Research map / Large language models / Reasoning and chain of thought
Learning Math Reasoning from Self-Sampled Correct and Partially-Correct Solutions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Learn from self-sampled fully-correct and partially-correct solutions instead of only the reference solution | GPT-Neo models · GSM5.5K-Python | pass@100: 35.5% to 44.5% | Helped |
| Learn from self-sampled fully-correct and partially-correct solutions instead of only the reference solution | GPT-Neo models · MathQA-Python-Filtered | pass@80: 27.6% to 36.2% | Helped |
| Add partially-correct solution prefixes to self-sampled fully-correct solutions (ablation) | Self-sampling method · GSM5.5K-Python | pass@100: +3.0% | Helped |
| Add partially-correct solution prefixes to self-sampled fully-correct solutions (ablation) | GPT-Neo 125M · MathQA-Python-Filtered | pass@k | No clear effect |
| Use MLE-Aug rather than MML for learning from fully-correct and partially-correct targets (ablation) | GPT-Neo 125M · GSM5.5K-Python dev | pass@100: 18.7 -> 35.0 | Helped |
| Use MLE-Aug rather than β-MML for learning from fully-correct and partially-correct targets (ablation) | GPT-Neo 125M · GSM5.5K-Python dev | pass@100: 27.9 -> 35.0 | Helped |
| Use MLE-Aug rather than MLE for learning from the single reference solution (ablation) | GPT-Neo 125M · GSM5.5K-Python dev | pass@100: 22.7 -> 32.3 | Helped |
| Use MML with FCS + PCS rather than single-reference MLE (ablation) | GPT-Neo 125M · GSM5.5K-Python dev | pass@100: 22.7 -> 18.7 | Hurt |
| Learn from multiple sampled solutions with self-sampling | GPT-Neo models · MathQA-Python-Filtered and GSM5.5K-Python | pass@1 | No clear effect |
| Train with self-sampled solutions rather than only a single reference | GPT-Neo 125M · GSM5.5K-Python | unique solutions among 100 samples: 20.8% -> 30.5% | Helped |
Closest papers by meaning
- Physics of Language Models: Part 2.2, How to Learn From Mistakes on Grade-School Math Problems (ICLR 2025 · similarity 0.92)
- MathCoder2: Better Math Reasoning from Continued Pretraining on Model-translated Mathematical Code (ICLR 2025 · similarity 0.91)
- Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages (ICLR 2025 · similarity 0.91)
- MARGE: Improving Math Reasoning with Guided Exploration (ICML 2025 · similarity 0.90)
- MathFimer: Enhancing Mathematical Reasoning by Expanding Reasoning Steps through Fill-in-the-Middle Task (ICLR 2026 · similarity 0.89)
- NFT: Bridging Supervised Learning and Reinforcement Learning in Math Reasoning (ICLR 2026 · 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.