Code2Math is a code-agent framework for synthesizing harder mathematical problems through executable exploration. Starting from seed problems, an evolution agent searches for deeper variants with Python tools, while two verifier agents filter candidates for mathematical solvability and meaningful difficulty increase.
This repository contains the public Code2Math data, prompt templates, and two full code-agent pipeline implementations.
- Code released: full code-agent pipelines are now included under
code2math/. - Data released: seed problems, evolved problem sets, demonstrations, and prompt templates are available in this repository.
- Paper:Code2Math: Can Your Code Agent Effectively Evolve Math Problems Through Exploration?
Dadi Guo*, Yuejin Xie*, Qingyu Liu*, Weixian Huang*, Jiayu Liu, Zhiyuan Fan, Qihan Ren, Shuai Shao, Tianyi Zhou, Jianjie Feng, Wenze Su, Yujiu Yang, Dongrui Liu (corresponding author), Yi R. (May) Fung (corresponding author)
*Equal contribution.
Affiliations: Hong Kong University of Science and Technology, Tsinghua University, Zhejiang University, Nanjing Tech University, Shanghai Jiao Tong University, University of Michigan, Independent Researcher.
Code2Math/
code2math/ # reusable pipeline package
scripts/ # runnable entry points
prompts/ # prompt templates
math_demonstrations/ # few-shot evolution demonstrations
evolved_problems/ # released evolved problem sets
original_problems.json # 100 seed problems
Code2Math decomposes problem evolution into three agents:
- Evolution Agent - proposes harder variants of seed problems through thought, empirical inquiry, and Python-backed exploration.
- Solvability Verification Agent - rejects ill-formed, inconsistent, or unsolvable candidates.
- Difficulty Verification Agent - checks whether the evolved problem increases conceptual discovery burden rather than only adding computation.
The code tools support symbolic computation, graph algorithms, combinatorial enumeration, numerical checks, and constraint-style exploration.
We release two complete implementations of the three-stage workflow.
| Backend | Entry Point | What It Does |
|---|---|---|
| Standard CodeAgent | scripts/run_code_agent_pipeline.py | Uses the standard smolagents.CodeAgent execution loop. |
| Interleaved Thinking | scripts/run_interleaved_pipeline.py | Preserves OpenAI-style messages, tool_calls, tool responses, and returned reasoning content in messages.json. |
python scripts/run_code_agent_pipeline.py --max-workers 5python scripts/run_interleaved_pipeline.py --max-workers 5The interleaved backend is implemented as a small compatibility layer in this repository, so users do not need to install a patched smolagents fork.
The implementation builds on and follows the design of Hugging Face smolagents, especially its CodeAgent, tool abstraction, Python execution tool, and OpenAI-compatible model wrappers. Code2Math adds the math-evolution prompts, three-agent orchestration, output parsing, released datasets, and an interleaved-thinking backend for preserving tool-call trajectories from thinking-model runs.
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .envEdit .env with an OpenAI-compatible endpoint:
URL=https://your-openai-compatible-endpoint/v1KEY=your_api_keyEVOLVE_MODEL=your-evolver-modelVERIFY_MODEL=your-verifier-modelQwen-style Basic-auth endpoints are also supported through API_AK, API_SK, QWEN_THINKING_URL, and QWEN_INSTRUCT_URL.
By default, generated results are written under:
adapted_problems/
codeagent/
interleaved/
Agent logs are written under logs/. These generated outputs are ignored by git.
Each result entry records the pipeline status, failure stage if any, evolved problem, verifier outputs, and final difficulty judgment.
Seed problems:
{
"problem_description": "...",
"solution_steps": "...",
"answer": "..."
}Evolved problem records:
{
"status": "success",
"result_data": {
"status": true,
"new_problem": {
"new_problem": "...",
"new_solution_steps": "...",
"new_answer": "..."
},
"solvability_verifier_output": {},
"difficulty_verifier_output": {}
}
}This release focuses on the two full code-agent workflows. The later single-turn no-code pipeline used in internal experiments is intentionally not included.
Please do not commit .env, logs, or generated run outputs.
If you use Code2Math data, prompts, or code, please cite:
@misc{guo2026code2mathcodeagenteffectively,
title={Code2Math: Can Your Code Agent Effectively Evolve Math Problems Through Exploration?},
author={Dadi Guo and Yuejin Xie and Qingyu Liu and Weixian Huang and Jiayu Liu and Zhiyuan Fan and Qihan Ren and Shuai Shao and Tianyi Zhou and Jianjie Feng and Wenze Su and Yujiu Yang and Dongrui Liu and Yi R. Fung},
year={2026},
eprint={2603.03202},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2603.03202},
}MIT License. See LICENSE.