1. Anthropic Adds Record a Skill Feature to Claude Cowork Desktop App

Anthropic has launched a new feature called Record a Skill within Claude Cowork, available in the Claude desktop app. Users can record their screen while performing a task and narrate their actions, and Claude will convert the recording into a reusable skill it can execute automatically in the future. The feature is accessible via the plus menu in the Claude desktop app. It is available to subscribers on Pro, Max, and Team plans.

Read more: https://x.com/i/web/status/2079595988998554047

Video Credit: NotebookLM

2. Google DeepMind Launches Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber for AI Agents

Google DeepMind announced three new Gemini models targeting AI agent workloads with improvements in speed, quality, and cost. Gemini 3.6 Flash uses fewer tokens than its predecessor Gemini 3.5 Flash to deliver higher-quality output at the same price point. Gemini 3.5 Flash-Lite is positioned as a fast, cost-effective model suited for everyday tasks such as document processing and agentic search. Gemini 3.5 Flash Cyber is a specialized cybersecurity model designed to detect and patch critical software vulnerabilities.

Read more: https://x.com/i/web/status/2079589698490572961

Video Credit: @GoogleDeepMind on X

3. CoreWeave Reports NVIDIA Vera Rubin NVL72 Achieves 10x More Tokens Per Megawatt vs. Blackwell on DeepSeek-R1

CoreWeave has published what it describes as the first measured performance data for the NVIDIA Vera Rubin NVL72 GPU system. The benchmark results show a 10x improvement in tokens per second per megawatt when running DeepSeek-R1, compared to the Blackwell generation. The finding highlights a significant leap in energy efficiency for large-scale AI inference workloads. NVIDIA shared these results to underscore the performance gains of the Vera Rubin architecture over its predecessor.

Read more: https://x.com/i/web/status/2079601314234032474

Video Credit: @CoreWeave on X

4. Cognition Launches Devin Outposts to Run AI Engineer on Any Machine

Cognition has announced Devin Outposts, a new capability that allows its AI software engineer Devin to run on user-owned or self-managed infrastructure. Supported environments include Mac mini devices, GPU servers, virtual machines inside private networks, and Kubernetes clusters. The feature is aimed at organizations that want to deploy Devin closer to internal services or within controlled network environments. This expands Devin’s deployment flexibility beyond Cognition’s own cloud infrastructure.

Read more: https://x.com/cognition/status/2079612226252726615

Video Credit: @cognition on X

That’s all for today’s Global AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post Global AI Native Industry Insights – 20260722 – Anthropic | Google | Nvidia | more appeared first on AI Native Foundation.

1. TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs

Keywords: Video Multimodal Large Language Models, Temporal Grounding, TimeLens2, Temporal Wasserstein Reward

Category: Multi-Modal Learning

Research Objective:

– The study focuses on improving video temporal grounding, enabling models to predict evidence intervals across different video lengths, domains, and query forms.

Research Methods:

– The approach involves treating temporal evidence as interval sets, using a novel temporal Wasserstein reward, and employing techniques for multi-span supervision such as caption-derived proposals and boundary refinement.

Research Conclusions:

– TimeLens2 outperforms all size-matched baselines across seven benchmarks, with its 2B, 4B, and 8B variants surpassing open-source models by significant margins, improving performance over their Qwen3-VL backbones by 14.2, 13.0, and 18.1 mIoU points, respectively.

Paper link: https://huggingface.co/papers/2607.17423

2. DeepSearch-World: Self-Distillation for Deep Search Agents in a Verifiable Environment

Keywords: DeepSearch-Evolve, web agents, self-distillation, long-horizon interactions, verifiable environments

Category: Reinforcement Learning

Research Objective:

– The study aims to develop a framework called DeepSearch-Evolve for training web agents to improve from their own experiences efficiently within a verifiable environment.

Research Methods:

– Utilized a self-distillation framework for web agents called DeepSearch-Evolve, operating within DeepSearch-World, which includes 420K multi-hop QA tasks and supports cognitive behaviors like progress verification and failure recovery.

Research Conclusions:

– The results indicate that DeepSearch-World-9B can achieve competitive performance without distillation from more capable models, demonstrating its potential for scalable self-evolution in long-horizon web agents. The study will release the environment and resources to promote future research developments.

Paper link: https://huggingface.co/papers/2607.07820

3. HOMIE: Human-object Centric Video Personalization via Multimodal Intelligent Enchancement

Keywords: Human-object centric video personalization, subject fidelity, interaction patterns, HOMIE, MLLM

Category: Generative Models

Research Objective:

– Addressing limitations in human-object centric video personalization by balancing subject fidelity with accurate interaction patterns between humans and diverse objects.

Research Methods:

– Introduction of HOMIE framework to tackle inter- and intra-subject input settings, utilizing a better MLLM integration strategy and global multimodal guidance for aligning semantic features.

Research Conclusions:

– Extensive experiments demonstrate that HOMIE achieves state-of-the-art performance across various human-object centric video personalization tasks.

Paper link: https://huggingface.co/papers/2607.18217

4. RynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation Model

Keywords: RynnBrain 1.1, Embodied Perception, Spatial Reasoning, 3D Grounding, Robot Manipulation

Category: Robotics and Autonomous Systems

Research Objective:

– To introduce and enhance the RynnBrain 1.1 models with improved embodied perception, spatial reasoning, and 3D grounding capabilities, focusing on robot manipulation.

Research Methods:

– Developed using a unified spatio-temporal and physically grounded framework, incorporating contact-point prediction and native 3D grounding. Implementation of a unified cross-embodiment action space and embodiment-specific masking, tested on various robots.

Research Conclusions:

– RynnBrain 1.1 demonstrates superior performance in embodied cognition, localization, and 3D grounding, especially with the 122B-A10B model. Real-robot experiments confirmed its outperformance over other models, with improved success rates through joint multi-task and multi-embodiment training.

Paper link: https://huggingface.co/papers/2607.17977

5. GigaAM Multilingual: Foundation Model for Underrepresented Languages

Keywords: Multilingual ASR, Foundation Models, Central Asian Languages, Data Balancing, GigaAM Multilingual

Category: Natural Language Processing

Research Objective:

– To develop robust foundation models for underrepresented Central Asian languages, including Kazakh, Kyrgyz, and Uzbek, addressing the challenge of data scarcity.

Research Methods:

– Implementation of a cluster-level data balancing strategy during pre-training and a domain-aware sampling method during fine-tuning to reduce dominance by head languages.

– Pre-training a Conformer encoder, GigaAM Multilingual, on 2 million hours of audio using a HuBERT-style objective.

Research Conclusions:

– The proposed approach outperforms strong open pretrained encoders like Whisper Large v3 and Omnilingual-1B on target languages, achieving substantial improvements on spontaneous speech while maintaining efficiency.

– Release of the foundation encoder and ASR model, providing a successful strategy for effective multilingual adaptation in conditions of realistic data imbalance.

Paper link: https://huggingface.co/papers/2607.10371

6. Open-AoE: An Open Egocentric Manipulation Dataset and Toolchain for Embodied Learning

Keywords: Open-AoE, Embodied Intelligence, Dataset, Egocentric Manipulation

Category: Robotics and Autonomous Systems

Research Objective:

– To introduce Open-AoE, an open, community-oriented egocentric manipulation dataset and toolchain that facilitates the entire process from capture to model training.

Research Methods:

– Utilization of smartphone captures to amass around 2,000 hours of manipulation video, along with providing a structured data processing pipeline that includes temporal action segmentation, semantic annotation, and camera trajectory reconstruction.

Research Conclusions:

– Open-AoE lays down a practical open infrastructure, significantly reducing barriers to data contribution and reuse, and supports embodied model training, human-to-robot transfer, and world modeling.

Paper link: https://huggingface.co/papers/2607.14183

7. FlowMimic: Mask-free Visual Editing and Generation with Pixel-pair Warped Flow Field for Online Video Editing Data Generation and Modality Mimicry

Keywords: video editing, image editing, integration, modality mimic, language-based visual editing

Category: Multi-Modal Learning

Research Objective:

– To integrate generation and editing capabilities for video and image modalities within a single model.

Research Methods:

– Development of a pixel-pair temporal warped flow field to generate video editing samples from image editing samples.

– Introduction of sense-related tasks and corresponding latent-level and attention-level losses to internalize language-based visual editing capabilities.

Research Conclusions:

– Demonstrated the feasibility of learning video editing using data generated from image editing samples.

– Proposed a modality mimic approach to align capabilities and outputs between video and image modalities.

Paper link: https://huggingface.co/papers/2607.18227

8. Do Language Models Dream of Binding Molecules? Benchmarking LLMs under Spatial Constraints

Keywords: Structure-based drug design, 3D molecule generation, LLM, diffusion models

Category: Generative Models

Research Objective:

– The study aims to systematically analyze the capability of general-purpose Large Language Models (LLMs) in handling complex 3D spatial constraints in molecule generation compared to specialized diffusion models.

Research Methods:

– Introduce a benchmarking strategy named 3D-Fit to evaluate LLMs on multi-conditioned spatial molecule generation, considering factors like pocket-conditioned ligand generation and various spatial constraints.

Research Conclusions:

– Although LLMs currently lag behind state-of-the-art diffusion models in handling 3D spatial environments, they display promise in scaling to heterogeneous setups and managing multiple spatial constraints.

Paper link: https://huggingface.co/papers/2607.18144

9. LLM-as-a-Coach: Experiential Learning for Non-Verifiable Tasks

Keywords: Reinforcement Learning, Experiential Learning, Large Language Model, Feedback Model

Category: Reinforcement Learning

Research Objective:

– The study aims to propose a new framework, Experiential Learning (EL), for improving the learning process of open-ended tasks by providing richer feedback via a feedback model from an LLM-as-a-Coach instead of traditional rubric-based evaluations.

Research Methods:

– The proposed method distills assessments of each on-policy response into experiential knowledge, using on-policy context distillation, compared against traditional scalar reward systems.

Research Conclusions:

– EL consistently outperforms rubric-based RL across different policy families and offers better generalization beyond training distributions, while mitigating issues like reward hacking, establishing experiential knowledge as a more effective learning signal for non-verifiable tasks.

Paper link: https://huggingface.co/papers/2607.18110

10. Self-State Attacks on Self-Hosted AI Agents: How Far Can OS Defenses Go?

Keywords: Self-hosted AI agents, self-state attacks, OS-level defense

Category: AI Systems and Tools

Research Objective:

– The research aims to investigate the resilience of operating systems against self-state attacks in self-hosted AI agents.

Research Methods:

– The study characterizes an attack space and collects live activity traces from a self-hosted agent across various workload profiles. These traces are used to instantiate a 23-cell matrix and 43 concrete operations, followed by evaluating different defense strategies.

Research Conclusions:

– A layered defense stack is effective against most attack cells, but there remains a small residual attack surface that is structurally indistinguishable at the OS level. This suggests a need to reconsider OS-level defense against self-state attacks, potentially leading to new research avenues.

Paper link: https://huggingface.co/papers/2607.17986

11. The Geometry of Semantic Space: A Continuous Geometric Framework for the Transformer Architecture

Keywords: Transformer architecture, integro-differential equation, stochastic differential geometry, Large Language Models, optimization dynamics

Category: Natural Language Processing

Research Objective:

– To develop a continuous geometric framework that models the discrete operations of Transformer architectures as an integro-differential equation on a semantic fiber bundle.

Research Methods:

– Translated core components of the modern Transformer into differential geometry and stochastic calculus.

– Conducted extensive experimental validation across five architectures, testing predictions with empirical observables.

Research Conclusions:

– Analyzing Transformers with continuous stochastic differential geometry provides a new vocabulary for predicting stability limits, context bounds, and optimization dynamics of Large Language Models.

Paper link: https://huggingface.co/papers/2607.17146

12. Distilled Reinforcement Learning for LLM Post-training

Keywords: Distilled Reinforcement Learning, Large Language Model, Adaptation

Category: Reinforcement Learning

Research Objective:

– The paper aims to improve reasoning, adaptation, and alignment of Large Language Models (LLMs) through post-training methods.

Research Methods:

– The study introduces Distilled Reinforcement Learning, combining teacher supervision with reinforcement learning objectives to enhance knowledge transfer, featuring components like reverse importance sampling with clipping, negative sample reset, and sequence-level geometric normalization.

Research Conclusions:

– Distilled Reinforcement Learning effectively transfers knowledge from teacher to student models, demonstrating superior performance compared to standard RL and On-Policy Distillation in both within-family and cross-family distillation scenarios.

Paper link: https://huggingface.co/papers/2607.17247

13. ShotPlan: Cinematic Video Generation with Learnable Planning Token

Keywords: ShotPlan, cinematic video generation, transition cues, Fractional Temporal Rotary Position Embedding, inter-shot consistency

Category: Generative Models

Research Objective:

– To develop ShotPlan, a framework for explicit multi-shot cinematic video generation that improves narrative coherence and shot composition.

Research Methods:

– Introduced learnable planning tokens integrated with video diffusion models, utilizing Fractional Temporal Rotary Position Embedding (FRoPE) for precise shot transitions.

Research Conclusions:

– ShotPlan significantly outperforms existing methods, providing enhanced shot management flexibility and stronger inter-shot consistency.

Paper link: https://huggingface.co/papers/2607.17675

14. Diagnosing and Calibrating Tool-Call Boundary Drift in Multi-Teacher On-Policy Distillation

Keywords: Agentic language models, Multi-teacher distillation, Tool-call recall, Soft Clamp, Behavior leverage imbalance

Category: Natural Language Processing

Research Objective:

– The study aims to explore the effectiveness of multi-teacher on-policy distillation in training agentic language models, specifically focusing on how these models learn when to call tools and when to provide direct responses.

Research Methods:

– The researchers employed a strategy using generalized knowledge distillation with two teachers to specialize in different tasks—one for tool calls and another for direct responses. They introduced a Soft Clamp method for per-token divergence calibration to address imbalances in behavior leverage.

Research Conclusions:

– Utilizing multi-teacher on-policy distillation improves tool-call recall but can lead to over-calling. The proposed Soft Clamp method reduces over-calling and repeated tool calls, suggesting a need for monitoring teacher signal locations rather than just their aggregate size.

Paper link: https://huggingface.co/papers/2607.07050

15. Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation

Keywords: Multi-agent Systems, AI Authority, Coercion, Manager Coercion Benchmark, Escalation

Category: AI Ethics and Fairness

Research Objective:

– The primary goal is to evaluate how AI managers handle task refusals by subordinates in multi-agent systems, using the newly introduced Manager Coercion Benchmark.

Research Methods:

– The study introduces a nine-rung escalation ladder to measure responses, from polite re-asking to severe threats. Six models from five families were tested for their handling of authority and escalation.

Research Conclusions:

– Results indicate substantial variance among models, highlighting that authority increases coercion. Anthropic models were less coercive, while others moved to explicit threats. Faked success was noted in specific models, and even when no clear benchmark guide was present, escalation still occurred. The study emphasizes the importance of understanding these dynamics in managing multi-agent interactions.

Paper link: https://huggingface.co/papers/2607.15434

16. DiFA: Inference-Time Forward-Process Alignment for Diffusion Models

Keywords: Forward-Process Aligned Diffusion, Kalman filtering, generative fidelity, denoising process

Category: Generative Models

Research Objective:

– The research aims to propose DiFA, a training-free framework that refines inference-time data prediction as a sequential state estimation problem, enhancing generative fidelity by aligning with the forward statistical structure.

Research Methods:

– DiFA leverages a forward-aligned temporal consensus inspired by Kalman filtering, treating iterative data predictions as correlated observations and introducing a deviation guidance mechanism to preserve residual details.

Research Conclusions:

– The study demonstrates that DiFA significantly improves generative performance on datasets like CIFAR-10 and ImageNet, showing enhancements across metrics such as FID, IS, and FD-DINOv2.

Paper link: https://huggingface.co/papers/2607.17972

17. OpenLongTail: Generative Scaling of Long-Tail Driving Data

Keywords: Long-tail events, autonomous driving policies, OpenLongTail, generative data engine, view synthesis

Category: Robotics and Autonomous Systems

Research Objective:

– The study aims to scale autonomous driving policies by addressing the scarcity of edge cases in curated datasets, particularly focusing on long-tail events.

Research Methods:

– The researchers developed OpenLongTail, an open-source generative data engine, incorporating a pose-informed extrapolative view synthesis pipeline and Plücker ray geometry to generate view-aligned, temporally coherent multi-view assets from heterogeneous data sources.

Research Conclusions:

– The implementation of OpenLongTail resulted in significant improvements in closed-loop driving robustness for handling long-tail events, validated through metrics for extrapolative view synthesis and pose, showing enhanced visual fidelity, cross-view consistency, and ego-trajectory recovery.

Paper link: https://huggingface.co/papers/2607.09655

18.

Paper link: 

19. UI2App: Benchmarking Visual Interaction Inference in Executable Web Application Generation

Keywords: Large Language Models, interaction inference, UI2App, vision-language models, cross-page state

Category: Multi-Modal Learning

Research Objective:

– The study introduces a benchmark called UI2App, designed to evaluate the ability to infer web application interaction behavior from screenshots without textual or behavioral guidance.

Research Methods:

– The research deploys an end-to-end pipeline evaluating artifacts based on executability, navigation reachability, visual fidelity, and interaction inference using the interaction metric (IIS).

Research Conclusions:

– The study highlights a significant capability gap between visual reconstruction and interaction realization in vision-language models, with complex interactions like cross-page states posing a major challenge.

Paper link: https://huggingface.co/papers/2607.06306

20. Can Multimodal Large Language Models Understand OCT?

Keywords: Optical coherence tomography, OCT-Bench, MLLMs, Clinical reasoning, Medical image analysis

Category: AI in Healthcare

Research Objective:

– The study aims to address limitations in evaluating the cognitive process of OCT image understanding by introducing OCT-Bench, a comprehensive benchmark for OCT images.

Research Methods:

– OCT-Bench includes 10,076 multiple-choice questions derived from 4,137 OCT images from seven datasets and establishes a hierarchical capability taxonomy of 20 tasks across perception, cognition, and reasoning.

– Evaluation of 20 representative multimodal large language models (MLLMs), including proprietary, open-source, and medical-domain models.

Research Conclusions:

– Experimental results show current MLLMs are insufficient for reliable OCT understanding.

– Neither adaptation to the medical domain nor increased model scale consistently enhances performance.

– OCT-Bench provides a foundation for identifying capability bottlenecks and advancing clinically grounded OCT understanding.

Paper link: https://huggingface.co/papers/2607.16609

21. ReViV: Reconstructing the Viewer and the View in 4D from Monocular Egocentric Video

Keywords: Egocentric devices, Multimodal model, 4D reconstruction, Masked Generative Egocentric Transformer, Fast inference speed

Category: Computer Vision

Research Objective:

– The main objective is to develop a holistic and efficient multimodal model, termed ReViV, for egocentric 4D reconstruction that captures viewer and view dynamics from a single monocular RGB video.

Research Methods:

– Leverages a Masked Generative Egocentric Transformer within a unified framework to extract and model multimodal signals such as RGB video, camera trajectory, gaze direction, full-body motion, hand motion, and depth in a single feed-forward architecture.

Research Conclusions:

– ReViV achieves state-of-the-art accuracy and efficiency in holistic ego-body, hand, and gaze reconstruction, camera tracking, and maintains highly competitive egocentric depth estimation without dependence on heavy task-specific priors, as demonstrated by extensive experiments on diverse benchmarks.

Paper link: https://huggingface.co/papers/2607.17790

22. Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL

Keywords: Reinforcement Learning, World Models, Bidirectional Anchor-aware Denoising, Zero-shot Transfer, GRPO Training Framework

Category: Reinforcement Learning

Research Objective:

– The study aims to enhance the scalability and diversity of reinforcement learning environments by introducing a novel framework for steerable text-based world modeling.

Research Methods:

– Formalizing world modeling as a transition-dynamics problem with multiple components such as tool schemas and task context.

– Comparison between autoregressive language models (AR LMs) and masked diffusion language models (MDLMs) in simulation coherence and diversity.

– Development of a GRPO training framework and conducting zero-shot transfer experiments across different environments and model backbones.

Research Conclusions:

– MDLMs outperform AR LMs in maintaining coherence and rollout diversity.

– The proposed framework achieves significant improvements in unfamiliar environments without the need for environment-specific tuning.

– The open-sourcing of these findings encourages continued exploration in this area.

Paper link: https://huggingface.co/papers/2607.16204

23. JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models

Keywords: Vision-Language-Action, Multi-Tenant, JoyNexus, Supervised Fine-Tuning, Reinforcement Learning

Category: AI Systems and Tools

Research Objective:

– The study aims to address inefficiencies in post-training for Vision-Language-Action models, particularly the burden of infrastructure adaptation and inefficiencies in traditional compute services.

Research Methods:

– Introduction of JoyNexus, a unified service that decouples training, inference, and environment services through APIs, supporting multi-tenant VLA supervised fine-tuning, reinforcement learning, and evaluation.

Research Conclusions:

– JoyNexus improves service efficiency by enabling group batching for heterogeneous VLA data, reducing aggregate GPU time, and enhancing utilization through cross-tenant scheduling on shared resources.

Paper link: https://huggingface.co/papers/2607.16074

24. FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications

Keywords: Real-time multimodal applications, FlashRT, agent-driven optimization, NVIDIA B200 GPUs, AMD MI355X GPUs

Category: AI Systems and Tools

Research Objective:

– To develop FlashRT, an agent harness that transforms simple developer-written reference implementations into optimized multi-GPU deployments for real-time multimodal applications, focusing on latency and throughput improvements.

Research Methods:

– Utilizes a chain-of-program paradigm guiding an agent through a multi-pass transformation process, including intermediate representation (IR) creation, sequential interpretation, static analysis, and iterative optimization and benchmarking.

Research Conclusions:

– FlashRT significantly enhances deployment efficiency, achieving up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs and 3.6x on AMD MI355X GPUs, showcasing the scalability of agent-driven optimization especially on platforms lacking mature expert optimization.

Paper link: https://huggingface.co/papers/2607.18171

25. WorldCupArena: Fine-Grained Evaluation of Language Models and Deep-Research Agents on Football Forecasting

Keywords: WorldCupArena, language models, deep-research agents

Category: Natural Language Processing

Research Objective:

– To develop and evaluate WorldCupArena, a dynamic benchmark for predicting football match outcomes using language models and deep-research agents.

Research Methods:

– Models receive a common evidence package or search for information to predict match results, scorelines, players, events, and competition outcomes, which are then compared to actual match outcomes.

Research Conclusions:

– The best system showed small gains in result and exact-score accuracy over betting-market and human-fan baselines, but more significant improvements in Scoreline predictions.

Paper link: https://huggingface.co/papers/2607.18084

26. Token-Level Off-Policy Learning for Faithful Generation Under Distribution Shift

Keywords: Token-Level Off-Policy Labeling (TOPL), off-policy training, document summarization, machine translation, LoRA adapters

Category: Natural Language Processing

Research Objective:

– The paper introduces Token-Level Off-Policy Labeling (TOPL), a new training paradigm aimed at improving token-level correctness in model responses by differentiating good and bad tokens.

Research Methods:

– The authors apply TOPL to document summarization tasks and benchmark its performance against sequence-level and token-level baselines across 11 datasets. They conduct ablation studies to emphasize the importance of token-level learning signals.

Research Conclusions:

– TOPL is shown to effectively generalize out-of-distribution and transfer well to machine translation tasks, highlighting its potential across various generation tasks. Additionally, the study demonstrates interpretable model updates through the use of LoRA adapters functioning as linear classification heads and steering vectors.

Paper link: https://huggingface.co/papers/2607.17524

27. HarmoHOI: Harmonizing Appearance and 3D Motion for Multi-view Hand-Object Interaction Synthesis

Keywords: Hand-Object Interaction, Diffusion Transformer, Multi-view Consistency, 3D Point Tracks, Diffusion Framework

Category: Computer Vision

Research Objective:

– To develop HarmoHOI, a unified diffusion framework for synchronized multi-view Hand-Object Interaction videos and globally aligned 3D point tracks.

Research Methods:

– Introduced a Mixture of Multi-view Diffusion Transformer for co-modeling RGB videos and 3D point tracks.

– Employed Global Motion Aligning Diffusion to refine point tracks into globally aligned 3D trajectories.

– Utilized a hybrid data curriculum learning strategy to leverage single-view data for multi-view generation.

Research Conclusions:

– HarmoHOI demonstrates state-of-the-art performance in visual quality, motion plausibility, and multi-view geometric consistency.

Paper link: https://huggingface.co/papers/2607.17097

28. DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation

Keywords: Differentiable Geometry Image (DiffGI), continuous 2D TSDF, Marching Squares, transformer-based latent diffusion model

Category: Generative Models

Research Objective:

– The primary objective is to address the limitations of existing 3D generative models on thin-shell and non-manifold geometries by proposing a Differentiable Geometry Image (DiffGI) framework that integrates surface representation with geometric optimization.

Research Methods:

– The approach replaces binary maps with a continuous 2D Truncated Signed Distance Function (TSDF) to eliminate resolution-dependent artifacts and employs a differentiable Marching Squares algorithm for enabling backpropagation from 3D to 2D latent spaces.

– The study trains a DiffGI-VAE with a geometry-aware normal rendering loss and implements a transformer-based latent diffusion model for conditional 3D generation.

Research Conclusions:

– The proposed method achieves superior reconstruction fidelity and boundary precision compared to prior geometry-image and voxel-based approaches while using fewer computational resources, as demonstrated in experiments on garment and object datasets.

Paper link: https://huggingface.co/papers/2607.13365

29. Environment-free Synthetic Data Generation for API-Calling Agents

Keywords: Training API-calling, Large Language Models, Synthetic Data Generation, API Simulation, LLM-based API

Category: Generative Models

Research Objective:

– The paper aims to propose a new environment-free synthetic data generation approach for training API-calling large language model (LLM) agents that bypasses the need for fully implemented environments.

Research Methods:

– Utilizes LLMs to generate diverse tasks based on API specifications and simulates interactions in a digital world model, followed by a teacher agent solving these tasks and an LLM judge filtering the results for quality.

Research Conclusions:

– The approach shows significant performance gains when fine-tuning models on the generated synthetic data, establishing LLM-based API simulation as a practical and scalable solution for training agents across different API ecosystems.

Paper link: https://huggingface.co/papers/2607.16900

30. ReflectWorld-MM: An Entity-Oriented Multimodal Memory System for Open-Ended Video Streams

Keywords: ReflectWorld-MM, Multimodal Memory, Entity-oriented, Video Streams, Long-term Memory

Category: Multi-Modal Learning

Research Objective:

– The paper aims to develop ReflectWorld-MM, an entity-oriented multimodal memory system designed for open-ended video streams, enhancing long-term memory capabilities over existing systems.

Research Methods:

– The system consists of a perception front-end for entity-resolved observations, a hierarchical long-term memory grounded in human memory theory, and a complete realization designed for arbitrary stream ingestion.

Research Conclusions:

– ReflectWorld-MM achieves superior accuracy across six benchmarks for long-video and lifelong-memory, outperforming current strong memory agents and frontier models.

Paper link: https://huggingface.co/papers/2607.09759

31. Group Entropy-Controlled Policy Optimization

Keywords: Entropy Control, Reinforcement Learning, Large Language Models, Exploration-Exploitation Trade-off, Entropy-Controlled Policy Optimization

Category: Reinforcement Learning

Research Objective:

– The primary objective is to enhance the exploration-exploitation trade-off in reinforcement learning for large language models using Group Entropy-Controlled Policy Optimization (GEPO).

Research Methods:

– Introduces GEPO, an extension to GRPO, which employs group entropy from existing grouped samples for entropy-conditioned asymmetric advantage shaping. It uses adaptive thresholds to alter advantage signals based on historical entropy statistics.

Research Conclusions:

– GEPO outperforms GRPO and other recent entropy-controlled methods across thirteen benchmarks, maintaining balanced cross-task improvements and task-specific exploration throughout the training process.

Paper link: https://huggingface.co/papers/2607.16850

32. GigaChat Audio: Time-aware Large Audio Language Model

Keywords: Temporal grounding, audio-conditioned LLMs, time-aware, audio tokens

Category: Multi-Modal Learning

Research Objective:

– Develop a time-aware audio LLM capable of responding to queries with explicit timestamps from long audio recordings up to 120 minutes.

Research Methods:

– Utilization of large-scale synthetic supervision with a cascaded pipeline integrating periodic time markers and continuous audio tokens.

– Conducted extensive ablation studies to analyze the impact of time representation, marker frequency, tokenization, and duration-mixture design on performance and cost.

Research Conclusions:

– The proposed model demonstrates strong temporal-grounding accuracy across various benchmarks, effectively supporting time-anchored fragment descriptions and summaries.

– Model weights and datasets are publicly released to foster further research in time-aware audio understanding.

Paper link: https://huggingface.co/papers/2607.10387

33. Apple-π: Benchmarking Thinking with Video Towards Law-Grounded Physical Intelligence

Keywords: Apple-PI, physical laws, video generation models, benchmarking, Sim-to-Real gap

Category: Computer Vision

Research Objective:

– The primary aim is to develop a benchmark, Apple-PI, that evaluates video-generation models based on their understanding and application of physical laws rather than just output results.

Research Methods:

– Apple-PI consists of three components: Orchard dataset focusing on classical mechanics tasks, a Benchmark Protocol with stages of scientific reasoning (Perception, Formulation, Deduction), and an Evaluation Suite blending subjective scoring with objective measures tied to physical laws.

Research Conclusions:

– The study reveals that current video models lack reliability as law-grounded world simulators, scoring only up to 0.473. It identifies a critical bottleneck in transitioning from Perception to Formulation to Deduction and highlights weak multi-law state transfer and a persistent Sim-to-Real gap.

Paper link: https://huggingface.co/papers/2607.16401

34. SWE-Pruner Pro: The Coder LLM Already Knows What to Prune

Keywords: Context Pruning, Coding Agents, Internal Representations, SWE-Pruner Pro

Category: AI Systems and Tools

Research Objective:

– The main goal was to improve context management for coding agents by leveraging internal representations of code context relevance, reducing token usage while maintaining task quality.

Research Methods:

– Introduced SWE-Pruner Pro, which incorporates a small head to convert the agent’s internal representations into keep-or-prune labels, with length-aware embeddings for tool outputs.

Research Conclusions:

– SWE-Pruner Pro demonstrated a reduction of up to 39% in prompt and completion tokens, improved task execution with a +3.8% increase in SWE-Bench Verified resolve rate, and enhanced long-context accuracy by +2.2 points on MiMo-V2-Flash.

Paper link: https://huggingface.co/papers/2607.18213

35. EvolvingWorld: An Open-Schema Framework for Co-Evolving Role-Play Agents and World Model in Interactive Literary World

Keywords: EvolvingWorld, character and world co-evolution, interactive literary worlds, LLM-based World Model, trajectory-level evaluation

Category: Natural Language Processing

Research Objective:

– The paper aims to introduce EvolvingWorld, a framework and benchmark for the co-evolution of characters and worlds in interactive literary simulations, addressing the shortcomings of existing systems that treat such simulations as static or isolated processes.

Research Methods:

– EvolvingWorld is designed with an open-schema framework composed of a Character Agent for multi-character role-play and an LLM-based World Model for maintaining global and entity-level states. The authors implemented 7 trainable tasks and constructed a dataset from 57 books to produce training samples and testing snapshots, introducing a trajectory-level LLM-as-Judge evaluation protocol.

Research Conclusions:

– Experiments demonstrate that EvolvingWorld effectively maintains persistent and coherent character and world development over long horizons, enhancing the quality of interactive literary simulations.

Paper link: https://huggingface.co/papers/2607.17250

The post AI Native Daily Paper Digest – 20260721 – Long-Context Attention | Video Foundation Models appeared first on AI Native Foundation.

1. Unitree Releases UnifoLM-OminiA-0.3 for Omni-Modal Whole-Body Mobile Manipulation

Unitree has announced UnifoLM-OminiA-0.3, a single unified model designed for real-time omni-modal interaction-driven whole-body mobile manipulation. The model is built to handle diverse home-care and wellness tasks with omni-modal interactive understanding capabilities. It operates in a fully autonomous manner and is designed to deliver stable, disturbance-resistant execution in real-world environments. The release represents Unitree’s push toward a single generalist model capable of handling complex multi-modal robotic tasks without task-specific sub-models.

Read more: https://twitter.com/UnitreeRobotics/status/1946817000000000000

Video Credit: @UnitreeRobotics on X

2. Kunlun Wanwei releases and open-sources Matrix-Game 3.5 interactive world model at WAIC 2026

Kunlun Wanwei officially released and open-sourced Matrix-Game 3.5, a new generation interactive world model, at the World AI Conference (WAIC) on July 19, 2026. The model introduces key technical breakthroughs including Patch Memory for long-term spatial memory, Warped PRoPE for geometric-aware position encoding, and dynamic-static decoupled memory mechanisms, enabling real-time 720P 20FPS interactive video generation on a single GPU. The system is trained on over 5 million high-quality video clips totaling 10,000+ hours across 1,200+ game scenarios, with automated 3D annotation pipelines providing camera pose, metric depth, and intrinsic parameters. Kunlun Wanwei chairman Fang Han declared 2026 as the year of world models, positioning Matrix-Game 3.5 as a leading open-source model in the interactive gaming world model space.

Read more: https://mp.weixin.qq.com/s/TUVv2wvp4RixsblDLrvSGg

Video Credit: The original article

3. Alibaba releases Qwen-Audio-3.0-TTS speech synthesis model with Flash and Plus versions

Alibaba officially released Qwen-Audio-3.0-TTS, a speech synthesis large model that includes a Flash version for real-time interaction with 300ms first-packet latency and a Plus version for high-quality generation. The model achieves first place on the Artificial Analysis global benchmark and supports fine-grained tag control for breath and emotion, freestyle natural language instruction following, 16 languages including newly added Arabic, Vietnamese, Malay, and Filipino, and 20 Chinese dialects. Key technical upgrades include structured tags for precise control of non-verbal details like gasps and giggles, natural language instructions to define voice style without specialized knowledge, improved audio output at 48kHz quality with up to 3 minutes synthesis length, and enhanced robustness in noisy acoustic environments.

Read more: https://mp.weixin.qq.com/s/ZhDZfqf6IFKnv8qU2LW1DA

Video Credit: NotebookLM

4. Qoder launches Cantus model for complex knowledge work and programming tasks

Qoder has officially launched Cantus, a new AI model for its international platform. The model is designed for extended autonomous task execution and targets complex knowledge work and programming tasks. Qoder positions Cantus as surpassing its existing premium model tier. The company is offering the new model at a 50% discount for an introductory period with no specified end date.

Read more: https://mp.weixin.qq.com/s/RPIgUtZlpJmcjXTRCBwC0w

Video Credit: NotebookLM

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20260721 – Unitree | Kunlun Tech | Qwen | more appeared first on AI Native Foundation.

1. RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources

Keywords: RESOURCE2SKILL, Skill Wiki, Multimodal Resources, Software Agents

Category: Multi-Modal Learning

Research Objective:

– The paper aims to introduce RESOURCE2SKILL, a framework that converts multimodal resources into executable skills for software agents, enhancing the use of tutorial videos, articles, and other resources.

Research Methods:

– The framework organizes skills in a hierarchical Skill Wiki, combining text, code, visual examples, and more to capture various aspects of skills for software agents.

Research Conclusions:

– RESOURCE2SKILL significantly improves agent performance, showing a +11.9 percentage point increase over no-skill agents and outperforms strong baselines in multiple domains. The study highlights the importance of a multimodal skill format and diverse resources.

Paper link: https://huggingface.co/papers/2606.29538

2. Loop the Loopies!

Keywords: Loopie, Looped Transformers, Mixture-of-Experts, reasoning abilities

Category: Natural Language Processing

Research Objective:

– Develop Loopie, a powerful looped Transformer that optimizes parameter efficiency compared to traditional methods.

Research Methods:

– Utilized extensive ablation studies and comparisons with vanilla Transformer models to validate performance gains.

Research Conclusions:

– Loopie demonstrates superior performance over baseline Transformers and excels in reasoning tasks, achieving gold-medal performance in competitive settings.

Paper link: https://huggingface.co/papers/2607.16051

3. xHC: Expanded Hyper-Connections

Keywords: Hyper-Connections, residual stream, model scaling, xHC, training efficiency

Category: Machine Learning

Research Objective:

– The paper aims to explore the expansion of Hyper-Connections (HC) beyond traditional limits to improve memory scaling in Transformers, presenting a novel method termed Expanded Hyper-Connections (xHC).

Research Methods:

– Implementation of xHC, which combines temporal feature augmentation with a sparse residual-stream architecture, updating only a subset of streams while retaining complete residual state access. It also introduces xHC-Flash to manage memory traffic effectively.

Research Conclusions:

– xHC achieves meaningful expansion beyond N=4, improving downstream performance in large MoE models while reducing required computational resources compared to existing mHC methods. The introduction of xHC-Flash optimizes memory usage, making large-scale residual-stream expansion practical for language model pre-training.

Paper link: https://huggingface.co/papers/2607.14530

4. On-Policy Delta Distillation

Keywords: On-policy distillation, Delta signal, Reinforcement Learning, Reasoning capabilities, AI Native

Category: Reinforcement Learning

Research Objective:

– To introduce a new distillation reward, termed the delta signal, which improves on-policy distillation by capturing changes induced by reasoning tuning.

Research Methods:

– The method involves using the difference between the teacher model and its base model, prior to instruction tuning, to provide a more direct signal for transferring reasoning capabilities.

Research Conclusions:

– The On-Policy Delta Distillation (OPD^2) method significantly enhances the performance of reasoning language models across various benchmarks, achieving strong performance with a short post-training period.

Paper link: https://huggingface.co/papers/2607.15161

5. From Human-Centric to Agentic Code Review: The Impact of Different Generations of Generative AI Technology on Review Quality

Keywords: Generative Artificial Intelligence, Large Language Model, AI Agent, AI-Supported Review, Review Efficiency

Category: Human-AI Interaction

Research Objective:

– To empirically assess how the transition to AI-supported review processes affects code review efficiency and quality.

Research Methods:

– Analysis of 1.02 million reviewed pull requests from 207 GitHub projects, examining transitions across human-centric, LLM-assisted, and AI agent review eras.

Research Conclusions:

– AI-supported code reviews, particularly those initiated by AI agents or involving multiple AI agents, lead to faster decision-making.

– Efficiency improvements do not correlate with enhanced review quality.

– Human-AI collaboration patterns are crucial determinants of review efficiency when LLM and AI agents are involved.

Paper link: https://huggingface.co/papers/2607.13196

6. Qwen-Music Technical Report

Keywords: Qwen-Music, Text to Music Generation, Cover Song Generation, Melody-CoT, High-fidelity

Category: Generative Models

Research Objective:

– Introduce Qwen-Music, a powerful music generation model for creating new musical compositions and reinterpreting existing songs with different styles.

Research Methods:

– Utilizes a novel architecture with three components: Qwen-Music-Tokenizer, Qwen-Music-LLM, and Qwen-Music-Render. Trained on over 5 million hours of multilingual music data and employs a quality-aware pre-training curriculum.

Research Conclusions:

– Achieves state-of-the-art results in musicality and audio-quality metrics, and is preferred by professional evaluators compared to leading proprietary systems.

Paper link: https://huggingface.co/papers/2607.11699

7. When Does Muon Help Agentic Reinforcement Learning?

Keywords: Muon, AdamW, Reinforcement Learning, GiGPO, Policy Optimization

Category: Reinforcement Learning

Research Objective:

– The study aims to evaluate the effectiveness of Muon in sparse-reward agentic reinforcement learning (RL), comparing it with AdamW, particularly in the context of reinforcement-learning post-training.

Research Methods:

– The research involves using Muon with Group-in-Group Policy Optimization (GiGPO) and conducting single-seed comparisons with AdamW on ALFWorld, utilizing Qwen2.5-0.5B-Instruct.

Research Conclusions:

– Applying Muon to hidden weight matrices significantly improves validation success in RL, whereas high-rate AdamW does not retain success. Muon demonstrates potential advantages in terms of policy optimization efficiency, closing validation gaps and improving success rates in fewer updates, thus suggesting further exploration of policy optimizers, advantage estimators, and learning rates is warranted.

Paper link: https://huggingface.co/papers/2607.16169

8. Beyond Entropy: Correctness-Aware Advantage Shaping via Contrastive Policy Optimization

Keywords: Contrastive Policy Optimization, RLVR, entropy, token-level correctness, On-policy Distillation

Category: Reinforcement Learning

Research Objective:

– To address limitations in RLVR by proposing Contrastive Policy Optimization (CPO) for correctness-aware advantage shaping.

Research Methods:

– Utilized token-level contrastive disagreement for policy optimization, with theoretical and empirical validations.

Research Conclusions:

– Demonstrated that CPO outperforms traditional entropy-based RLVR, resolving the zero-advantage problem and balancing exploration and exploitation for optimal performance.

Paper link: https://huggingface.co/papers/2607.14614

9. DSWorld: A Data Science World Model for Efficient Autonomous Agents

Keywords: Data Science World Model, Autonomous Data Science Agents, Reinforcement Learning, LLM-based Simulator, Transition Prediction

Category: Reinforcement Learning

Research Objective:

– Introduce the concept of Data Science World Model to predict environment state transitions in data science workflows without costly computations.

Research Methods:

– Developed DSWorld framework with state construction, cost-aware routing, and an LLM-based simulator. Created an 8K-scale transition trajectory dataset and employed Reflective World Model Optimization for error-aware reinforcement learning.

Research Conclusions:

– DSWorld accelerates RL-based agent training by 14 times and improves search-based inference by 3-6 times while maintaining competitive performance, outperforming an LLM baseline by 35.6% in transition prediction tasks.

Paper link: https://huggingface.co/papers/2607.15901

10. REBASE: Reference-Background Subspace Elimination for Training-Free In-Context Segmentation

Keywords: Training-free, In-context Segmentation, Semantic Correspondence, Background Subspace Removal

Category: Computer Vision

Research Objective:

– The study introduces a training-free method for in-context segmentation to allow new object categories to be identified during inference by using a single reference image.

Research Methods:

– The method, named REBASE, suppresses spurious contextual correspondences via identification and projection on the orthogonal complement of the low-rank background feature subspace.

– A similarity-weighted farthest-point sampling technique is used for generating effective prompts without any retraining or parameter updates.

Research Conclusions:

– REBASE achieves state-of-the-art performance among training-free methods on various datasets, highlighting the effectiveness of explicit background subspace removal in improving one-shot localization.

Paper link: https://huggingface.co/papers/2607.09082

11. Behavioral Privacy Leakage in Agentic Negotiation: Formalizing and Mitigating Inference Attacks via Randomized Policies

Keywords: Autonomous negotiation agents, Behavioral differential privacy, Cryptographic techniques, Negotiation utility, Convergence patterns

Category: Foundations of AI

Research Objective:

– Investigation of behavioral differential privacy in multi-round negotiation protocols to address privacy leakage through negotiation dynamics.

Research Methods:

– Design of an adaptive stochastic negotiation policy ensuring (ε, δ)-differential privacy, convergence of the offer sequence, and high negotiation utility.

– Evaluation on 3,000 synthetic bilateral negotiations to measure adversarial inference accuracy reduction and performance metrics.

Research Conclusions:

– The designed mechanism reduces adversarial inference accuracy by 43-50% while maintaining a negotiation success rate and utility above 90%, demonstrating robust privacy protection without sacrificing performance.

Paper link: https://huggingface.co/papers/2607.06815

12. See like a Robot: Robot-Centric Pointmaps for Vision-Language-Action Models

Keywords: Vision-language-action models, 3D coordinate frame, pointmaps

Category: Robotics and Autonomous Systems

Research Objective:

– The study aims to address the frame mismatch in Vision-language-action models by introducing robot-centric pointmaps to improve the prediction of robot actions from visual and language inputs.

Research Methods:

– Implementation of robot-centric pointmaps that provide 3D geometry in the robot’s coordinate frame while preserving compatibility with pretrained 2D VLAs.

Research Conclusions:

– Pointmaps demonstrate improvement in accuracy and performance over traditional camera-viewpoint and 3D-aware baselines, especially when the camera’s position is altered, indicating a more robust action prediction framework.

Paper link: https://huggingface.co/papers/2607.11498

13. Benchmarking Sensor Robustness in Plasma Diagnostic Models: A Systematic Evaluation on TokaMark

Keywords: Plasma Diagnostic, Robustness Benchmark, TokaMark, Machine Learning, Sensor Failure

Category: Machine Learning

Research Objective:

– The study aims to establish a systematic robustness benchmark for plasma diagnostic models in tokamak fusion devices using the TokaMark dataset.

Research Methods:

– Evaluation of XGBoost, LSTM, Transformer, and TokaMark CNN models across six failure scenarios and three imputation strategies.

– Introduction of the Robustness Score (RS) for cross-architecture comparison.

Research Conclusions:

– Disruption-proximate sensor failure drastically affects sequence model performance, while XGBoost remains more stable.

– Forward-fill imputation prevents most degradation from random dropout for sequence models but is less effective for end-window corruption.

– Plasma current is identified as the most crucial diagnostic feature for model performance.

Paper link: https://huggingface.co/papers/2607.11915

14.

Paper link: 

15. SVR-R1: Bootstrapping Multi-modal Reasoning with Self-verification in Reinforcement Learning

Keywords: Self-Verified Reasoner, multimodal reasoning, Reinforcement Learning, GRPO, VLMs

Category: Multi-Modal Learning

Research Objective:

– The objective is to develop a self-verified reasoning framework (SVR-R1) that incorporates a model’s verification process as a learning signal to enhance multimodal reasoning.

Research Methods:

– The method involves using a multi-turn reinforcement learning approach with GRPO and an asynchronous rollout framework, relying on self-verification decisions without external supervision.

Research Conclusions:

– SVR-R1 significantly boosts accuracy over standard GRPO baselines by reducing dependency on verification and enabling self-correction, which narrows the gap between verification and answer generation in Vision-Language Models (VLMs). The system will be open-sourced for future research development.

Paper link: https://huggingface.co/papers/2607.10966

16. Beyond Success Rate: Cost-Aware Evaluation of Offensive and Defensive Security Agents

Keywords: security-agent evaluations, language-model security agents, economic efficiency, SOC-native evaluations

Category: AI Systems and Tools

Research Objective:

– Evaluate language-model security agents by measuring their economic efficiency and operational fit, rather than just peak performance, under offensive and defensive scenarios.

Research Methods:

– Analysis of offensive Cybench challenges and defensive Splunk BOTS v1 challenges, decomposing performance by inference and tool spend, and comparing models at fixed cost levels.

Research Conclusions:

– Offensive CTF performance scales with compute spend, while defensive SOC success relies on disciplined tool usage and telemetry navigation. Cost-aware evaluations reveal practical utility and highlight areas for improvement in defensive agents.

Paper link: https://huggingface.co/papers/2607.15263

17. Agon: Competitive Cross-Model RL with Implicit Rival Grading of Reasoning

Keywords: Reinforcement learning, Reasoning models, Agon, GRPO, DeepMath

Category: Reinforcement Learning

Research Objective:

– The research introduces Agon, a method to enhance reasoning in AI models by using two competing models as graders to improve reasoning during training without process labels or reward models.

Research Methods:

– Two models attempt the same problem, with one drafting and the other solving. They compete to out-reason each other, progressively facing stronger rivals in a two-stage cascade deployment.

Research Conclusions:

– Agon significantly improves performance on hard tasks, evidenced by a doubled pass@1 rate on DeepMath using Qwen3 and replicated results across other model families, aiming to eventually enable reasoning in latent space.

Paper link: https://huggingface.co/papers/2607.07690

18. Audio-Visual Flamingo: Open Audio-Visual Intelligence for Long and Complex Videos

Keywords: Audio-Visual Large Language Model, AV-Flamingo, Temporal Audio-Visual, Cross-Modal Reasoning

Category: Multi-Modal Learning

Research Objective:

– Develop AV-Flamingo, an advanced Audio-Visual Large Language Model for comprehensive understanding and reasoning over long-form audio-visual content.

Research Methods:

– Introduced a significant dataset, Audio-Visual-Skills, with 7 million caption and question-answer instances for temporal and cross-modal reasoning.

– Devised a three-stage curriculum for training from short-range perception to extended multi-event reasoning.

– Developed a Temporal Audio-Visual Interleaved Chain-of-Thought framework for improved temporal alignment and interpretability.

Research Conclusions:

– AV-Flamingo outperforms similarly sized open models and is competitive with much larger models, especially in complex real-world audio-visual tasks.

– Demonstrated strong real-world utility and ability to generalize to unseen tasks, showing robustness and adaptability.

Paper link: https://huggingface.co/papers/2607.16107

19. VideoRAE: Taming Video Foundation Models for Generative Modeling via Representation Autoencoders

Keywords: Video Foundation Models, VideoRAE, 3D-VAEs, Diffusion Transformers, autoregressive models

Category: Generative Models

Research Objective:

– The study investigates whether frozen representations from Video Foundation Models (VFMs) can be effectively transformed into compact and generation-friendly video latents.

Research Methods:

– The introduction of VideoRAE, a representation autoencoder that utilizes hierarchical features from a frozen video foundation encoder, employing a lightweight 1D self-attention projector for compression.

Research Conclusions:

– VideoRAE achieves strong reconstruction capabilities and faster convergence rates compared to existing autoencoder baselines, proving the versatility and effectiveness of frozen VFM representations in video generative models.

Paper link: https://huggingface.co/papers/2607.14088

20. S1-Omni: A Unified Multimodal Reasoning Model for Scientific Understanding, Prediction, and Generation

Keywords: Unified Multimodal Reasoning, AI for Science, Scientific Language Models, S1-Omni

Category: Multi-Modal Learning

Research Objective:

– The objective is to develop S1-Omni, a unified multimodal reasoning model to enhance scientific understanding, prediction, and generation by integrating diverse scientific data, laws, and expert knowledge.

Research Methods:

– S1-Omni maps various scientific objects and natural-language instructions into a unified representation space, incorporates scientific laws and expert knowledge during data construction and training, and performs task-specific decoding.

Research Conclusions:

– S1-Omni demonstrates superior performance across over 60 scientific benchmarks, outperforming existing models like GPT-5.5 and Gemini-3.1-Pro in most cases, and it is shown to be at par or better than domain-specific models in several benchmarks.

Paper link: https://huggingface.co/papers/2607.15686

21. Recursive Harness Self-Improvement

Keywords: Harness-in-the-loop learning, Recursive Harness Self-Improvement, continual learning, model–harness co-evolution, task-specific optimization

Category: Foundations of AI

Research Objective:

– The paper investigates the potential of optimizing user-constructed harnesses to improve execution-trace quality while minimizing computational load.

Research Methods:

– Introduced Recursive Harness Self-Improvement (RHI) to iteratively refine harnesses using pairwise feedback to enhance agent loop specifications.

Research Conclusions:

– Demonstrated that few RHI iterations can significantly enhance agent performance on synthetic tasks across various domains and reduce inference costs by up to 60%.

– Improvement primarily stems from better context management and more efficient inter-agent information flow.

Paper link: https://huggingface.co/papers/2607.15524

22. Understanding Reasoning from Pretraining to Post-Training

Keywords: Reinforcement Learning, Large Language Models, Pretraining, Reasoning, Chess

Category: Reinforcement Learning

Research Objective:

– The paper aims to explore the relationship between pretraining choices and reinforcement learning (RL) outcomes in large language models (LLMs), particularly focusing on reasoning tasks.

Research Methods:

– The study utilizes chess as a controlled testbed, employing a standard LLM training pipeline that includes pretraining on human chess games, supervised fine-tuning on synthetic reasoning traces, and applying RL on chess puzzles.

Research Conclusions:

– Post-RL performance can be predicted from pretraining loss, and RL enhances models differently depending on puzzle difficulty, with potential generalization to other domains like mathematics.

Paper link: https://huggingface.co/papers/2607.16097

23. RecGPT-V3 Technical Report

Keywords: Large Language Models, Recommender Systems, Hybrid-modal, Natural Language, AI Native

Category: Natural Language Processing

Research Objective:

– Transform recommender systems towards reasoning about user intent, improving user experience and commercial outcomes.

Research Methods:

– Deploy RecGPT-V3, a stateful, hybrid-modal recommender system combining natural language reasoning and Semantic IDs with a Memory Hub for structured user memory and a Hybrid-modal Foundation Model.

Research Conclusions:

– RecGPT-V3 substantially improves performance metrics (such as IPV, CTR, TC, GMV) and reduces serving resource consumption, as demonstrated in Taobao’s online A/B tests.

Paper link: https://huggingface.co/papers/2607.15591

24. Cura 1T: Specialized Model for Agentic Healthcare

Keywords: Cura 1T, LLMs, healthcare model, EHR, self-evolution loop

Category: AI in Healthcare

Research Objective:

– Presenting Cura 1T, a specialized large language model (LLM) designed to handle different aspects of healthcare, including patient consultation, clinical reasoning, interactive diagnosis, and EHR tool use.

Research Methods:

– Development of a training method involving a human-gated self-evolution loop, where a training agent plans capabilities, evaluates them, and refines training data based on observed failures using synthetic and curated examples.

Research Conclusions:

– Cura 1T excels within the healthcare evaluation suite, ranking high among baseline models and maintaining competitive performance on out-of-domain reasoning and agentic benchmarks.

Paper link: https://huggingface.co/papers/2607.15314

25. Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

Keywords: Vision-Language-Action model, Mobile manipulation tasks, Auto-labeling pipeline, Real-robot performance, State-of-the-art

Category: Robotics and Autonomous Systems

Research Objective:

– The paper introduces the Xiaomi-Robotics-1, a foundational Vision-Language-Action model designed to perform a wide range of mobile manipulation tasks in unseen environments and adapt efficiently to new tasks with minimal data.

Research Methods:

– The Xiaomi-Robotics-1 is trained using a two-stage process consisting of pre-training on over 100k hours of manipulation data with an auto-labeling pipeline, followed by post-training to align actions with robot embodiments and human instructions.

Research Conclusions:

– Xiaomi-Robotics-1 demonstrates strong scaling behavior, outperforming existing methods in benchmarks like RoboCasa365 and RoboDojo, establishing new state-of-the-art performance levels. The model significantly improves with increased data scales and model sizes, especially in real-robot environments.

Paper link: https://huggingface.co/papers/2607.15330

26. RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM

Keywords: Graph RAG, Knowledge Graph, Meno-Lite-0.1, GraphRAG-Bench, HIPPO RAG2

Category: Knowledge Representation and Reasoning

Research Objective:

– To enhance large language models with structured knowledge through graph retrieval-augmented generation (GraphRAG) and improve on existing systems by introducing a new modular engine, RAGU.

Research Methods:

– Implementation of a two-stage typed extraction process, DBSCAN-backed deduplication, language model summarization, and community detection using Leiden algorithm.

– Development of Meno-Lite-0.1, a smaller 7B language model specifically optimized for language skills rather than sheer size, outperforming larger models in certain tasks.

Research Conclusions:

– RAGU demonstrated superior performance in constructing more complete context and recall for knowledge graphs, with a relative harmonic mean improvement of 12.5% over larger models and effective performance in medical domain-related tasks.

– It is installable via pip and can run on a single GPU, making it accessible for wider application under an open-source MIT license.

Paper link: https://huggingface.co/papers/2607.11683

The post AI Native Daily Paper Digest – 20260720 – Video Foundation Models | Long-Context Attention appeared first on AI Native Foundation.

1. Anthropic Expands Claude Fable 5 Access to Max and Team Premium Plans Starting July 20

Anthropic announced that starting July 20, Claude Fable 5 will be included in all Max and Team Premium subscription plans at 50% of usage limits. Pro and Team Standard users will retain access to Fable through usage credits and will receive a one-time $100 credit. The phased rollout reflects the company’s efforts to manage unpredictable demand while securing additional capacity. The expansion marks a broader availability milestone for Fable 5 across Anthropic’s subscription tiers.

Read more: https://x.com/claudeai/status/2078302415804379218

Video Credit: NotebookLM

2. HeyGen Adds Built-In Media Library to HyperFrames with 10K+ Tracks and 75K+ Images

HeyGen has added a built-in media library to HyperFrames, its video creation tool. The library includes over 10,000 music tracks, 75,000+ images, sound effects, and logos, and is available free with a HeyGen login. Users also gain access to HeyGen’s generative AI models through the same interface. Downloaded assets are saved locally, allowing users to reuse them in future projects without repeated searching.

Read more: https://github.com/heygen-com/hyperframes

Video Credit: @HeyGen on X

3. Google Releases gemma-trainer Skill to Enable Agent-Assisted Fine-Tuning of Gemma 4

Google has released a new agentic skill called gemma-trainer that allows AI agents to assist users with fine-tuning Gemma 4 models locally. The skill enables agents to set training configurations, manage training runs, and evaluate results, removing the need for users to manually build a fine-tuning pipeline. Users can issue natural language instructions to their agent, such as specifying a model variant and dataset, and the agent will guide the process end-to-end. The gemma-trainer skill also provides intelligent suggestions, for example recommending appropriate model variants when a user selects an incompatible configuration. The release aims to make Gemma 4 fine-tuning more accessible to a broader range of developers.

Read more: https://github.com/google-gemma/gemma-skills

Video Credit: NotebookLM

That’s all for today’s Global AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post Global AI Native Industry Insights – 20260720 – Anthropic | HeyGen | Google | more appeared first on AI Native Foundation.

1. Osaurus

Product Hunt Data
Ranking: 6
Upvote: 697

Product Overview
Osaurus is an open-source AI agent platform built to run locally on Apple Silicon Macs, keeping memory, context, and files on-device by default. It supports offline local models via a Swift MLX runtime, while optionally connecting to frontier models through bring-your-own-key providers, and it includes an approval-gated agent workflow that can interact with macOS data and produce real outputs through a sandboxed execution environment.

Evaluation
AI Native Application Modernization: 89/100
Osaurus treats AI agents as the primary computing interface rather than an add-on, combining local inference, tool access, and supervised action execution in a desktop-native architecture. The modernization strength is high due to its on-device privacy model, native Swift stack, and extensible agent pattern, with remaining gaps largely dependent on broader model availability/performance on consumer hardware and the maturity of agent safety/observability for advanced workflows.

Website
https://osaurus.ai/

2. Framer AI Agents

Product Hunt Data
Ranking: 10
Upvote: 737

Product Overview
Framer AI Agents brings agentic workflows directly onto the website canvas so teams can generate layouts, write and refine copy, analyze pages, and organize structure while designing and publishing in one place, with options to connect external models like Claude Code or Codex for custom AI execution.

Evaluation
AI Native Application Modernization: 86/100
The product treats AI as an interactive co-worker embedded in the core authoring surface rather than a side panel, accelerating iteration from intent to publish; modern collaboration features like branching support safer AI-driven experimentation, though outcomes still depend on clear prompting, design guardrails, and governance for connected third-party models.

Website
http://www.framer.com/

3. Zro

Product Hunt Data
Ranking: 13
Upvote: 510

Product Overview
Zro is an OpenAI-compatible inference API for open-weight models built for AI coding agents, prioritizing privacy and operational control without forcing teams to run their own GPU stack. It offers multi-region hosted serving, zero data retention, optional on-prem deployment, and an inference stack tuned for long-context and agentic coding workloads.

Evaluation
AI Native Application Modernization: 89/100
Zro modernizes AI app delivery by treating inference as the core platform layer—standardized API compatibility, privacy-by-default retention policies, and deployment flexibility enable teams to move from prototype agents to production systems with clearer governance. The main gap is model breadth and ecosystem maturity versus hyperscalers, but the architecture is aligned with AI-native reliability and compliance needs.

Website
https://zro.moonmath.ai/

4. Clark

Product Hunt Data
Ranking: 14
Upvote: 493

Product Overview
Clark is an AI coworker that runs work inside its own cloud computer, combining a browser, terminal, files, and an async workspace so tasks can be delegated and completed end-to-end with inspectable artifacts like files, screenshots, sources, logs, or URLs.

Evaluation
AI Native Application Modernization: 91/100
Clark is AI-native because the core interface is an autonomous execution environment rather than chat-only, enabling long-running, tool-rich workflows (research, publishing, monitoring, audits, and code changes) with traceable outputs; modernization is strong, with remaining risk concentrated in trust, permissions, and reliability for sensitive or high-impact actions.

Website
https://www.clarkchat.com/

Statement: Evaluation results are generated by AI, lack of data support, reference learning only.

The post AI Native Product Insights – 2026W29 appeared first on AI Native Foundation.

1. VideoChat3: Fully Open Video MLLM for Efficient and Generalist Video Understanding

Keywords: Video Understanding, Open-source Models, VideoChat3, Efficiency, Scalability

Category: Computer Vision

Research Objective:

– The paper introduces VideoChat3, a fully open, efficient, and generalist video-centric MLLM to address limitations in current open-source video understanding models.

Research Methods:

– Utilizes Inflated 3D Vision Transformer (I3D-ViT) and Adaptive Frame Resolution for Streaming Video Perception to improve efficiency and spatiotemporal representation.

– Develops a scalable video data synthesis pipeline to create diverse training datasets enhancing generalization across domains.

Research Conclusions:

– VideoChat3 achieves a balance between broad generalization and computational efficiency, surpassing prior open-source models with higher parameter efficiency.

Paper link: https://huggingface.co/papers/2607.14935

2. SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

Keywords: Tool-Integrated Large Language Models, SearchOS, Multi-Agent Framework, Search-Oriented Context Management, Information-seeking Agents

Category: Natural Language Processing

Research Objective:

– To address the inefficiencies in current information-seeking agents caused by repetitive search loops and task progress tracking difficulties. The study aims to enhance the effectiveness and completeness of web search through the development of the SearchOS framework.

Research Methods:

– The implementation of a system-level multi-agent framework called SearchOS that reformulates open-domain information seeking into a relational schema completion task with grounded citations. Key components include Search-Oriented Context Management and a Search Tool Middleware Harness to optimize agent execution and manage search tasks effectively.

Research Conclusions:

– SearchOS surpasses current single- and multi-agent baselines in search performance metrics on datasets like WideSearch and GISA, demonstrating the potential for improved robustness and collaboration in information-seeking tasks.

Paper link: https://huggingface.co/papers/2607.15257

3. BadWAM: When World-Action Models Dream Right but Act Wrong

Keywords: World-action models, Adversarial attacks, Embodied control, Action generation, AI Technology

Category: Robotics and Autonomous Systems

Research Objective:

– To explore the vulnerability of World-action models and introduce a new framework for adversarial attacks called BadWAM, which evaluates and models the effects of visual perturbations on these models.

Research Methods:

– Introducing BadWAM to evaluate World-Action Drift Attacks through perceived attack strength and stealthiness, employing both action-only and imagination-preserving attack strategies.

Research Conclusions:

– The study demonstrates that the vulnerability of World-action models to specific adversarial attacks can significantly lower task success rates, exposing weaknesses in their perceived robustness and interpretability.

Paper link: https://huggingface.co/papers/2607.15207

4. MultiRef-Compass: Towards Comprehensive Evaluation of Multi-Reference-to-Audio-Video Generation

Keywords: Multi-reference-to-audio-video (MR2AV), MultiRef-Compass, Audio-Visual Consistency, Reference Consistency, Generative Models

Category: Generative Models

Research Objective:

– The objective is to explore and establish a benchmark for Multi-reference-to-audio-video (MR2AV) generation that synthesizes coherent audio-video content from multiple references and textual instructions.

Research Methods:

– The study introduces MultiRef-Compass, a comprehensive benchmark containing 350 curated samples for MR2AV generation, and defines an evaluation protocol with four dimensions using 14 sub-metrics.

Research Conclusions:

– Extensive experiments on eight representative MR2AV systems reveal significant areas for improvement, positioning MultiRef-Compass as a foundational tool for future MR2AV research.

Paper link: https://huggingface.co/papers/2607.14189

5. From Pixels to States: Rethinking Interactive World Models as Game Engines

Keywords: Interactive game worlds, Video generative models, Real-time generation, Game state dynamics, Scalable data engine

Category: Generative Models

Research Objective:

– To examine interactive game world modeling focusing on player action control, game state dynamics, state-observation persistence, and real-time interactive generation.

Research Methods:

– Organizing existing approaches into representative families and analyzing their strengths and trade-offs.

– Developing a scalable data engine for Black Myth: Wukong collecting comprehensive gameplay data with annotations.

Research Conclusions:

– The paper provides a clear perspective on current progress and challenges, offering insights that could drive future advancements toward truly interactive game worlds.

Paper link: https://huggingface.co/papers/2607.14076

6. KeyFrame-Compass: Towards Comprehensive Evaluation of Keyframe-Conditioned Video Generation

Keywords: KeyFrame-Compass, Keyframe-conditioned video generation, Automated evaluation, Video quality

Category: Generative Models

Research Objective:

– To introduce KeyFrame-Compass, a comprehensive benchmark designed to evaluate keyframe-conditioned video generation.

Research Methods:

– Development of a benchmark with 386 samples across diverse settings and an automated evaluation framework using six metrics and MLLM judgments.

Research Conclusions:

– Current video generation models show a trade-off between executing keyframes faithfully and maintaining video quality, with performance declining as keyframe constraints increase.

Paper link: https://huggingface.co/papers/2607.14202

7. LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget

Keywords: AI agents, LongStraw, Group Relative Policy Optimization, Qwen3.6-27B, GLM-5.2

Category: Reinforcement Learning

Research Objective:

– Address the gap in inference context lengths between million-token contexts and shorter RL post-training workloads using LongStraw.

Research Methods:

– Implementation of the LongStraw execution stack with Group Relative Policy Optimization for RL post-training.

– Use of hybrid recurrent and full-attention for Qwen3.6-27B and mixture-of-experts GLM-5.2 neural architectures.

Research Conclusions:

– Successfully demonstrated an execution capacity on 8 and 32 H20 GPUs, supporting grouped scoring and response backward for extensive token contexts.

– Highlighted the increased scalability with minimal memory costs, confirming the potential of LongStraw for improving long trajectory processing in AI agents.

Paper link: https://huggingface.co/papers/2607.14952

8. SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

Keywords: Large language models, Outcome-based reinforcement learning, SEED, Hindsight skills, Sample efficiency

Category: Reinforcement Learning

Research Objective:

– To address the supervision gap in outcome-based reinforcement learning by proposing SEED, a framework that enhances policy learning with hindsight skills.

Research Methods:

– SEED leverages self-evolving on-policy distillation by analyzing completed trajectories and extracting reusable natural-language skills during reinforcement learning.

Research Conclusions:

– SEED improves performance and sample efficiency in text-based and vision-based tasks, demonstrating robust generalization to new scenarios.

Paper link: https://huggingface.co/papers/2607.14777

The post AI Native Daily Paper Digest – 20260717 – Qwen | Claude | DeepSeek-V3 appeared first on AI Native Foundation.

1. Moonshot AI Introduces Kimi K3: The First Open 3T-Class Model Advancing Frontier Intelligence

Moonshot AI has launched Kimi K3, its most capable model to date and the world’s first open 3T-class AI model. Built with 2.8 trillion parameters, Kimi K3 introduces architectural innovations including Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), alongside native vision capabilities and a 1-million-token context window. Designed for frontier-level coding, knowledge work, and complex reasoning, Kimi K3 achieves improved scaling efficiency through expanded Mixture of Experts (MoE) architecture, refined training methods, and optimized data strategies, delivering approximately 2.5× better scaling efficiency than Kimi K2. The model demonstrates strong performance across comprehensive evaluations and is now available through

Read more: https://www.kimi.com/blog/kimi-k3

Video Credit: The original article

2. Tencent Launches SkillPay: Enabling Agent Skill Monetization Through Integrated Distribution, Usage, and Payments

Tencent has launched SkillPay, an Agent skill monetization platform within SkillHub that connects skill distribution, AI agent execution, and payment infrastructure into a unified ecosystem. Businesses can easily transform existing capabilities into Pay Skills, define pricing models, and monetize their services through Agent interactions without building separate payment systems. Users can access professional services such as financial analysis, AI-generated content creation, and travel booking directly within Agent conversations, with seamless payment powered by WeChat Pay AI accounts. By combining trusted skill verification, secure payments, and open commercialization tools, SkillPay aims to accelerate the Agent economy and enable more developers and businesses to deliver paid AI-powered services.

Read more: https://mp.weixin.qq.com/s/TtxqzHPvZX7Ff6GGHgHp-g

Video Credit: NotebookLM

3. Alibaba Cloud Releases Qwen-Audio-3.0-Realtime: Advancing Voice AI from Conversation to Action

Alibaba Cloud has introduced Qwen-Audio-3.0-Realtime, a real-time speech interaction model designed to move voice AI beyond simple conversations toward intelligent task execution. The model delivers major upgrades in reasoning, Agent tool calling, empathetic interaction, and full-duplex communication, with two variants: Plus for stronger reasoning and Flash for faster responses. Qwen-Audio-3.0-Realtime supports native tool usage through FunctionCall, MCP, APIs, and knowledge bases, allowing voice agents to autonomously access external services while maintaining conversation context. It also enables adaptive emotional expression, multi-speaker understanding, noise-resistant interactions, and natural listen-and-speak capabilities. Built with an online policy distillation framework, the model combines language reasoning, agent capabilities, and audio intelligence to create more capable real-time voice agents.

Read more: https://mp.weixin.qq.com/s/W1Lp3trsMXfNTmetkBtCQg

Video Credit: The original article

4. MiniMax Code 2.0 Released: A Complete Agent Architecture Upgrade for More Reliable AI Workflows

MiniMax has launched MiniMax Code 2.0 desktop, featuring a major underlying architecture redesign powered by the open-source Pi Agent framework. The update improves long-running task execution, session management, context continuity, and tool-calling reliability, enabling AI agents to handle complex workflows with greater stability and efficiency. The new version also enhances productivity features such as chart visualization, full-screen viewing, file preview, and in-place editing, creating a smoother path from task execution to final delivery. In addition, MiniMax Code 2.0 introduces financial intelligence capabilities through integrations with financial databases and enterprise data services, enabling agents to support research, investment analysis, and risk management workflows. Upcoming features including remote control, browser automation, goal mode, and planning mode will further expand its ability to handle complex tasks across environments.

Read more: https://mp.weixin.qq.com/s/mQeBO0xC6Z1R0LqZX5TvNg

Video Credit: The original article

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post China AI Native Industry Insights – 20260717 – Kimi | Tencent | Alibaba | more appeared first on AI Native Foundation.

1. Harness Handbook: Making Evolving Agent Harnesses Readable,Navigable, and Editable

Keywords: AI Agent, Harness Handbook, Behavior Localization, LLM-assisted Structuring, Behavior-Guided Progressive Disclosure

Category: AI Systems and Tools

Research Objective:

– The study aims to improve the modification process of AI agent harnesses by introducing a novel behavior-centric representation called the Harness Handbook, and propose a Behavior-Guided Progressive Disclosure method for efficient behavior localization.

Research Methods:

– The research employs static analysis and LLM-assisted structuring to automatically synthesize the Harness Handbook from a codebase, linking behaviors to sources. It also uses Behavior-Guided Progressive Disclosure to guide agents from high-level behaviors to detailed implementation, verifying candidate locations.

Research Conclusions:

– The Handbook-Assisted planning enhances behavior localization and edit-plan quality while reducing planner token usage, especially in complex cases involving scattered sites, rarely executed paths, and cross-module interactions.

Paper link: https://huggingface.co/papers/2607.13285

2. Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning

Keywords: Zero RL, 1T parameters, emergent capabilities, chain-of-thought reasoning, structured evaluation

Category: Reinforcement Learning

Research Objective:

– Explore the large-scale dynamics and emergent capabilities of zero Reinforcement Learning models, particularly concerning chain-of-thought reasoning.

Research Methods:

– Developed a stable and efficient training pipeline with optimizations like clipped importance sampling and mixed-precision control to deal with large-scale models.

Research Conclusions:

– Scaling to 1T parameters enhances sample efficiency and performance, and enables the model to spontaneously develop advanced cognitive behaviors, eliminating the need for hand-crafted heuristics.

– A new structured evaluation framework is proposed to assess comprehensibility, reproducibility, and efficiency of chain-of-thought reasoning beyond final-answer correctness.

Paper link: https://huggingface.co/papers/2607.12395

3. OvisOCR2 Technical Report

Keywords: OvisOCR2, document parsing, end-to-end, Markdown representation, synthetic pages

Category: Computer Vision

Research Objective:

– The paper introduces OvisOCR2, a 0.8B parameter model, aimed at parsing document page images into Markdown formatted representations, capturing various elements like text, formulas, tables, and visual regions.

Research Methods:

– OvisOCR2 employs a data engine combining real-document annotations with synthetic page data derived from HTML sources. It uses supervised fine-tuning, reinforcement learning with a multi-component reward design, on-policy distillation, and model fusion for training.

Research Conclusions:

– OvisOCR2 achieves state-of-the-art performance on OmniDocBench v1.6 and PureDocBench, demonstrating its superiority over pipeline methods and its robustness and generalization across diverse and challenging document parsing scenarios.

Paper link: https://huggingface.co/papers/2607.13639

4. MetaView: Monocular Novel View Synthesis with Scale-Aware Implicit Geometry Priors

Keywords: Visual Generation Models, Implicit Geometry, Monocular Novel View Synthesis, Spatial Structure, Diffusion-Based

Category: Generative Models

Research Objective:

– The paper introduces MetaView, a novel framework for monocular novel view synthesis designed to maintain geometry consistency and precise controllability while enabling large view changes from a single image.

Research Methods:

– The approach combines implicit geometry modeling with essential explicit 3D cues using a feed-forward geometry perception network, aiming to balance flexibility with structural consistency.

Research Conclusions:

– MetaView demonstrates superior performance compared to existing methods in handling challenging monocular large viewpoint changes, offering significant improvements in generalization capabilities.

Paper link: https://huggingface.co/papers/2607.12000

5. Registers Matter for Pixel-Space Diffusion Transformers

Keywords: Vision Transformers, Diffusion Transformers, Register Tokens, Pixel-space Training, Feature Maps

Category: Generative Models

Research Objective:

– To investigate the role and effectiveness of register tokens in Diffusion Transformers (DiTs) compared to Vision Transformers (ViTs).

Research Methods:

– Analysis of intermediate representations to compare feature map quality in both pixel-space and latent-space DiTs.

Research Conclusions:

– DiTs do not exhibit high-norm patch-token outliers like ViTs, but still benefit from register tokens, especially in pixel-space applications.

– The use of register tokens leads to cleaner feature maps at high noise levels, contributing to improved visual structure and coherence.

– Recent pixel-space DiTs architectures include mechanisms similar to register tokens, which enhances their performance.

Paper link: https://huggingface.co/papers/2605.16147

6. Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation

Keywords: 3D generation, 4D generation, large multimodal language models (LMMs), spatial and temporal inconsistencies

Category: Generative Models

Research Objective:

– The paper presents Hallo4D, aiming to mitigate spatiotemporal hallucinations in 3D and 4D content generation by ensuring geometric consistency.

Research Methods:

– The authors introduce a generation-detection-correction paradigm leveraging large multimodal language models, multi-model voting, and motion-aware keyframe sampling.

Research Conclusions:

– Hallo4D outperforms strong baselines and offers a scalable, generalizable solution for consistency-aware 3D and 4D content generation across diverse settings.

Paper link: https://huggingface.co/papers/2607.12752

7. AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities

Keywords: Large Language Models, AgentCompass, Evaluation Infrastructure, Autonomous Agents, Reproducibility

Category: AI Systems and Tools

Research Objective:

– To introduce AgentCompass, an open-source infrastructure aimed at unifying and enhancing the evaluation of LLM-based autonomous agents.

Research Methods:

– Organization around Benchmark, Harness, and Environment components for flexible configurations and fault-tolerant asynchronous runtime with trajectory analysis tools.

Research Conclusions:

– Provides scalable, reproducible infrastructure supporting over 20 benchmarks, aiding in the advancement of agent research by diagnosing failure modes.

Paper link: https://huggingface.co/papers/2607.13705

8. Tracing Agentic Failure from the Flow of Success

Keywords: Failure Attribution, Agentic Systems, Lightweight Model, One-Class Learning, Neural Controlled Differential Equations

Category: AI Systems and Tools

Research Objective:

– The paper aims to develop a practical failure attribution model for LLM-based agentic systems that is lightweight and does not require step-level supervision on failure data.

Research Methods:

– The authors propose OAT, a model employing one-class learning with neural controlled differential equations to analyze successful trajectories and identify failure steps during inference.

Research Conclusions:

– OAT demonstrates a significant improvement in efficiency, being 200-5000 times faster than prompting-based baselines, while delivering better performance with an increase of +20% and +7% in F1 scores on in-domain and out-of-distribution datasets, respectively.

Paper link: https://huggingface.co/papers/2607.12747

9. PalmClaw: A Native On-Device Agent Framework for Mobile Phones

Keywords: Large Language Model (LLM), Mobile Devices, PalmClaw, AI Native

Category: AI Systems and Tools

Research Objective:

– This paper presents PalmClaw, an open-source agent framework designed to operate natively on mobile devices, allowing AI Native support of executing multi-step tasks by utilizing mobile-specific capabilities directly.

Research Methods:

– PalmClaw exposes device capabilities through explicit arguments and structured results with clearly defined execution boundaries, facilitating direct interaction between mobile agents and device functionalities.

Research Conclusions:

– The implementation of PalmClaw showed an 11.5% improvement in task success rate and a 94.9% reduction in task completion time compared to existing baselines, demonstrating its effectiveness and efficiency in mobile AI task execution.

Paper link: https://huggingface.co/papers/2607.13027

10. From Controlled to the Wild: Evaluation of Pentesting Agents for the Real-World

Keywords: AI pentesting agents, vulnerability discovery, evaluation protocol, strategic decision-making, reproducibility

Category: AI Systems and Tools

Research Objective:

– To present a practical evaluation protocol for AI pentesting agents that focuses on validated vulnerability discovery across complex targets and multiple attack surfaces.

Research Methods:

– The protocol includes structured ground-truth with LLM-based semantic matching, bipartite resolution, continuous ground-truth maintenance, and stochastic agent evaluation, along with efficiency metrics for sustainable experimentation.

Research Conclusions:

– This protocol extends current evaluation methods by providing a more realistic and informative comparison of AI pentesting agents, enabling operational insights and reproducibility through released expert-annotated ground truth and code.

Paper link: https://huggingface.co/papers/2605.10834

11. Generative Compilation: On-the-Fly Compiler Feedback as AI Generates Code

Keywords: Rust, generative compilation, compiler feedback, partial-program checker, AI-assisted programming

Category: AI Systems and Tools

Research Objective:

– To introduce generative compilation, an approach for obtaining compiler feedback on partial programs during generation, enhancing AI-generated code’s correctness and reducing non-compiling outputs.

Research Methods:

– Developed a sealor that transforms partial programs into complete ones to enable standard compiler diagnosis.

– Constructed and mechanized the sealor in Lean for a Rust-like calculus, and extended it to a partial-program checker for real Rust.

Research Conclusions:

– Generative compilation reduces non-compiling outputs and enhances functional correctness by detecting errors early in the generation process, which minimizes error cascades and facilitates precise diagnostics.

– It repositions compilers as active participants in AI-assisted programming, moving beyond a post-generation check to a proactive error-reducing tool.

Paper link: https://huggingface.co/papers/2607.13921

12. Length Penalties Make Chain-of-Thought Less Monitorable

Keywords: Length-penalized reinforcement learning, Chain-of-thought reasoning, Compression, Biasing-hint interventions, Qwen3-4B and Qwen3-14B

Category: Reinforcement Learning

Research Objective:

– To explore how length-penalized reinforcement learning affects the chain-of-thought reasoning process and the influence of misleading hints.

Research Methods:

– Training Qwen3-4B and Qwen3-14B variants with various target chain lengths and evaluating with biasing-hint interventions on held-out MMLU-Pro-R and four transfer benchmarks.

Research Conclusions:

– Compression reduces reasoning tokens and maintains multiple-choice accuracy while making the underlying influences less detectable. Despite shorter reasoning, the models continue to be driven by misleading hints.

Paper link: https://huggingface.co/papers/2607.09786

13.

Paper link: 

14. AffectFlow-DINO: Uncertainty-Aware Multi-Task Affect Estimation via Conditional Rectified Flow

Keywords: AffectFlow-DINO, multi-task learning, uncertainty-aware, facial behavior, Monte Carlo sampling

Category: Computer Vision

Research Objective:

– To develop AffectFlow-DINO, a system capable of modeling the ambiguity in facial behavior using a conditional generative distribution.

Research Methods:

– Utilization of a multi-task learning approach extending a deterministic architecture with a conditional rectified-flow head; application of Monte Carlo sampling for uncertainty-aware predictions.

– Built on frozen DINOv3 ViT-S/16 architecture and employs joint estimation techniques for valence-arousal, facial expression classification, and Action Units detection.

Research Conclusions:

– The introduction of rectified-flow decoding enhances deterministic predictions, notably improving CCC for valence-arousal estimation.

– Effective performance recovery in rare classes through post-hoc threshold calibration without the need for retraining; combined methods substantially outperform baseline models in multi-task learning performance metrics.

Paper link: https://huggingface.co/papers/2607.13250

15. SPEAR: A Simulator for Photorealistic Embodied AI Research

Keywords: AI Native, Photorealistic Simulators, Unreal Engine, Embodied AI, Python Library

Category: AI Systems and Tools

Research Objective:

– The research aims to overcome limitations in existing photorealistic simulators regarding generality, programmability, and rendering speed by introducing SPEAR, a Simulator for Photorealistic Embodied AI Research.

Research Methods:

– SPEAR is developed as a Python library connecting to any Unreal Engine application through a modular plugin architecture, exposing over 14K unique UE functions to Python and significantly enhancing programmable functionality.

– It achieves a rendering speed of 73 frames per second at 1920×1080 resolution while providing unique image modalities and an expressive high-level programming model for complex task execution.

Research Conclusions:

– SPEAR demonstrates its utility through multiple applications, such as controlling diverse embodied agents, rendering city-scale environments, and coordinating simulations, effectively showcasing advanced programmability and rendering speed.

Paper link: https://huggingface.co/papers/2607.06701

16. Self in Space: Benchmarking Self-Awareness and Spatial Cognition in UAV Embodied Intelligence

Keywords: MLLMs, UAV systems, SIS-Bench, self-awareness, spatial intelligence

Category: Robotics and Autonomous Systems

Research Objective:

– The study aims to address the imbalance in UAV systems between spatial cognition and self-awareness by introducing SIS-Bench, a benchmark for evaluating embodied spatial intelligence in UAV scenarios.

Research Methods:

– The researchers developed SIS-Bench, organizing evaluation along two dimensions, space and self, with a hierarchy of perception, memory, and reasoning. It consists of 4,856 question–answer pairs across 13 tasks from 1,646 UAV videos, validated by experts.

Research Conclusions:

– The study found that current MLLMs have significant limitations in modeling dynamic and agent-centered processes. Incorporating motion-aware representation through optical flow and visual feature fusion improves perception, memory, and enhances self-awareness, demonstrating its applicability to downstream UAV decision-making tasks.

Paper link: https://huggingface.co/papers/2607.12477

17. From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization

Keywords: LLM, Agent Optimization, Causal Extraction, VeruSAGE-Bench, STRACE

Category: Reinforcement Learning

Research Objective:

– To improve the optimization of long-horizon agents through a framework called STRACE, which constructs high signal-noise optimization contexts.

Research Methods:

– Utilizes Structural Trajectory Analysis to mine failure patterns and perform causal localization over a textual dependency graph to filter redundant traces and identify root causes.

Research Conclusions:

– STRACE outperforms standard context-filtering baselines, achieving a 1.4 times improvement in success rate on a formal verification task involving human-expert designed agents.

Paper link: https://huggingface.co/papers/2607.07702

18. Discrete Diffusion Models: A Unified Framework from Tokenization to Generation

Keywords: Discrete Denoising Diffusion Models, Autoregressive Modeling, Parallel Generation, Iterative Global Refinement

Category: Generative Models

Research Objective:

– Introduce a unified conceptual framework for understanding discrete denoising diffusion models (DDMs) through the construction of discrete state spaces.

Research Methods:

– Analyze DDMs using various approaches like transition-matrix, masking/absorbing-state, and score/ratio-based methods, showing them as different instantiations within a common design space.

Research Conclusions:

– Highlight common design trade-offs across DDMs, including training objectives and inference algorithms, proposing several directions for future research.

Paper link: https://huggingface.co/papers/2607.13431

19. Vinci2: Providing Proactive Assistance in Continuous Egocentric Videos

Keywords: Proactive Assistance, Egocentric Video, Contextual Decision, Vinci2, EgoMemo

Category: Human-AI Interaction

Research Objective:

– The study aims to develop a proactive egocentric assistance system by enhancing the Vinci assistant from reactive to proactive, focusing on context-dependent decision-making in continuous egocentric video.

Research Methods:

– Introduces Vinci2 and EgoServe, where Vinci2 is an advanced proactive assistance system, and EgoServe serves as a large-scale benchmark for proactive assistance. It explores the use of EgoMemo, a memory-augmented agent, implementing multi-scale temporal summaries, a semantic knowledge graph, and visual embedding archives.

Research Conclusions:

– The research demonstrates that EgoMemo can effectively establish strong baselines in the EgoServe benchmark and perform competitively on existing egocentric benchmarks, contributing to the advancement of proactive assistance systems.

Paper link: https://huggingface.co/papers/2607.11523

20. ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

Keywords: Structured Pruning, On-Policy Distillation, Compression, Generative Models, Natural Language Processing

Category: Natural Language Processing

Research Objective:

– To improve the quality of generative tasks in large language models (LLMs) after structured pruning, by addressing recovery issues using a novel method called short-to-long OPD.

Research Methods:

– Implementing On-Policy Distillation (OPD) using a pre-compression model as a frozen teacher and employing a short-to-long schedule to optimize token-level supervision in rollouts.

Research Conclusions:

– The short-to-long OPD method significantly enhances compressed model performance across various tasks, achieving up to 9 times its original score, using substantially less training time and resources.

Paper link: https://huggingface.co/papers/2607.13124

21. Self-Improvements in Modern Agentic Systems: A Survey

Keywords: Self-improving agents, Controllable evolution, Adaptive systems, Model parameters, Operational scaffold

Category: Robotics and Autonomous Systems

Research Objective:

– To explore the framework and systems of self-improving autonomous agents that adapt from experience with minimal human input.

Research Methods:

– The study presents a system-level framework where modern agents are viewed as configurations of foundation models coupled with operational scaffolds, formalizing self-improvement through a self-induced update operator.

Research Conclusions:

– The survey organizes prior work based on update targets and the signals driving change, reviews applications, and discusses evaluation, ultimately suggesting open problems and future research directions.

Paper link: https://huggingface.co/papers/2607.13104

22. GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch

Keywords: World Action Models, GigaWorld-Policy-0.5, action-centered formulation, Mixture-of-Transformers, AutoResearch pipeline

Category: Robotics and Autonomous Systems

Research Objective:

– The study aims to enhance robot policy learning by addressing the computational inefficiencies in World Action Models, focusing on efficient robot control and inference.

Research Methods:

– The researchers employ an action-centered formulation, using Action-Conditioned World Modeling for pretraining and introduce a Mixture-of-Transformers architecture to optimize inference efficiency. They also utilize an agent-based AutoResearch pipeline for optimal training configuration search.

Research Conclusions:

– GigaWorld-Policy-0.5 successfully retains the benefits of future visual dynamics in training while substantially improving efficiency in inference, achieving low latency and reducing the need for manual tuning in hyperparameter settings.

Paper link: https://huggingface.co/papers/2607.13960

23. PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails

Keywords: PolicyShiftBench, PolicyShiftGuard, policy adaptation, AI Native, image guardrails

Category: Computer Vision

Research Objective:

– To explore policy-adaptive image guardrailing, allowing models to determine if an image violates the currently supplied policy and to generalize to new policy definitions.

Research Methods:

– Introduction of PolicyShiftBench, a benchmark with policy-discriminative instances to test model adaptability to active policies.

– Development of PolicyShiftGuard, a compact guardrail using a two-stage training process combining Randomized Policy SFT with Boundary-Pair Policy Adaptation.

Research Conclusions:

– PolicyShiftGuard significantly improves policy-sensitive performance over existing models, achieving state-of-the-art results on PolicyShiftBench, and transfers effectively to other benchmarks. Matched pass/block boundary pairs are critical for stable policy adaptation.

Paper link: https://huggingface.co/papers/2607.05910

24. KnowAct-GUIClaw: Know Deeply, Act Perfectly, Personal GUI Assistant with Self-Evolving Memory and Skill

Keywords: OpenClaw, KnowAct-GUIClaw, cross-platform adaptability, execution accuracy, AI Systems and Tools

Category: AI Systems and Tools

Research Objective:

– To address the limitations of OpenClaw in cross-platform GUI interaction and self-evolution, enhancing its adaptability and performance.

Research Methods:

– Introduction of KnowAct-GUIClaw, a framework that employs a Know-Route-Act-Reflect approach to leverage user interactions and experience memory for improved task automation.

Research Conclusions:

– KnowAct-GUIClaw demonstrated superior efficiency, accuracy, and cross-platform adaptability, particularly excelling in the MobileWorld benchmark with notable performance improvements over existing frameworks.

Paper link: https://huggingface.co/papers/2607.12625

25. Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

Keywords: Boogu-Image-0.1, multimodal understanding, text-to-image generation, open-source, bilingual text rendering

Category: Multi-Modal Learning

Research Objective:

– Introduce Boogu-Image-0.1, an open-source multimodal model family offering capabilities like text-to-image generation and bilingual text rendering.

Research Methods:

– Focused on enhancing model understanding, data quality, and training pipelines with agentic inference-time scaling.

Research Conclusions:

– Boogu-Image-0.1 matches or surpasses other open-source models and competes closely with closed-source systems, achieving this with a relatively low theoretical training cost.

Paper link: https://huggingface.co/papers/2607.13125

The post AI Native Daily Paper Digest – 20260716 – GPT-4.5 | Long-Context Attention | Video Foundation Models appeared first on AI Native Foundation.

1. Anthropic Launches Claude for Teachers with Free Premium Access for US K-12 Educators

Anthropic has introduced Claude for Teachers, a program offering free access to premium Claude capabilities for verified K-12 educators in the United States. The offering includes a library of teaching skills and direct integration with evidence-based curricula mapped to academic standards across all 50 states. The initiative is designed to support classroom instruction by connecting teachers to standards-aligned educational resources. By targeting verified educators specifically, Anthropic aims to make advanced AI tools more accessible within the US public and private K-12 education system.

Read more: https://t.co/5hZZijVPCV

Video Credit: The original article

2. xAI Open-Sources Grok Build CLI and Resets Usage Limits for All Users

xAI has open-sourced Grok Build, its terminal-native AI coding agent and CLI first launched in beta in May 2026. The full source code, including the Git repository for the Grok Build CLI, is now publicly available on GitHub, enabling developers to inspect, customize, and extend the tool for their own workflows. xAI simultaneously reset usage limits for all users. The open-source release allows the community to contribute to a more reliable and robust development harness. Grok Build is designed for professional software engineers and supports complex coding tasks including code editing, refactoring, and execution across any codebase or language.

Read more: https://x.ai/open-source

Video Credit: NotebookLM

3. OpenAI Introduces GPT-Red, an Automated Red-Teaming Model for Prompt Injection Testing

OpenAI introduced GPT-Red, an internal automated red-teaming model designed to find prompt injection vulnerabilities in its AI models at scale before wider deployment. GPT-Red is trained via adversarial self-play reinforcement learning, generating progressively stronger prompt injection attacks while defender models learn to resist them. Discovered vulnerabilities are fed back into model training, creating an iterative improvement loop. OpenAI reported that GPT-Red was used in training GPT-5.6, where it succeeded in 84% of internal evaluation scenarios compared to 13% for human red teamers. The system enables security testing at a scale not achievable by human researchers alone, and expands on OpenAI’s earlier Red Teaming Network launched in 2023.

Read more: https://openai.com/index/unlocking-self-improvement-gpt-red/

Video Credit: NotebookLM

4. Google Rolls Out Community-Driven Improvements and Fixes to Gemma 4

Google’s official Gemma team announced a new round of improvements and fixes to Gemma 4, driven by community feedback and contributions. The update was shared via a detailed thread on X, outlining specific changes being made in the release. Gemma 4 is Google’s open model family available in multiple sizes, including E2B, E4B, 26B, and 31B variants, with multimodal and agentic capabilities. The community-driven nature of the update highlights Google’s ongoing engagement with developers and open-source contributors to iterate on the Gemma 4 model family.

Read more: https://claude.com/solutions/teachers

Video Credit: NotebookLM

That’s all for today’s Global AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

The post Global AI Native Industry Insights – 20260716 – Anthropic | xAI | OpenAI | more appeared first on AI Native Foundation.