation learning and embodied AI systems.

Paper link: https://huggingface.co/papers/2607.28625

35. Flux-OPD: On-Policy Distillation with Evolving Contexts

Keywords: Large language model, evolving contexts, Flux-OPD, open-ended domains, teacher supervision

Category: Natural Language Processing

Research Objective:

– The study aims to develop a mechanism for more effective supervision in large language model training across open-ended domains, addressing the issue of lacking verifiable rewards and formalized task preferences.

Research Methods:

– The research introduces Flux-OPD, a novel OPD paradigm that leverages evolving contexts as a form of in-training supervision. It analyzes the effect of context through the reverse KL objective and uses context-conditioned teachers to provide contextual difference signals.

Research Conclusions:

– Flux-OPD demonstrates superiority over existing OPD paradigms, effectively capturing task preferences and enhancing performance in open-ended tasks by combining evolving contexts with teacher supervision.

Paper link: https://huggingface.co/papers/2607.28022

36. Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Keywords: Agentic Visual Reasoning, Multimodal Large Language Models, Mode Adaptiveness, Tool Effect, Beacon

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance the success rate of multimodal large language models (MLLMs) in complex tasks by rethinking agentic visual reasoning, focusing on Mode Adaptiveness and Tool Effect.

Research Methods:

– Comprehensive analysis to quantify Mode Adaptiveness and Tool Effect in existing models, followed by the proposal of a new model named Beacon, featuring Necessity-Aware Adaptive Reward and Hint-Guided Capability Expansion.

Research Conclusions:

– The research concludes that existing models have limited Mode Adaptiveness, and tool benefits are often neutralized by their drawbacks. Beacon, however, shows improved overall performance, notably enhancing Mode Adaptiveness and achieving genuine gains from tool use.

Paper link: https://huggingface.co/papers/2607.28595

37. VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Keywords: Text-to-video, VideoCoCo, Blender program, spatiotemporal process, generative video engine

Category: Generative Models

Research Objective:

– The study aims to address the limitations of text-to-video models in generating physically consistent dynamics by introducing a new framework called VideoCoCo.

Research Methods:

– A dual-engine framework is employed, where a coding agent synthesizes executable Blender code to specify scene evolution and produce deterministic drafts, followed by a generative video engine for photorealistic video transformation.

Research Conclusions:

– VideoCoCo effectively improves performance on benchmarks PhyGenBench and VBench-2.0 by providing an executable, controllable, and inspectable representation for video generation, achieving the best average scores on both benchmarks.

Paper link: https://huggingface.co/papers/2607.27380

38. PhiZero: A World Model Built Around Physical Language

Keywords: PhiZero, physical language, world model, zero-shot motion transfer

Category: Generative Models

Research Objective:

– The primary aim is to develop PhiZero, a model that abstracts predictive structure from visual experience to explicitly reason about physical world evolution using physical language.

Research Methods:

– PhiZero employs a reason-then-render paradigm where it infers future world evolution sequences in physical language and renders transitions into video through self-supervision from in-the-wild videos.

Research Conclusions:

– The experiments demonstrate PhiZero’s capability to model physically coherent world evolution and highlight its potential for realistic interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

Paper link: https://huggingface.co/papers/2607.28624

39. Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Keywords: GUI agents, Qwen-UI-Agent, Reinforcement Learning, MobileWorld, Online RL

Category: AI Systems and Tools

Research Objective:

– The paper aims to advance GUI agents for real-world applications, enabling reliable operation across various devices and platforms, proactive service initiation, and autonomous capability improvements.

Research Methods:

– Introduces Qwen-UI-Agent, utilizing a unified action space that combines GUI operations with CLI execution, and employs an AutoResearch-style data flywheel and Online Reinforcement Learning (RL) for training and advancing agent performance.

Research Conclusions:

– Qwen-UI-Agent demonstrates state-of-the-art performance in mobile-use benchmarks and competitive results in computer- and browser-use tasks, achieving high scores on MobileWorld, AndroidDaily, and WebArena among others.

Paper link: https://huggingface.co/papers/2607.28227

The post AI Native Daily Paper Digest – 20260923 – AskChem | Metis | Chimera appeared first on AI Native Foundation.

1. AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Keywords: AskChem, claim-centered infrastructure, GPT-5.5, AI agents, chemistry search

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective of the research was to develop AskChem, a claim-centered infrastructure for conducting cross-paper chemistry searches, which shifts the retrieval focus from whole papers to provenance-carrying claims.

Research Methods:

– AskChem converts papers into atomic, typed claims grounded by a source DOI and quotes or evidence locators, enabling hierarchical retrieval and exploration through a faceted taxonomy and evidence graph.

Research Conclusions:

– AskChem indexes 2.4M claims from 147K papers, improving the retrieval capability with 100% resolvable DOIs for a GPT-5.5 reader and provides superior citation density compared to other systems.

Paper link: https://huggingface.co/papers/2607.28618

2. Metis: Memory Foundation Model

Keywords: AI Native, native memory, memory foundation models, Metis

Category: Foundations of AI

Research Objective:

– Introduce memory foundation models to enable native memory capabilities in AI systems.

Research Methods:

– Proposed Metis, a new architecture with a native memory state for storing historical information.

– Developed large-scale memory-specific training data and optimization objectives for native memory procedures.

Research Conclusions:

– Metis demonstrates native memory capabilities with advantages in architecture and efficiency.

– The study provides a detailed analysis of strengths, limitations, and behaviors of Metis.

Paper link: https://huggingface.co/papers/2607.26760

3. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Keywords: Recursive self-improvement, AI4AI, OpenMLE, Machine learning engineering, Reinforcement learning

Category: AI Systems and Tools

Research Objective:

– The paper aims to study Recursive self-improvement by introducing OpenMLE, an open full-stack system for AI4AI research in machine learning engineering.

Research Methods:

– Utilizes a combination of execution feedback environments (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) to post-train a meta-evolution agent, Frontis-MA1, with atomic program-evolution operators.

Research Conclusions:

– The system significantly improves performance on MLE-Bench Lite, surpassing other models in efficiency and capability with OpenMLE-Evo-Max, and demonstrates transferable capabilities on NatureBench Lite.

Paper link: https://huggingface.co/papers/2607.28568

4. Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Keywords: Memory Decoder, scalability, parameter efficiency, Faiss, pretraining

Category: Natural Language Processing

Research Objective:

– To scale memory models up to 6.9B parameters and improve memory capacity independently for better performance of language models.

Research Methods:

– Implementation of a distributed pipeline for Faiss indexing and retrieval combined with sparse, batch-wise loading of kNN distributions.

– Comparison of parameter-performance tradeoff by allocating more parameters to memory across different model scales.

Research Conclusions:

– Independently scaling pretrained memory models can achieve better parameter efficiency and enhance language model performance, as demonstrated by outperforming larger models with fewer parameters.

Paper link: https://huggingface.co/papers/2607.27919

5. BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms

Keywords: Retrieval-Augmented Generation, BM25, Dense Retrieval, Lexical Retrieval, Agentic Reasoning

Category: Natural Language Processing

Research Objective:

– To conduct a controlled study evaluating retrieval-augmented generation (RAG) approaches across varying corpus sizes and determine their accuracy-cost scaling.

Research Methods:

– The study varies corpus size across 28 nested tiers, analyzing accuracy, construction, query tokens, and latency under a single reader model and judging protocol.

Research Conclusions:

– Different approaches show varying effectiveness based on corpus size, with BM25 eventually leading as the corpus grows, while agentic search incurs high query token costs at larger scales.

– Lexical retrieval emerges as the most scalable default, outperforming dense retrieval and graph-based indexing at larger corpus sizes, highlighting global candidate ranking’s importance.

Paper link: https://huggingface.co/papers/2607.26497

6. MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Keywords: Text-to-image, personalized editing, MPIE-Bench, MPIE-Eval, multi-person mesh reconstruction

Category: Computer Vision

Research Objective:

– The research aims to address issues in synthesizing high-fidelity images featuring multiple named people in shared contact actions, such as embrace or carry.

Research Methods:

– Introduced MPIE-Bench, a benchmark of 2,500 samples from video-mined editing triplets covering various scenes, interactions, and contact densities.

– Proposed MPIE-Eval, which evaluates contact-time geometry using a public multi-person mesh reconstruction framework.

Research Conclusions:

– The study finds that current visual language model (VLM) metrics often fail to capture anatomical and geometric errors, instead showing that MPIE-Eval aligns more closely with human judgment in evaluating anatomy and interaction accuracy.

Paper link: https://huggingface.co/papers/2607.27616

7. Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Keywords: Role-playing agents, Evaluation, User simulators, Personalized rubrics, Multi-turn conversation

Category: Human-AI Interaction

Research Objective:

– The paper aims to improve the evaluation of Role-playing agents (RPAs) by addressing the limitations of existing benchmarks and introducing a more personalized and realistic evaluation method through user simulators.

Research Methods:

– Implementation of PALATE, a scalable RPA benchmark using 300 character profiles and personalized rubrics, enabling free-form, multi-turn conversations with user simulators to assess RPAs’ conversational abilities.

Research Conclusions:

– PALATE provides interpretable evaluations of RPAs’ performance in specific user interactions, demonstrating higher agreement with human judgments compared to traditional rubrics, and addresses both generic quality and personalized user experience.

Paper link: https://huggingface.co/papers/2607.27816

8. SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Keywords: Vision-language models, SpatialCLI, Embodied agents, Compositional perception

Category: Multi-Modal Learning

Research Objective:

– The paper aims to address the capability mismatch in Vision-language models (VLMs) between task-level reasoning and capturing visual details, proposing a new framework to bridge this gap.

Research Methods:

– The proposed SpatialCLI framework operates in three stages: exposing specialist vision models as spatial tools, improving tool use with Cold-Start SFT and agentic RL, and internalizing successful tool-use trajectories.

Research Conclusions:

– SpatialCLI significantly improves VLM performance on the MindCube benchmark, raising the accuracy from 29.3% to 84.6% with tools, and maintaining 73.8% accuracy even without tools after internalization, surpassing existing models.

Paper link: https://huggingface.co/papers/2607.27703

9. Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Keywords: Visual Generation, Chimera, Kimi Delta Attention, Multi-head Latent Attention, Sparse Mixture-of-Experts

Category: Generative Models

Research Objective:

– To develop a hybrid visual diffusion backbone, Chimera, for efficient handling of high-resolution images, long videos, and multimodal context.

Research Methods:

– Introduction of Chimera that utilizes Kimi Delta Attention, Multi-head Latent Attention, and Sparse Mixture-of-Experts layers, along with HeteroP for scaling the architecture effectively.

Research Conclusions:

– Chimera’s dense backbone is 1.7x more compute-efficient than the Wan-2.1 baseline, and the complete system achieves 7.3x efficiency.

– Chimera can extrapolate from 5-second training clips to 30-second videos with minimal performance degradation.

– Fitted compute-optimal laws indicate balanced compute division in image pretraining, with a slight model size preference in video pretraining at higher budgets.

Paper link: https://huggingface.co/papers/2607.28611

10. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Keywords: On-policy self-distillation, policy optimization, reinforcement learning, reasoning performance, mathematical reasoning

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance reasoning language models using a refined version of on-policy self-distillation, termed β-OPSD, which introduces a versatile regularization parameter β for better policy optimization.

Research Methods:

– The paper introduces β-OPSD by positioning it within a broader policy-optimization family. It employs reinforcement learning for distilling the closed-form solution through efficient token-level logit mixing to approximate costly optimization processes.

Research Conclusions:

– Experiments demonstrate that β-OPSD consistently outperforms the standard OPSD in mathematical reasoning tasks, achieving improved optimization stability and better reasoning performance.

Paper link: https://huggingface.co/papers/2607.28582

11. Can Large Language Models Execute Parent Orders?

Keywords: Parent-order execution, Large Language Models (LLMs), PACE, Shenzhen Stock Exchange, Execution decisions

Category: AI in Finance

Research Objective:

– The study aims to enhance parent-order execution in algorithmic trading using Large Language Models (LLMs), addressing limitations of current methods like reliance on market assumptions and lack of adaptability.

Research Methods:

– Introduction of PACE (Plan-Ahead Controlled Execution), a hierarchical framework that separates parent-order execution into long-term planning and short-term execution, bypassing the need for specific market assumptions and training.

Research Conclusions:

– PACE demonstrated superior performance compared to existing TWAP, Almgren-Chriss, and learning-based strategies during tests on Shenzhen Stock Exchange data. It was observed that LLMs navigate execution decisions differently from humans, with higher confidence correlated with improved performance and earlier trades instead of late actions.

Paper link: https://huggingface.co/papers/2607.28410

12. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Keywords: Multimodal Agents, Visual Question Answering, Provenance-Constrained, Structured Evidence, LedgerMind

Category: Multi-Modal Learning

Research Objective:

– The study aims to improve the evaluation of multimodal agents that handle visual question answering by considering not only answer accuracy but also the process through which answers are derived.

Research Methods:

– Introduction of LedgerMind, a system based on a provenance-constrained state machine, to standardize tool outputs and support structured evidence.

– Utilization of a Three-Layer Grounding Protocol, Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to manage reasoning depth and ensure provenance.

Research Conclusions:

– LedgerMind effectively addresses common failure patterns such as unsupported reasoning and entity hallucinations, enhancing both answer accuracy and the reliability of the reasoning trajectory.

Paper link: https://huggingface.co/papers/2607.28374

13. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Keywords: LLM agents, multi-agent systems, Σ-Mem, reliability memory, interaction content

Category: AI Systems and Tools

Research Objective:

– The objective is to introduce Σ-Mem, a novel online reliability memory system, addressing limitations in existing memory systems within LLM agents, particularly in multi-agent environments.

Research Methods:

– Σ-Mem models historical competence and peer relationship evidence, using real symmetric states updated from correctness feedback while ensuring stable online adaptation without requiring model retraining.

Research Conclusions:

– Σ-Mem effectively adapts to shifts in counterfactual reliability and generalizes to new peers and task domains, outperforms traditional voting methods, and improves with increased correctness feedback, establishing a solid foundation for adaptive coordination in multi-agent LLM systems.