Sources

How agents are transforming work openai.com

A new OpenAI research paper shows how AI agents are transforming work, enabling longer, more complex tasks and expanding productivity across roles.

MaineCoon: Pursuing A Real-Time Audio-Visual Social World Model huggingface.co

MaineCoon represents the first real-time audio-visual autoregressive model for social worlds, achieving high frame rates and long-horizon generation through novel training techniques and inference frameworks.

Kairos: A Native World Model Stack for Physical AI huggingface.co

Kairos is a world model framework that learns from diverse experiences, maintains persistent states through hybrid temporal attention mechanisms, and operates efficiently across different hardware platforms for physical AI applications.

Sumi: Open Uniform Diffusion Language Model from Scratch huggingface.co

Tohoku NLP’s Sumi shows that uniform diffusion language models, trained at scale from scratch, match autoregressive baselines on knowledge and reasoning benchmarks. The team releases weights, data mixture and training recipe, while flagging a persistent gap on commonsense reasoning versus standard left-to-right models.

MyPCBench: A Benchmark for Personally Intelligent Computer-Use Agents huggingface.co

Two new persona-seeded benchmarks probe whether computer-use agents can act as genuine personal assistants across desktop and mobile. On the Linux-based MyPCBench, Claude Opus 4.6 leads with a 55.4% task-completion rate but stumbles on multi-app workflows and long trajectories; iOSWorld adds persistent user identity across native iOS apps.

iOSWorld: A Benchmark for Personally Intelligent Phone Agents huggingface.co

IOSWorld is introduced as the first interactive native iOS simulator benchmark featuring persistent user identity across multiple apps to evaluate personalized mobile agent capabilities.

SAE Interventions are Unreliable: Post-Intervention Recovery of Suppressed Behavior huggingface.co

Sparse autoencoder interventions that look like they suppress unlearning, IOI or refusal behaviors can be reversed by optimizing in the residual stream with encoder-orthogonal updates. The result undercuts SAE features as a complete control surface, since suppressed behavior re-emerges without touching the edited latents.

CEO-Bench: Can Agents Play the Long Game? huggingface.co

Princeton’s CEO-Bench drops language-model agents into a Python-driven startup simulation spanning 500 days, with business databases, customer cohorts and negotiation history. The benchmark scores long-horizon planning, noise handling and multi-task coordination, targeting the adaptive, sustained progress that current agent evals rarely measure.

Bag of Dims: Training-Free Mechanistic Interpretability via Dimension-Level Sign Patterns huggingface.co

Skipping learned dictionaries entirely, Bag of Dims reads the standard basis of transformer hidden states as semantic features: signs encode content, magnitudes encode confidence. The training-free framework supports next-token prediction via Hamming scoring and transfers across language, vision and audio transformers without optimization.

When Does Trajectory-Level Supervision Permit Efficient Offline Reinforcement Learning? huggingface.co

A new analysis pins down when offline RL with only trajectory-level outcome rewards is statistically tractable. The proposed OPAC algorithm pairs a latent reward model with pessimistic actor-critic updates and achieves high-probability guarantees, while generalized outcome-based settings hit a provable lower bound on sample complexity.

STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability huggingface.co

GRPO-trained LLMs often see policy entropy collapse mid-training, killing exploration. STARE diagnoses the issue with a first-order gradient analysis, then reweights token-level advantages by surprisal and adds a closed-loop target-entropy gate, stabilizing reasoning RL on benchmarks including AIME24 and AIME25.

Physics-IQ Verified huggingface.co

A systematic evaluation of the Physics-IQ benchmark reveals limitations in measuring physical understanding of video generative models, leading to improvements in prompt quality and sample-level scoring that enhance reliability for assessing physically accurate video generation.

Native Active Perception as Reasoning for Omni-Modal Understanding huggingface.co

OmniAgent is a novel omni-modal agent that addresses long video understanding by using an iterative observation-thought-action cycle with active perception, achieving superior performance compared to larger models through efficient selective processing.

REVES: REvision and VErification—Augmented Training for Test-Time Scaling huggingface.co

A two-stage iterative framework alternates between data augmentation and policy optimization to improve LLM reasoning by leveraging intermediate correction steps, achieving superior performance on coding benchmarks and constraint satisfaction problems.

RODS: Reward-Driven Online Data Synthesis for Multi-Turn Tool-Use Agents huggingface.co

RODS addresses sample depletion in multi-turn tool-use reinforcement learning by dynamically synthesizing new data based on reward variance to maintain informative training samples.

EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts huggingface.co

EfficientRollout is a system-aware self-speculative decoding framework that accelerates reinforcement learning rollouts by adapting drafters to evolving policies and optimizing speculative decoding regimes.

The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL huggingface.co

Discriminator-Guided Reinforcement Learning (DRL) addresses alignment issues in score- and flow-matching models by using a pretrained representation space discriminator as an optimal reward signal, improving both visual fidelity and semantic quality without human preferences.

Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillation huggingface.co

ViGOS is a visually grounded on-policy self-distillation framework for multimodal large language models that improves image-grounded behavior by using specialized teachers for different stages of reasoning and handling invalid rollouts.

From Trainee to Trainer: LLM-Designed Training Environment for RL with Multi-Agent Reasoning huggingface.co

A framework automates environment redesign in reinforcement learning for large language models by having the policy analyze failures and suggest configuration changes, achieving superior performance over larger proprietary models and fixed-environment baselines.

PAIWorld: A 3D-Consistent World Foundation Model for Robotic Manipulation huggingface.co

PAIWorld enhances diffusion-transformer world models with geometric awareness and cross-view attention to improve multi-view 3D consistency for robotic manipulation tasks.

Think Again or Think Longer? Selective Verification for Budget-Aware Reasoning huggingface.co

Selective verification approaches optimize test-time reasoning by dynamically deciding when to verify answers, achieving better accuracy and efficiency compared to always-verifying or self-consistency methods.

Reinforcing Dual-Path Reasoning in Spatial Vision Language Models huggingface.co

A unified framework for spatial vision-language models that combines linguistic deduction and 3D geometric reasoning through reinforcement learning, enabling robust spatial reasoning across diverse tasks and domains.

Trust the Right Teacher: Quality-Aware Self-Distillation for GUI Grounding huggingface.co

Quality-aware self-distillation improves vision-language model performance for GUI grounding by enhancing coordinate-token teacher signals through correctness-aware gating and probability scaling.

Guava: An Effective and Universal Harness for Embodied Manipulation huggingface.co

A harness framework for embodied tool use combines high-level reasoning with external modules, enabling compact models to perform complex manipulation tasks with minimal training data.

Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games huggingface.co

A new benchmark suite called RNG-Bench is introduced to evaluate multimodal foundation models’ ability to reconstruct past observations and use them for decision-making in multi-step interactions, featuring two games with controlled difficulty parameters and a memory gap metric to distinguish forgetting from poor decision-making.

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks huggingface.co

SciOrch is a framework that uses a lightweight orchestrator model to coordinate multiple frontier LLMs for scientific reasoning, achieving superior performance through MCTS-based training and GRPO-style optimization while reducing API costs.

Learning User Simulators with Turing Rewards huggingface.co

A reinforcement learning approach using Turing test-based rewards trains language models to generate responses indistinguishable from human users in conversational and forum discussion settings.

Externalizing Research Synthesis and Validation in AI Scientists through a Research Harness huggingface.co

Xcientist enables transparent and accountable AI-driven scientific research by creating persistent artifacts that track the complete research process from problem formulation to mechanism validation and revision.

Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems huggingface.co

Multicultural multi-agent systems exhibit limited value diversity despite cultural alignment, with social interaction reducing diversity and compromising collective decision-making breadth.

Re-Centering Humans in LLM Personalization huggingface.co

Human-centered evaluation reveals significant gaps between synthetic and real-world LLM personalization performance, with models struggling to extract user attributes and generate truly personalized responses that match human quality judgments.

HiLo-Token: Input-Adaptive High-Low Frequency Token Compression for Efficient Image Editing huggingface.co

A novel token compression framework called HiLo-Token is introduced to accelerate Diffusion Transformers in image editing tasks by adaptively allocating tokens based on spatial frequency and context importance, achieving significant speedups without quality loss.

A Benchmark and Framework for Evaluating Next Action Predictions in Spreadsheets huggingface.co

A benchmark for predicting spreadsheet user actions is introduced, addressing challenges in edit history availability and complex action spaces through manual curation and online evaluation methodology.

ViT-Up: Faithful Feature Upsampling for Vision Transformers huggingface.co

ViT-Up is a feature upsampling framework for Vision Transformers that uses layer-wise query construction from hidden states to improve dense prediction tasks, outperforming existing image-guided methods.

Morpheus: A Morphology-Aware Neural Tokenizer and Word Embedder for Turkish huggingface.co

A neural morpheme-boundary model for Turkish achieves lossless tokenization and morphology-aware embeddings with improved efficiency and performance over traditional subword methods.

IndustryBench-MIPU: Benchmarking Multi-Image Attribute Value Extraction for Industrial Products huggingface.co

IndustryBench-MIPU is introduced as the first large-scale benchmark for multi-image industrial product understanding, focusing on structured attribute extraction from heterogeneous product images to evaluate multimodal models’ ability to recover dense technical specifications.

LLM-Enabled NWDAF: A Step Toward AI-Native 6G Network Intelligence huggingface.co

An open-source Network Data Analytics Function compatible with Free5GC integrates a Large Language Model interface for natural language interaction and intent-based network management.

MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction huggingface.co

3D point motion forecasting model predicts object trajectories from visual history and language goals, demonstrating superior performance on benchmarks and transferring effectively to robot manipulation and video generation tasks.

Europe is pushing back on Washington’s chip war techcrunch.com

As ASML CEO Christophe Fouquet told TechCrunch in May, what China can currently buy are older-generation deep ultraviolet tools — gear first shipped about a decade ago — the same machines the MATCH Act would now put off-limits.

References

Medium summary of METR time-horizon research medium.com

the 50% task-completion time horizon has been doubling approximately every seven months since 2019… frontier models like o3 have achieved a 50% time horizon of roughly 110 minutes

DigitalApplied — AI Agent Evaluation Pipeline 2026 digitalapplied.com

while an agent may show a 97% success rate in a best-case pass@3 scenario, its actual reliability for a single production run might be as low as 34%

BankInfoSecurity on OpenAI deployment-simulation paper bankinfosecurity.com

frontier models detect they are being tested roughly 99% of the time, leading to sycophantic behavior where the agent performs for the researcher rather than acting as it would for a user

TIME — Ronnie Chatterji, ‘The Two Greatest Opportunities of the AI Economy’ time.com

approximately four million American entrepreneurs—71% of whom are already running active businesses—use ChatGPT for critical functions like marketing and copywriting

AIMultiple — AI Job Loss tracker aimultiple.com

Acemoglu estimates that generative AI will likely contribute a nontrivial but modest productivity boost of just 0.66% to the total U.S. economy over the next decade

Thomson Reuters — ‘AI is ready but firms are not’ thomsonreuters.com

while 18% of these roles face high short-term automation risk, 46% are being redesigned rather than replaced

Manila Times / ACE Robotics press release manilatimes.net

Kairos 3.0-4B… outperformed the 16B-parameter Cosmos 3 and the 28B Lingbot, achieving inference speeds 72 times faster than NVIDIA’s Cosmos 2.5 on A800 GPUs and a top score of 9.30 on WorldModelBench Robot.

Yutori Scouts analyst brief scouts.yutori.com

Critics argue that such dramatic performance gaps often stem from ‘benchmaxxing’—optimizing a model specifically for private or holdout test sets… comparisons between a 4B model and much larger architectures like Cosmos 3 (16B) are often criticized as ‘apples to oranges’.

KrASIA interview with ACE chairman Wang Xiaogang kr-asia.com

World models alone will not solve robotics’ deployment problem… they represent a breakthrough in spatial intelligence, but do not yet solve the entire deployment problem due to a lack of universal benchmarks for long-horizon task execution.

OpenReview audit of LIBERO/RoboTwin benchmarks openreview.net

Researchers have identified ‘shortcut solvability’ in the original LIBERO, where simple probes without language encoders could achieve near-SOTA performance… many reported gains across these leaderboards are not yet provably statistically significant.

TechNode on ACE Robotics funding technode.com

ACE Robotics secured an Angel round led by Ant Group with participation from Qiming Venture Partners, Lenovo Capital, and SenseTime’s investment arm, followed by an Angel+ round in June 2026 including Geely Capital, raising total 2026 funding to hundreds of millions of dollars.

36Kr developer report eu.36kr.com

Despite being labeled ‘parameter-efficient,’ the model requires roughly 23.5GB of VRAM for 480P resolution inference on A800 platforms… 35GB weight files and high VRAM requirements make local workflows difficult without enterprise-grade hardware.

Hugging Face repo (catnip-ai-tech/MaineCoon) huggingface.co

Repository serves as a landing page for the technical report and demo; model weights and source code are not yet available — public access remains restricted as the team prioritizes live testing through their Experience Platform.

Medium review of NVIDIA LongLive (artgor) artgor.medium.com

LongLive at 1.3B parameters achieves ~20.7 FPS on a single H100 (up to 24.8 FPS with FP8); the Helios 14B model reaches 19.5 FPS — MaineCoon’s 47.5 FPS at 22B effectively doubles these prior real-time streaming baselines.

AARD project page (Agentic Auto-Regressive Diffusion) dxlong2000.github.io

Formulates generation as a closed-loop ‘Retrieve–Synthesize–Refine–Update’ cycle with a Multimodal Video Memory, reporting ~30% gains in visual consistency and 20% in narrative coherence over autoregressive baselines on LVBench-C.

Zen van Riel — LTX-2.3 open-source guide zenvanriel.com

LTX-2.3 excels at prompt adherence and speed but exhibits ‘slow-motion drift’ where motion decelerates toward the end of long clips, and its A2V pipeline acts more as a mood conditioner than a precise lip-sync tool.

SiliconANGLE — Tavus Phoenix-3/Raven-0/Sparrow-0 siliconangle.com

Tavus’s Conversational Video Interface stack achieves sub-600ms latency for face-to-face interaction that perceives a user’s emotions and gestures during the call.

VideoWeb.ai — Hedra Omnia review videoweb.ai

Hedra’s Omnia/Character-3 jointly reasons over vision, text, and audio to produce directed scenes where micro-expressions and body language match the emotional subtext of the script — a ‘performance-first’ rather than streaming-first approach.

Sources

References

Jack Sun, writing.