LADDER: Self-Improving LLMs Through Recursive Problem Decomposition

arxiv:2503.00735

Original source · Paper JSON

Post-training · 2025-03-02

Cites (3) · Cited by (1)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

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

Contents

• LADDER: Self-Improving LLMs Through Recursive Problem Decomposition
• Abstract
• 1 Introduction
• 2 Related Work
• Self-Improvement and Automated Curriculum Generation in LLMs.
• Test-Time Compute Scaling and Reasoning Improvements.
• Reinforcement Learning Approaches for LLM Self-Improvement.
• 3 Methodology
• 3.1 Algorithm Design (LADDER & TTRL)
• 3.1.1 LADDER
• 3.1.2 Variant Generation
• 3.1.3 Solution Verification
• 3.1.4 Reinforcement Learning Protocol
• 3.1.5 Test-Time Reinforcement Learning
• 3.2 Experiments
• 3.2.1 Llama 3B Experiments
• 3.2.2 MIT Integration Bee (LADDER)
• 3.2.3 MIT Integration Bee (LADDER + TTRL)
• 4 Results
• 4.1 Llama 3B Experiments
• 4.2 MIT integration Bee (LADDER)
• 4.3 MIT Integration Bee (LADDER + TTRL)
• 5 Discussion
• 5.1 A Form of Test-Time Compute Scaling
• 5.2 Extension to Other Verifiable Domains
• 5.3 Future Work
• 6 Conclusion
• Appendix: MIT Integration Bee Failed Questions
• Instructions for reporting errors