A research-led Monday after a quiet weekend. The agent-reliability thread sharpened: When Tools Fail benchmarks dynamic replanning and anomaly recovery, ResearchClawBench evaluates end-to-end autonomous research, SubtleMemory tests fine-grained relational memory in long-horizon agents, and OpenSkill proposes open-world self-evolution for LLM agents past the usable-learning-loop assumption. Spatial reasoning came back into focus with Stream3D-VLM and Imaginative Perception Tokens. Robots Need More than VLA and World Models argues against the policy-scaling-is-enough framing for generalist robot intelligence. No press cycle on the day.
Four benchmarks tighten what production-grade agent reliability means: dynamic replanning when tools fail, end-to-end autonomous scientific research, fine-grained relational memory over long horizons, and open-world self-evolution where the learning loop isn't guaranteed.
Benchmarks how LLM agents replan and recover when tools misbehave — covering the failure modes 'happy-path' tool-use evals systematically miss.
Evaluates AI coding agents on end-to-end autonomous scientific research, not on isolated subtasks.
Tests whether long-horizon agents can keep fine-grained relational distinctions in memory across accumulating interactions.
Self-evolving agents that adapt after deployment without the curated learning loops most prior approaches assume.
Stream3D-VLM brings 3D understanding online with incremental geometry priors, Imaginative Perception Tokens enhance VLM spatial reasoning, and Thinking with Imagination uses world simulators for agentic visual spatial reasoning. The week's earlier 'VLMs don't actually know spatial' critique is being answered.
Brings 3D scene understanding online for 3D multimodal models — incremental geometry priors instead of requiring the whole scene up front.
Imaginative perception tokens give VLMs structured intermediate state for spatial reasoning when info isn't visible.
Couples VLMs with world simulators to extend their visual spatial reasoning beyond observation.
A position paper argues generalist robot intelligence is not just a policy-scaling problem; AnchorWorld extends interactive world modeling with view-based evolution; Direct 3D-Aware Object Insertion ships a clean compositing tool.
Argues generalist robot intelligence requires more than scaling vision-language-action and world models — task structure, planning, and abstraction matter.
Embodied egocentric world simulation that supports customizable view-based evolution.