RSI-Anything: Recursive Improvement Through World Models

Introduction

A system that improves recursively must do more than generate a better answer once. It needs to turn experience from repeated interaction into updates that remain useful later. This is the goal of recursive self-improvement, or RSI. In practice, however, external evaluation can be expensive, while a world model trained for one policy may become unreliable after that policy evolves.

RSI-Anything addresses this problem with a predictive world-model framework built on JEPA-Anything. Rather than tying the method to one application, the framework uses predictions to estimate the consequences of possible actions and to select learning data. Real feedback is then used to improve the policy or workflow, and, when needed, to adapt the world model itself.

How the framework works

• Prediction-guided action selection: The system uses its world model to prioritize actions that are expected to be informative or effective, reducing reliance on indiscriminate interaction.
• Feedback-based improvement: Observed outcomes can update an agent policy or an entire workflow, allowing the method to operate beyond conventional parameter-level learning.
• Conditional model adaptation: The world model is adapted when policy changes make its predictions less reliable, rather than being updated automatically after every iteration.
• Retention checks: Component-level and compatibility checks determine whether an updated policy-model pair should be kept, helping prevent a local improvement from destabilizing the larger system.

Reported results

The paper evaluates the approach across several computational and scientific settings. In code repair, with the coder held fixed, success rises from 26.3% with a static verifier to 29.2% after predictive selection, and to 31.2% after world-model adaptation, with comparable test usage. In GUI evaluations, task completion improves while the average number of actions falls by 39.4%, suggesting that prediction can reduce unnecessary interaction as well as guide behavior.

Laboratory-planning scores increase from 21.80% to 46.01% across five successive policy versions. In reaction optimization, under the same budget of 40 experiments, RSI-Anything reports a mean best observed yield of 96.15%, compared with 82.16% for Bayesian optimization. The authors also describe an end-to-end immune-science application that connects workflow evolution and predictive modeling with therapeutic hypothesis generation.

Why it matters

The important contribution is the proposed loop rather than any single benchmark number. The world model helps allocate scarce interaction, real feedback corrects its assumptions, and retention checks turn successful iterations into reusable system versions. This combination could be useful for coding agents, laboratory automation, and scientific discovery systems that must operate over long horizons.

The available material does not include the full experimental protocols, datasets, training details, or statistical analyses. The results should therefore be read as evidence for a promising framework, not as proof that recursive improvement will transfer uniformly to every domain. Long-term prediction drift, policy-model compatibility, and stability across many iterations remain important questions.

Source: Hugging Face Daily Papers