RSI-Anything: World Models Enable Cross-Domain Agent Self-Improvement

Recursive Self-Improvement (RSI) is emerging as a hot research direction for AI agents. Researchers from PhAI Labs, the Chinese University of Hong Kong, and the University of Macau have open-sourced the RSI-Anything framework, which unifies world-model prediction with real-environment interaction feedback. The framework lets an agent predict the outcomes of candidate actions, select the most promising ones for execution, learn from the actual feedback, and retain verified improvements for the next iteration.

Framework Design

RSI-Anything builds on the earlier JEPA-Anything framework, which uses Orthogonal Prediction Factorization (OPF) to decompose complex states into complementary prediction factors, predict each separately, and recombine them into a complete latent state. The RSI loop operates as follows:

Candidate generation – the policy proposes actions.

Prediction – the world model forecasts the result of each candidate.

Selection – candidates are scored by predicted task reward, feasibility, constraints, and uncertainty; only the most promising are executed.

Execution & evaluation – real feedback is collected.

Learning – both policy and world model are updated.

Component check – after policy update, task performance and constraints are verified; after world-model update, prediction quality is verified.

Compatibility check – the new policy–world-model pair is tested for mutual compatibility.

Retention – only updates passing all checks are kept for the next round.

This mechanism addresses the co-adaptation problem: policy updates change the action distribution, potentially invalidating the world model, and world-model updates alter candidate selection, affecting the policy.

Multi-Domain Evaluation

Code Repair

Using a fixed 7B code generation model and identical candidate budget, the hidden test success rate rose from 26.3% (static verifier) to 29.2% (fixed world model) to 31.2% (adaptive world model). Online test calls remained ~3.7 per task, confirming gains come from smarter selection, not more trials. Results on APPS and LiveCodeBench also beat random selection and static verifiers.

GUI Agent

On MiniWoB++ checkbox-clicking tasks, success rate climbed from a baseline of 51% to 99% , average action count dropped 39.4% , repeated clicks fell from 237 to 0 , and erroneous submissions decreased sharply. On the more complex OSWorld benchmark, task success rate improved 18.4% relative to the initial agent.

Lab Planning

On 200 Level-2 tasks in LabHorizon, a single 4B planning policy iterated through five versions; comprehensive score rose steadily from 21.80% to 46.01% , surpassing an Actor-Simulator-Selector reference system. Level-1 next-action accuracy reached 0.700 , demonstrating cumulative planning improvement through retained policy versions.

Chemical Reaction Optimization

In two Suzuki–Miyaura coupling screens, RSI-Anything used world-model yield predictions to guide condition selection under the same experimental budget. One reaction reached 93.0% optimal yield after 20 experiments (vs. 87.5% for Bayesian optimization); the other averaged 96.15% after 40 experiments (vs. 82.16% ). This shows world-model-guided selection outperforms traditional Bayesian optimization when experiments are limited.

Immunology End-to-End Application

The team evolved a scientific-agent workflow with Agent Genesis , built an immunology world model, and combined model predictions with experimental evidence for self-correcting analysis. This produced a therapeutic hypothesis: "SIRPα inhibition + IL-36γ axis modulation" . The paper explicitly notes this hypothesis still requires wet-lab validation.

Conclusion

RSI-Anything provides a reusable cross-domain loop: propose candidates → predict consequences → select actions → execute & evaluate → learn from feedback → retain improvements . The same mechanism delivers measurable gains in code repair, GUI automation, lab planning, chemical optimization, and scientific hypothesis generation, demonstrating that world-model-guided recursive improvement can generalize across vastly different tasks.