AI
- Hugging Face Papers — RRSI — Google proposes “regularized recursive self-improvement” so an agent harness can iteratively rewrite/improve itself without the self-improvement loop degrading into unstable or overfit versions.
- Hugging Face Papers — Grounded Action Model — NUS researchers argue 3D grounding, not 2D vision-language alone, should be the foundation representation for general-purpose robot action models.
- Hugging Face Papers — WorldCrafter — Tencent ARC Lab’s video world model keeps long generations visually/spatially consistent via an implicit, updateable 3D-aware memory instead of frame-by-frame conditioning.
- arXiv cs.CL — OSWorld-Pro — Proposes scoring computer-use agents on the correctness of each intermediate step, not just final task success, to catch agents that “succeed” via wrong or unsafe actions along the way.
- arXiv cs.CL — DolphinBench — Benchmarks agent-memory systems on where they sit on the cost/latency/recall Pareto frontier, so teams can pick a memory design for their constraints rather than assuming one design wins outright.
- arXiv cs.CL — onPanda — Proposes token-level correction (vs. full re-labeling) as a cheaper way to annotate on-policy alignment data for LLMs and agents.
General
- Stack Overflow Blog — AMD’s VP of Software Anush Elangovan argues agentic AI is drastically lowering the barrier to entry for low-level GPU/hardware programming via ROCm, and that software and hardware development timelines are rapidly converging as a result.