Dream-RSI: Recursive Self-Improvement through Evolving Worlds

arxiv:2609.14858

Original source · Paper JSON

Agents, open-endedness, AGI · 2026-09-14

Cites (8) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• Dream-RSI: Recursive Self-Improvement through Evolving Worlds
• Abstract
• 1 Introduction
• 2 Motivation: Discovery History as a Replay Simulator
• 3 Dream-RSI : Recursive Self-Improvement through Evolving Worlds
• Discovery trees and the shared decision interface.
• Online rollout.
• Offline evaluation.
• Replay objective.
• Policy improvement and selection.
• 4 Experiments
• 4.1 Algorithm Engineering
• Baselines and Setup.
• Main Results.
• Recursive Discovery Dynamics.
• Discovered Solver Analysis.
• 4.2 Mathematics Optimization
• Results.
• 4.3 Kernel Engineering
• Results.
• 5 Further Analysis
• 5.1 Analysis of Historical Inductive Biases in Long-Horizon Discovery
• 5.2 Analysis of Evolution of Exploration Behavior
• 6 Related Work
• AI-Driven Scientific and Algorithmic Discovery.
• Self-Evolving Agents.
• Memory, History, and Experience Reuse.
• 7 Conclusion
• Appendix A Detailed Task Description
• Appendix B Prompts
• B.1 Exploration Prompt
• B.2 Replay-Based Policy Improvement Prompt
• Appendix C Discovered Programs
• Instructions for reporting errors