Researchers Taoyong Cui, Pinxian Zeng, Wanghan Xu and colleagues introduce RSI-Anything, a framework for recursive self-improvement that uses predictive world models to guide actions and select learning data. The paper was published on arXiv on October 9 and submitted to Hugging Face Papers on October 12. It builds on JEPA-Anything and aims to turn interaction feedback into persistent system improvements.

RSI-Anything uses observed feedback to improve a system’s policy or workflow and adapt its world model. Component and compatibility checks determine whether updated policy-model pairs are retained. The authors identify two challenges: external feedback can be costly, and world-model predictions may become unreliable as policies evolve.

In code repair with the coder held fixed, success rose from 26.3% with a static verifier to 29.2% with predictive selection, then to 31.2% after world-model adaptation, at comparable test usage. In graphical user interface evaluations, task completion was higher and average actions fell by 39.4%.

Across five successive policy versions, laboratory-planning scores increased from 21.80% to 46.01%. In reaction optimization, RSI-Anything reached a mean best observed yield of 96.15%, compared with 82.16% for Bayesian optimization under the same 40-experiment budget. The paper also reports an end-to-end immune-science application connecting workflow evolution and predictive modeling with therapeutic hypothesis generation.

Researchers Taoyong Cui, Pinxian Zeng, Wanghan Xu and colleagues introduce RSI-Anything, a framework for recursive self-improvement that uses predictive world models to guide actions and select learning data. The paper was published on arXiv on October 9 and submitted to Hugging Face Papers on October 12. It builds on JEPA-Anything and aims to turn interaction feedback into persistent system improvements.

RSI-Anything uses observed feedback to improve a system’s policy or workflow and adapt its world model. Component and compatibility checks determine whether updated policy-model pairs are retained. The authors identify two challenges: external feedback can be costly, and world-model predictions may become unreliable as policies evolve.

In code repair with the coder held fixed, success rose from 26.3% with a static verifier to 29.2% with predictive selection, then to 31.2% after world-model adaptation, at comparable test usage. In graphical user interface evaluations, task completion was higher and average actions fell by 39.4%.

Across five successive policy versions, laboratory-planning scores increased from 21.80% to 46.01%. In reaction optimization, RSI-Anything reached a mean best observed yield of 96.15%, compared with 82.16% for Bayesian optimization under the same 40-experiment budget. The paper also reports an end-to-end immune-science application connecting workflow evolution and predictive modeling with therapeutic hypothesis generation.