RSI-Anything: Unlocking Recursive Self-Improvement with World Models
The paper claims recursive self-improvement can be made more reliable by pairing policy updates with adaptive world-model checks.
RSI-Anything uses predictive world models to guide actions, choose learning data, and decide whether updated policy-model pairs should be kept. The authors report gains across code repair, GUI tasks, lab planning, reaction optimization, and an immune-science workflow. In the examples given, code-repair success rises from 26.3% to 31.2%, GUI actions fall by 39.4%, and lab-planning scores more than double across five policy versions. Source: ArXiv · AI/CL/LG's note.
RSI-Anything uses predictive world models to guide actions, choose learning data, and decide whether updated policy-model pairs should be kept. The authors report gains across code repair, GUI tasks, lab planning, reaction optimization, and an immune-science workflow. In the examples given, code-repair success rises from 26.3% to 31.2%, GUI actions fall by 39.4%, and lab-planning scores more than double across five policy versions. Source: ArXiv · AI/CL/LG's note.
score 4