RSI-Anything Uses World Models for Recursive Self-Improvement, Beats Bayesian Optimization on Reaction Yields
ylecun · x · 2026-10-12
Ling Yang's team introduces RSI-Anything (arXiv:2610.13912), a domain-general recursive self-improvement framework built on JEPA-Anything: predictive world models guide action selection and learning-data selection, observed feedback drives policy/workflow improvement plus conditional world-model adaptation, and component/compatibility checks decide which updated policy-model pairs are retained. Validated across four domains:
- Code repair: with the coder fixed, success rises from 26.3% (static verifier) to 29.2% (predictive selection) and 31.2% (world-model adaptation) at comparable test usage
- GUI agents: higher task completion with 39.4% fewer average actions
- Lab planning: scores more than double from 21.80% to 46.01% over five policy versions
- Reaction optimization: mean best observed yield of 96.15% vs 82.16% for Bayesian optimization under the same 40-experiment budget
An end-to-end immune-science application ties it together. Yann LeCun amplified the release. Paper and code are public.
Related event: RSI-Anything: World models enable recursive self-improvement across domains(4 posts)→
More from AGI Musings
- Vitalik releases 32-chapter sci-fi novel Snowmoon, introducing his self-invented 'Coordination Quotient' — devanshmehta · 2026-10-12
- One normal week of science now equals a decade of past progress — Dr_Singularity · 2026-10-12
- Rumors claim Google, OpenAI and Anthropic are all near recursive self-improvement — FinanceYF5 · 2026-10-12
- Going all-in on AI only makes sense for the top 0.1%, argues VC debate — vaibhavbetter · 2026-10-12
- From Poke to OpenAI's Dots: the personal messaging agent era is here — sujingshen · 2026-10-12
- IMF warns Thailand is aging before getting rich as AI gains come too late — AIFlow_ML · 2026-10-12