Towards Execution-Grounded Automated AI Research

arxiv:2601.14525

Original source · Paper JSON

Post-training · 2026-01-20

Cites (7) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4

Contents

• Towards Execution-Grounded Automated AI Research
• Abstract
• Keywords:
• 1 Introduction
• 2 Automated Idea Executor
• 2.1 Research Environments for Ideation
• 2.2 System design
• Implementer
• Scheduler
• Worker
• 3 Benchmarking LLM Ideators and Executors
• 3.1 End-to-End Ideation and Execution
• 3.2 Comparing Ideators with the Same Executor
• 4 Execution-Guided Evolutionary Search
• 4.1 Search Scaffold
• 4.2 Experiment Results
• 4.3 Comparison with Best-of-N
• 4.4 Analysis of Generated Ideas
• 5 Reinforcement Learning from Execution Reward
• 5.1 Reward Design and Experiment Setup
• 5.2 Experiment Results
• Positive Training Curves for Average Reward
• The Case of Max Reward
• 5.3 Analysis of Training Dynamics
• Thinking Length
• Diversity Collapse
• 6 Related Work
• AutoML
• LLM-based Research Agents
• AI for Research
• Execution Grounding for Code
• 7 Conclusion
• 8 Discussion
• Acknowledgment
• Appendix A Appendix
• A.1 Other RL Attempts
• Attempt 1: Dynamic Prompt
• Attempt 2: Length Reward
• Attempt 3: Diversity Reward
• A.2 Additional Idea Examples
• A.3 Code Execution Examples
• Instructions for reporting errors