Embodied agents and multi-agent orchestration
Key Questions
What is RotVLA and what benchmark does it lead?
RotVLA achieves 98.2% on LIBERO using continuous rotational latent actions represented as SO(n) matrices, establishing a new state-of-the-art for robot control.
How does Microsoft FastContext improve agent performance?
Microsoft FastContext reduces token waste during repository exploration by agents, enabling more efficient multi-agent orchestration in code-related tasks.
What is OpenClaw-Skill and its approach to skill libraries?
OpenClaw-Skill proposes tree search over flat distillation to build reusable agent skill libraries, improving efficiency in embodied AI systems.
What does the Critique of Agent Model paper address?
The paper delivers a theoretical and philosophical critique of current agent architectures, highlighting limitations in existing designs for embodied and multi-agent systems.
What new benchmarks target long-horizon agent tasks?
CEO-Bench focuses on long-horizon agent tasks, while PreAct explores reusing past reasoning for computer-using agents, alongside WEAVER and RepWAM for world modeling efficiency.
RotVLA achieves 98.2% LIBERO using continuous rotational latent actions (SO(n) matrices) — SOTA for robot control. Microsoft FastContext reduces token waste in agent repo exploration. WeaveBench (long-horizon computer-use agent benchmark). HarnessBridge (learnable bidirectional controller). Socratic-SWE (50.4% on SWE-bench Verified). Multi-agent ChatOps architecture. World model paper using 2D stick-figure skeletons. Qwen-VLA. NIST humanoid benchmark. SePO. Flash-WAM. OpenClaw-Skill proposes tree search over flat distillation for reusable agent skill libraries. Critique of Agent Model paper challenges current agent architectures. New: CEO-Bench targets long-horizon agent tasks. New: PreAct proposes reusing past reasoning for computer-using agents. WEAVER and RepWAM also contribute to embodied AI efficiency.