Research map / Large language models / Agents, code and math
On Code-Induced Reasoning in LLMs
Abdul Waheed, Zhen Wu, Carolyn Rose, Daphne Ippolito
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 |
|---|---|---|---|
| finetune on code or mixed code+NL data instead of NL-only or zero-shot | LLM SFT (14 bases) · natural language tasks | accuracy: 64% | Improved |
| finetune on code or mixed data | LLM SFT (14 bases) · math and code understanding tasks | accuracy: 86% | Improved |
| finetune on code or mixed data | LLM SFT (14 bases) · code generation | judge score | Improved |
| raise the fraction of code in mixed finetuning | Qwen3 SFT · math and other tasks | accuracy | Improved |
| apply structural perturbations (whitespace removal, pseudocode, flowcharts) to code data (ablation) | code-ft · math and code tasks | accuracy | Hurt |
| apply semantic perturbations (identifiers, keywords, comments) to code data (ablation) | code-ft · math and code tasks | accuracy | Hurt |
| replace code with pseudocode or flowchart abstractions (ablation) | code-ft · non-code-generation tasks | accuracy | No clear effect |
| replace code with natural language step-by-step procedures (ablation) | code-ft · reasoning tasks | accuracy | Hurt |
| reduce token density of code data (compact encodings) (ablation) | code-ft · reasoning tasks excluding code generation | accuracy | No clear effect |
| reduce token density of code data (ablation) | code-ft · code generation | judge score | Hurt |
| use low-interpretability corrupted code with misleading signals (ablation) | code-ft · reasoning tasks | accuracy | No clear effect |
| train on lower-level languages (Java, Rust) instead of high-scripting languages (ablation) | single-language SFT · math tasks | accuracy | Improved |
Closest papers by meaning
- Which Data Attributes Stimulate Math and Code Reasoning? An Investigation via Influence Functions (NeurIPS 2025 · similarity 0.92)
- At Which Training Stage Does Code Data Help LLMs Reasoning? (ICLR 2024 · similarity 0.91)
- CodeCrash: Exposing LLM Fragility to Misleading Natural Language in Code Reasoning (NeurIPS 2025 · similarity 0.91)
- To Code or Not To Code? Exploring Impact of Code in Pre-training (ICLR 2025 · similarity 0.90)
- Unveiling the Magic of Code Reasoning through Hypothesis Decomposition and Amendment (ICLR 2025 · similarity 0.90)
- DyCodeEval: Dynamic Benchmarking of Reasoning Capabilities in Code Large Language Models Under Data Contamination (ICML 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.
