RSI-Anything 在四类任务上量化递归自我改进效果

Recursive self-improvement is one of the most discussed ideas in AI right now.

RSI-Anything measures it in code repair, GUI agents, lab planning, and chemistry.

They find that the lab-planning score more than doubles from 21.80% to 46.01% across five policy versions.

The framework involves a predictive world model that decides which actions and which training data are worth spending costly real feedback on. The feedback then updates both the policy and the world model.

A policy update is kept only if the updated world model still fits it. That check stops the world model from drifting out of date as the policy changes.

In code repair with a fixed coder, success goes from 26.3% with a static verifier to 31.2% after world-model adaptation. In reaction optimization, it reaches a mean best yield of 96.15%, compared with 82.16% for Bayesian optimization with the same 40 experiments.

arxiv.org/abs/2610.13912

Chat with Paper: academy.dair.ai/papers/rsi-a…