Research map / Large language models / Agents, code and math
Theoretical Modeling of Large Language Model Self-Improvement Training Dynamics Through Solver-Verifier Gap
Yifan Sun, Yushan Liang, Zhen Zhang, Xin Liu, Jiaye Teng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add external DeepSeek-V3 data (Early) to self-improvement | Phi-4-mini self-improvement · Math (train, QE) | solver accuracy (%): 45.08 -> 46.33 | Improved |
| add external DeepSeek-V3 data (Early) to self-improvement | Phi-4-mini self-improvement · GSM8k (train, QE) | solver accuracy (%): 88.53 -> 88.54 | No clear effect |
| add external DeepSeek-V3 data (Early) to self-improvement | Llama-3.2-3B self-improvement · Math (train, QE) | solver accuracy (%): 49.16 -> 53.48 | Improved |
| add external DeepSeek-V3 data (Early) to self-improvement | Llama-3.2-3B self-improvement · GSM8k (train, QE) | solver accuracy (%): 88.66 -> 88.26 | No clear effect |
| average cross-improvement over baseline | Phi-4-mini / Llama-3.2-3B · GSM8k (train, QE) | solver accuracy (%): +0.10 (Phi-4-mini), -0.77 (Llama-3.2-3B) | Mixed |
| average cross-improvement over baseline | Llama-3.2-3B self-improvement · Math (train, QE) | solver accuracy (%): +3.35 | Improved |
| vary external-data timing: Early vs Late (ablation) | Llama-3.2-3B cross-improvement · Math (train, QE) | solver accuracy (%): 53.48 vs 51.31 | Mixed |
| vary external-data timing: Early vs Uniform vs Late (ablation) | Phi-4-mini cross-improvement · Math (train, QE) | solver accuracy (%): Max-Min 0.73 | No clear effect |
| vary external-data timing: Early vs Uniform vs Late (ablation) | Phi-4-mini cross-improvement · GSM8k (train, QE) | solver accuracy (%): Max-Min 0.46 | No clear effect |
| vary external-data timing: Early vs Uniform vs Late (ablation) | Llama-3.2-3B cross-improvement · GSM8k (train, QE) | solver accuracy (%): Max-Min 1.07 | No clear effect |
| early external data raises initial verifier accuracy (ablation) | Phi-4-mini · Math (train, QE) | solver accuracy (%): +5.87 | Improved |
| early external data raises initial verifier accuracy (ablation) | Llama-3.2-3B · GSM8k (train, QE) | solver accuracy (%): -1.58 | Hurt |
Closest papers by meaning
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models (ICLR 2025 · similarity 0.95)
- Self-Improvement in Language Models: The Sharpening Mechanism (ICLR 2025 · similarity 0.93)
- Progress or Regress? Self-Improvement Reversal in Post-training (ICLR 2025 · similarity 0.92)
- Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing (NeurIPS 2024 · similarity 0.90)
- Latent Principle Discovery for Language Model Self-Improvement (NeurIPS 2025 · similarity 0.89)
- Importance Weighting Can Help Large Language Models Self-Improve (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
