Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy

一句话要点

银河通用（Galbot）研究物理世界递归自我改进的来源：提出 PhysRSI-eval，从首次表现、效率、成功、预算、从失败中提升五个维度评测，在 18 个操作任务上比较不同视觉语言模型，并提出“物理木桶假说”。

Galbot investigates where physical recursive self-improvement comes from, introducing PhysRSI-eval with five axes (first-attempt performance, efficiency, success, budget, improvement from failure), evaluating VLMs on 18 manipulation tasks and proposing a Physical Barrel Hypothesis.