Hyperagents

arxiv:2603.19461

Original source · Paper JSON

Agents, open-endedness, AGI · 2026-03-19

Cites (14) · Cited by (5)

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
7. Part 7

Contents

• HyperAgents
• Abstract
• 1 Introduction
• 2 Related Work
• 3 Methods
• 4 Experiment Setup
• 4.1 Baselines
• 4.2 Domains
• 5 Results
• 5.1 Improving Task Performance
• 5.2 Improving the Ability to Improve
• 5.3 Compounding Self-Improvements
• 6 Safety Discussion
• 7 Limitations and Conclusion
• Acknowledgments
• Author Contributions
• Appendix A Algorithmic details
• A.1 Initial Agent
• A.2 Parent Selection
• A.3 Pseudocode
• A.4 Multi-domain Optimization
• Appendix B Baseline Details
• Appendix C Domain Details
• C.1 Polyglot
• C.2 Paper Review
• C.3 Robotics Reward Design
• C.4 Olympiad-level Math Grading
• Appendix D Experiment Details
• D.1 Hyperparameters for FMs
• D.2 Cost Estimate
• D.3 Improvement@k Metric
• D.4 Transfer Agent Selection
• Appendix E Additional Results
• E.1 Best Discovered Task Agents
• E.1.1 Paper Review
• E.1.2 Robotics Reward Design
• E.1.3 Olympiad-level Math Grading
• E.2 Qualitative: Improving Task Performance
• E.3 Qualitative: Improving the Ability to Improve
• E.3.1 Better Evaluation Analysis
• E.3.2 Compute-Aware Strategic Planning
• E.3.3 Automated Bias Detection and Correction
• E.3.4 Performance Tracking Infrastructure
• E.3.5 Prompt Template Systems
• E.3.6 Refinement Based on Multi-Generation Performance
• E.3.7 Persistent Memory
• E.4 Olympiad-level Math Graders
• E.5 Modifying Parent Selection
• Appendix F Additional Safety Discussion
• Instructions for reporting errors