[论文] RSI-Anything: Unlocking Recursive Self-Improvement with World Models

论文概要

研究领域: NLP
作者: Taoyong Cui, Pinxian Zeng, Wanghan Xu, Hao Xiao, Shuhan Xue, Qiang Gao, Wanli Ouyang, Pheng Ann Heng, Zhenfei Yin, Yingcheng Wu, Ling Yang
发布时间: 2026-10-09
arXiv: 2610.13912

中文摘要

递归自我改进（RSI）旨在将交互经验转化为持续的能力提升，但外部反馈成本高昂，且随着策略演进，世界模型的预测可能变得不可靠。我们提出 RSI-Anything——一个基于 JEPA-Anything 构建的领域通用框架，利用预测式世界模型来指导动作和学习数据的选择。观测反馈驱动策略或工作流的改进以及条件化的世界模型自适应，而组件与兼容性检查则管理更新后策略-模型对的保留。在多个计算与科学领域中，RSI-Anything 展现出显著增益。在代码修复任务中（编码器固定），静态验证器的成功率为 26.3%，通过预测选择提升至 29.2%，经世界模型自适应后进一步达到 31.2%，且测试用量相当。图形用户界面（GUI）评估显示任务完成率更高，同时平均动作数减少 39.4%。在连续五个策略版本中，实验室规划得分从 21.80% 翻倍至 46.01%。在反应优化中，在相同的 40 次实验预算下，RSI-Anything 达到 96.15% 的平均最优观测产率，而贝叶斯优化为 82.16%。一个端到端的免疫科学应用进一步将工作流演进与预测建模和治疗性假设生成联系起来。

原文摘要

Recursive self-improvement (RSI) aims to transform interaction experience into persistent improvements, yet external feedback is costly and world-model predictions may become unreliable as policies evolve. We introduce RSI-Anything, a domain-general framework built on JEPA-Anything that uses predictive world models to guide action and learning-data selection. Observed feedback drives policy or workflow improvement and conditional world-model adaptation, while component and compatibility checks govern the retention of updated policy-model pairs. Across diverse computational and scientific domains, RSI-Anything demonstrates substantial gains. In code repair, with the coder fixed, success increases from 26.3% with a static verifier to 29.2% through predictive selection and 31.2% after world-m...

自动采集于 2026-10-14

#论文 #arXiv #NLP #小凯