RSI-Anything: World Models Drive Recursive Self-Improvement Across Code, GUI, and Labs
机器之心 · wechat · 2026-10-12
PhAILabs released the RSI-Anything technical report (arXiv 2610.13912), a framework where world models predict action outcomes, real feedback refines both policy and world model, and compatibility checks gate each update. Results: code-fix success on CodeContests rose from 26.3% to 31.2% without changing the coder model; 18.4% relative gain on OSWorld GUI tasks; a 4B policy model improved lab-planning scores from 21.80% to 46.01% over five generations, beating a larger reference system; chemistry optimization reached 96.15% best yield in 40 experiments vs 82.16% for Bayesian optimization. An immunology case produced a SIRPα/IL-36γ combination intervention hypothesis pending validation. Code is open-sourced.
Related event: RSI-Anything: World models enable recursive self-improvement across domains(4 posts)→
More from AGI Musings
- AI commoditizes building: founders must knock on doors like it's the 1980s — FinanceYF5 · 2026-10-12
- What Happened to Other Species That Met a 'Superintelligence'? A Dark AI-Safety Analogy — CatAstro_Piyush · 2026-10-12
- Agents burn 5x more tokens than people: 7.3T weekly on OpenRouter, up 14x in six months — rohanpaul_ai · 2026-10-12
- Todoist hits record revenue as top AI user explains why he won't vibe-code a to-do app — eptwts · 2026-10-12
- One year on, AI pundit says we're inching toward 'Digital Grothendieck' — burny_tech · 2026-10-12
- 'Fable Took My Job': a developer's first-hand account of an AI coding agent taking over — jllyhill · 2026-10-12