RSI-Anything: Unlocking Recursive Self-Improvement with World Models

RSI-Anything. Unlocking Recursive Self-Improvement with World Models. By Taoyong Cui and ten others, from PhAI Labs, the Chinese University of Hong Kong, and six other institutions. Published on the ninth of October. Its code repository is worked out: about one hundred and seventy source files, with a readme and a licence.

Recursive self-improvement aims to turn interaction experience into lasting improvement. Two things get in the way. External feedback is costly to obtain. And the predictions of a world model can become unreliable as the policy it serves keeps changing. A system that improves itself therefore needs a way to spend scarce real feedback carefully, and a way to stop its own predictive model from drifting out of step with the policy it is meant to describe.

RSI-Anything is a domain-general framework, built on JEPA-Anything, that uses predictive world models to guide two choices: which action to take, and which data to learn from.

Observed feedback does double duty. It drives improvement of the policy or workflow, and it drives conditional adaptation of the world model. Updates are not simply kept, though. Component checks and compatibility checks govern whether an updated pair of policy and model is retained. The policy and the world model are treated as a pair, so a change to one is kept only if the two still fit together. That is the answer to the second problem: as the policy evolves, the world model is adapted alongside it, and the compatibility check guards the match.

The gains are reported across diverse computational and scientific domains. In code repair, with the coder held fixed, success rises from twenty-six point three percent with a static verifier, to twenty-nine point two percent through predictive selection, and to thirty-one point two percent after world-model adaptation, at comparable test usage. In graphical user interface tasks, completion is higher while the average number of actions falls by thirty-nine point four percent.

Over five successive policy versions, laboratory-planning scores more than double, from twenty-one point eight percent to forty-six point zero one percent. In reaction optimization, under the same budget of forty experiments, RSI-Anything reaches a mean best observed yield of ninety-six point one five percent, against eighty-two point one six percent for Bayesian optimization. An end-to-end application in immune science also links workflow evolution and predictive modelling with the generation of therapeutic hypotheses.

Two ingredients carry the argument: a world model that predicts which actions and data are worth the cost of real feedback, and a gate that only keeps updates which check out. Together, the authors argue, they let improvement compound across very different environments.