RSI-Anything uses world models to measure recursive self-improvement, doubling lab-planning scores
dair_ai · x · 2026-10-15
Recursive self-improvement (RSI) is one of the most discussed ideas in AI, and a new paper, RSI-Anything (PhAI Labs et al., arXiv 2610.13912), introduces a domain-general framework to measure and enable it across code repair, GUI agents, lab planning, and chemistry.
Method:
- A predictive world model decides which actions and training data are worth spending costly real feedback on
- Observed feedback updates both the policy and the world model
- A policy update is kept only if the updated world model still fits it, a compatibility check that prevents the world model from drifting out of date
Key results:
- Lab-planning score more than doubles from 21.80% to 46.01% across five policy versions
- Code repair with a fixed coder: 26.3% with a static verifier → 29.2% with predictive selection → 31.2% after world-model adaptation
- GUI evaluations show higher task completion with 39.4% fewer average actions
- Reaction optimization: 96.15% mean best yield vs 82.16% for Bayesian optimization under the same 40-experiment budget
- An end-to-end immunology application ties workflow evolution to therapeutic hypothesis generation
Code and models are open-sourced.
More from Research
- arXiv paper gives StopGrads theoretical grounding: convergence without an underlying loss — srush_nlp · 2026-10-23
- PEARL demo: an LLM agent that checks experiment results, revises plans, and re-tests — KevinKaichuang · 2026-10-23
- Berkeley's StyleStream 2.0 cuts streaming voice conversion latency to 520ms, 2x faster — Berkeley · 2026-10-23
- Math PhD student quits academia: AI is making math unrecognizable — pickover · 2026-10-23
- ICLR Submission Found Containing LLM's Own Meta-Comment Admitting Results Weren't Significant — _onionesque · 2026-10-23
- AKV Serving System Lets Agents Manage Their Own KV Cache, Cutting Cost 27.7% — _akhaliq · 2026-10-23