Research map / Large language models / Agents, code and math
What Really Improves Mathematical Reasoning: Structured Reasoning Signals Beyond Pure Code
Yuze Zhao, Junpeng Fang, Lu Yu, Zhenya Huang, Kai Zhang, Qing Cui, Qi Liu, JUN ZHOU, Enhong Chen
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 |
|---|---|---|---|
| include standalone pure code in pretraining mix (fixed token budget) (ablation) | 10T-token MoE · mathematical benchmarks average | relative score: -14.38% | Hurt |
| include pure code corpus (ablation) | 10T-token MoE · MATH and OlympiadBench | relative score: MATH -22.64%; OlympiadBench -47.16% | Hurt |
| include pure code corpus (ablation) | 10T-token MoE · GSM8K | relative score: +0.4% | No clear effect |
| include pure code corpus (ablation) | 10T-token MoE · PIQA and HellaSwag | relative score: -2.11%, -2.39% | Hurt |
| include pure code corpus (ablation) | 10T-token MoE · programming benchmarks | score | Improved |
| include math corpus (ablation) | 10T-token MoE · CodeForces and LiveCodeBench | relative score: CodeForces +37.11%; LiveCodeBench +11.26% | Improved |
| include math corpus (ablation) | 10T-token MoE · CruxEval, MBPP, HellaSwag | relative score: CruxEval -17.30%; MBPP -6.12%; HellaSwag -2.94% | Hurt |
| replace part of math data with structured cognitive-scaffold samples at fixed math budget | 10T-token MoE · mathematical reasoning average | relative score: +17.56% | Improved |
| add cognitive scaffolds | 10T-token MoE · OlympiadBench, MATH, College Math, MathBench | relative score: +47.78%, +23.17%, +30.05%, +14.51% | Improved |
| add cognitive scaffolds | 10T-token MoE · GSM8K and CMath | relative score: -6.29%, -2.00% | Hurt |
| add cognitive scaffolds | 10T-token MoE · programming benchmarks | score: about -1% | No clear effect |
| remove cognitive scaffolds (w/o cog) vs removing code or math (ablation) | MoE routing · expert-routing JS divergence | routing shift | No clear effect |
Closest papers by meaning
- Which Data Attributes Stimulate Math and Code Reasoning? An Investigation via Influence Functions (NeurIPS 2025 · similarity 0.93)
- Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages (ICLR 2025 · similarity 0.92)
- MathCoder2: Better Math Reasoning from Continued Pretraining on Model-translated Mathematical Code (ICLR 2025 · similarity 0.91)
- MathFimer: Enhancing Mathematical Reasoning by Expanding Reasoning Steps through Fill-in-the-Middle Task (ICLR 2026 · similarity 0.91)
- Unveiling the Magic of Code Reasoning through Hypothesis Decomposition and Amendment (ICLR 2025 · similarity 0.90)
- Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models (ICLR 2025 · similarity 0.90)
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.
