Researchers have introduced RSI-Anything, a domain-general framework designed to unlock recursive self-improvement (RSI) by using predictive world models to guide both action and the selection of learning data. The work, published as a paper on Hugging Face, builds on JEPA-Anything and addresses a core challenge in RSI: external feedback is costly, and world-model predictions can become unreliable as policies evolve.
The framework couples observed feedback with conditional world-model adaptation, allowing policies or workflows to improve over time. Component and compatibility checks govern whether updated policy-model pairs are retained, helping ensure that only reliable improvements persist.
Across diverse computational and scientific domains, RSI-Anything reports substantial gains. In code repair, with the coder fixed, success rose from 26.3% with a static verifier to 29.2% through predictive selection and 31.2% after world-model adaptation, at comparable test usage. In GUI evaluations, task completion improved while average actions decreased by 39.4%.
In laboratory planning, scores more than doubled across five successive policy versions, climbing from 21.80% to 46.01%. For reaction optimization, RSI-Anything achieved a mean best observed yield of 96.15% versus 82.16% for Bayesian optimization under the same 40-experiment budget. An end-to-end immune-science application further connects workflow evolution and predictive modeling with therapeutic hypothesis generation.
The authors say these results demonstrate how predictive world models and retained system updates can enable recursive improvement across diverse environments. Code is available at github.com/Gen-Verse/RSI-Anything.