RSI-Anything Lifts Code Repair to 31.2% Using World Models

A single framework claims to drive recursive agent self-improvement across code repair, GUI control, laboratory planning, and chemistry optimization by letting a predictive world model, not just external feedback, pick which actions to take and which experiences to train on. The authors call it RSI-Anything and posted the preprint to arXiv with code at Gen-Verse/RSI-Anything.

The gains are self-reported. On code repair with the coder fixed, "success increases from 26.3% with a static verifier to 29.2% through predictive selection and 31.2% after world-model adaptation," at what the authors call comparable test usage. On a laboratory-planning benchmark, scores "more than double from 21.80% to 46.01%" across five successive policy versions. GUI evaluations show higher task completion while "average actions decrease by 39.4%." For reaction optimization, the method reports "96.15% mean best observed yield versus 82.16% for Bayesian optimization under the same 40-experiment budget."

The paper describes the mechanism in domain-general terms: "RSI-Anything, a domain-general framework built on JEPA-Anything that uses predictive world models to guide action and learning-data selection." Observed feedback drives policy or workflow updates, while component and compatibility checks decide which policy-model pairs survive.

The paper does not break out per-domain compute cost or name the specific base coder, GUI controller, or lab-planning policy carrying each result, and the preprint has not been peer reviewed. An end-to-end immune-science application is also mentioned, applying the same loop to therapeutic hypothesis generation.

RSI-Anything joins a run of agent-stack preprints we've tracked this week, including Sakana's MASS, which recycles one 27B model across solver, judge, and optimizer roles.