A new paper on arXiv systematically examines physical recursive self-improvement (RSI) in LLM-based policies. The authors introduce PhysRSI-eval, a benchmark designed to quantify physical RSI through a five-axis metric, with the first axis being First-Attempt performance. The work aims to uncover where physical RSI capability comes from, rather than merely demonstrating it.
The study reports an observational scaling law for LLM-as-policy, connecting model scale or data scale to RSI capability in physical domains. The abstract also indicates a focus on generalization, suggesting the authors test whether RSI abilities transfer across tasks or environments.
Because the abstract is truncated, the full details of the five axes and the scaling law are not available in the source. The paper appears to be a step toward formalizing RSI evaluation, but readers should consult the full text for specifics on methodology and results.