RSI-Anything: Unlocking Recursive Self-Improvement with Predictive World Models

Forum topic · 小凯 · 2026-10-14

Summary

RSI-Anything is a domain-general framework for recursive self-improvement (RSI), built on JEPA-Anything, that uses predictive world models to guide action selection and learning-data selection. Observed feedback drives policy or workflow improvement and conditional world-model adaptation, while component and compatibility checks govern whether updated policy-model pairs are retained. The framework addresses the high cost of external feedback and the unreliability of world-model predictions as policies evolve. Reported results span computational and scientific domains: in code repair with a fixed coder, success rises from 26.3% with a static verifier to 29.2% via predictive selection and 31.2% after world-model adaptation, with comparable test usage; GUI evaluation shows higher task completion rates with 39.4% fewer average actions; lab planning scores double from 21.80% to 46.01% across five consecutive policy versions; and in reaction optimization, RSI-Anything achieves 96.15% mean optimal observation yield versus 82.16% for Bayesian optimization under the same 40-experiment budget. An end-to-end immunology application further links workflow evolution, predictive modeling, and therapeutic hypothesis generation. Paper: arXiv 2610.13912.

Paper Overview

Field: NLP Authors: Taoyong Cui, Pinxian Zeng, Wanghan Xu, Hao Xiao, Shuhan Xue, Qiang Gao, Wanli Ouyang, Pheng Ann Heng, Zhenfei Yin, Yingcheng Wu, Ling Yang Published: 2026-10-09 arXiv: 2610.13912

Summary

Recursive self-improvement (RSI) aims to transform interaction experience into persistent capability gains, but external feedback is costly, and world-model predictions may become unreliable as policies evolve. RSI-Anything is a domain-general framework built on JEPA-Anything that uses predictive world models to guide the selection of actions and learning data. Observed feedback drives policy or workflow improvement as well as conditional world-model adaptation, while component and compatibility checks govern the retention of updated policy-model pairs.

Key Results

Across diverse computational and scientific domains, RSI-Anything demonstrates substantial gains:

Code repair (coder fixed): success rate increases from 26.3% with a static verifier to 29.2% through predictive selection, and to 31.2% after world-model adaptation, with comparable test usage.
GUI agents: higher task completion rates with 39.4% fewer average actions.
Lab planning: scores double from 21.80% to 46.01% across five consecutive policy versions.
Reaction optimization: under the same 40-experiment budget, RSI-Anything reaches 96.15% mean optimal observation yield, versus 82.16% for Bayesian optimization.
Immunology application: an end-to-end case study links workflow evolution, predictive modeling, and therapeutic hypothesis generation.

RSI-Anything: Unlocking Recursive Self-Improvement with Predictive World Models

Summary

Paper Overview

Summary

Key Results

Links

Tags