AKV: Cache Engineering for Agents A new serving system that lets LLM agents directly manage their KV cache — dropping unused history and repositioning retained keys without rewriting or re-prefilling. Reduces average cost by 27.7% while improving task accuracy by 1.7 points.

Microsoft Research's WildAWM A scalable pipeline that synthesizes executable computer-use agent environments and tasks using only open-source models, yielding 341 long-horizon env-task pairs and Qwen agents that generalize to 8 unseen benchmarks.

Meta just released MemProBench on Hugging Face 800 frozen memory histories from real-world Ego4D activity, built to study memory-driven proactive assistance: when should an assistant speak up, not just what to say?

Sage-A1.5 Shanghai AI Laboratory just released a 35B MoE agent on Hugging Face that reaches trillion-parameter performance: it orchestrates parallel subagents, verifies its own progress, and conducts open scientific discovery.

Ropedia W0 A world model that learns loco-manipulation from 10k hours of human egocentric video, then transfers across embodiments, from robot arms to humanoids and from grippers to dexterous hands.

Skywork AI's Matrix-Game 4.0 A playable game world rendered by a video model in real time: 480p at 60 FPS on a single RTX 5090, with a key-press-to-display latency of just 112ms.

PAR: Progress-Aligned Recurrence for length generalization in looped LMs Loop k learns to predict the state after k operations, and recurrent states are rescaled to first-iteration norms. A 0.5B Qwen2.5 trained on 32 steps hits 95.5% at 512 steps—16x extrapolation.

NVIDIA just released a driving simulation evaluation dataset on Hugging Face 91 multiview driving scenes with 11 cameras each, featuring World Scenario Map controls and text prompts.