• Arbor runs the full research loop on its own, accumulating experience across rounds through a Hypothesis Tree. It topped all six real research tasks, with average relative gains over 2.5x those of Codex and Claude Code.

• Environments are becoming the new scaling axis. RACES treats verifiable environments as LEGO bricks and composes them recursively, so 50 base environments deliver roughly the training value of 300 independent ones.

• InternVideo3 brings the agentic playbook to long video. Observation, reasoning, tools, and memory share one evolving context, turning long-video understanding into a loop of gathering evidence and verifying it.

• Falling MTP acceptance rates trace back to rising entropy during RL. Bebop pairs rejection sampling with a TV loss to push acceptance as high as 95% and speed end-to-end training by 1.8x.

• Reach for a lighter lens before training an SAE. ICA Lens pulls readable directions straight from activation geometry without training any dictionary, matching public SAEs on SAEBench.

Also Notable

• Pretrained Video Generators Can Plan Without Text — World Model self-distillation distills task-solving ability and drops the dependence on detailed text descriptions.
• Stop Post-Training Diffusion Language Models With Random Masks — attention-guided denoising exploits the intrinsic dependencies between tokens better than random masking.
• VLA Models Aren't Robust to How Instructions Are Phrased — the first systematic multilingual evaluation finds language sensitivity surfaces step by step during execution.
• LLM Judges Hit a Ceiling on Scientific Novelty — this paper steps back to grade a cleaner upstream object: the research question itself.
• Multimodal ICL Is Stuck on Context Window and KV Cache Cost — task-aware structured memory offers a path to dynamic compression.
• Every Turn of a Conversation Carries a Swelling History — incremental compression with cross-turn memory sharing preserves more than naive truncation or summarization.
• Redundant, Unique, and Synergistic Information Shift Per Sample — an information-theoretic decomposition pulls these dynamics apart for the first time.
• VLMs Still Miss the World's Dynamics — NVIDIA's 4DP-QA turns 4D perception into scalable QA to quantify the gap.
• How to Build Agents That Refuse Responsibly — Google argues machine non-compliance comes in many distinct forms.
• A Scalable Metric for Language-Model Creativity — automatic evaluation across open-ended tasks measures creative potential systematically.

Read the full edition →