Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy
arXiv cs.AI · 20 Oct 2026, 03:59 AEST
The paper investigates the mechanisms and scaling behavior of physical Recursive self-improvement (RSI) in VLMs. It introduces a five-axis metric to evaluate different capabilities across manipulation tasks and proposes the Physical Barrel Hypothesis.
Key facts
- We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and Improvement from failure.
- Different VLMs are evaluated on 18 manipulation tasks under a fixed interaction budget.
- The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains.
- Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier.
Read aloud
The paper investigates the mechanisms and scaling behavior of physical Recursive self-improvement RSI in VLMs. It introduces a five-axis metric to evaluate different capabilities across manipulation tasks and proposes the Physical Barrel Hypothesis. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and Improvement from failure. Different VLMs are evaluated on eighteen manipulation tasks under a fixed interaction budget. The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains. Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier.
Sources in this cluster
- Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy
arXiv cs.AI · 20 Oct 2026, 03:59 AEST