Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents

arxiv:2505.22954

Original source · Paper JSON

Agents, open-endedness, AGI · 2025-05-29

Cites (13) · Cited by (13)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4
5. Part 5
6. Part 6

Contents

• Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
• Abstract
• 1 Introduction
• 2 Related Work
• 3 Darwin Gödel Machine
• 4 Experiments
• 4.1 Experiment Setup
• 4.2 Benchmarks
• 4.3 Baselines
• 4.4 Results
• 5 Safety Discussion
• 6 Conclusion and Limitations
• Ethics Statement
• Reproducibility Statement
• Acknowledgments
• Appendix
• Table of Contents
• Appendix A Additional Results
• A.1 Baselines on SWE-bench
• A.2 Generality across models on Polyglot
• A.3 Ablation of Parent Selection
• A.4 Additional Statistics of DGM runs
• Appendix B Additional Related Work
• Appendix C Algorithmic Details
• C.1 Initial Coding Agent
• C.2 Parent Selection
• C.3 Self-Improve Prompts
• C.4 Pseudocode
• Appendix D Experiment Details
• D.1 Hyperparameters for Foundation Models
• Appendix E Benchmark Details
• E.1 Cost Estimate
• E.2 SWE-bench Tasks
• E.3 Polyglot Tasks
• E.4 SWE-bench State-of-The-Art
• E.5 Polyglot Representative Agent
• Appendix F Best-Discovered Agents
• F.1 DGM on SWE-bench
• F.2 DGM on Polyglot
• Appendix G Similar Target Functionality, Different Implementations
• Appendix H Case Study: Solving Hallucination
• Appendix I Additional Safety Discussion
• Appendix J Additional Future Work Directions
• Instructions for reporting errors