Code2Math: Can Your Code Agent Effectively Evolve Math Problems Through Exploration?

2026-03-03 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, quick

Summary

A new multi-agent framework, Code2Math, investigates the potential of code agents to autonomously evolve existing math problems into more complex variations. This research addresses the scarcity of challenging, high-quality math problems needed for training and evaluating advanced large language models (LLMs) aiming for International Mathematical Olympiad (IMO) level capabilities. The framework is designed to perform problem evolution while simultaneously validating the solvability and increased difficulty of the generated problems. Experiments show that with sufficient test-time exploration, code agents can synthesize novel, solvable problems that are structurally distinct and more challenging than their original counterparts, demonstrating code-driven agents as a viable mechanism for generating high-difficulty mathematical reasoning problems.

Key takeaway

For research scientists developing advanced mathematical LLMs, the Code2Math framework offers a promising approach to overcome the bottleneck of scarce, high-quality training and evaluation problems. You should consider integrating code-driven agentic problem evolution into your data generation pipelines to create a continuous supply of challenging and structurally distinct mathematical reasoning problems, thereby enhancing model training and evaluation rigor.

Key insights

Code agents can autonomously evolve math problems, creating more complex and solvable variations for LLM training.

Principles

• Code execution offers a scalable environment for mathematical experimentation.
• Problem evolution requires validation of solvability and increased difficulty.

Method

A multi-agent framework is introduced to evolve math problems, validating solvability and increased difficulty through test-time exploration to synthesize new, challenging problems.

In practice

• Generate diverse math problems for LLM training.
• Evaluate LLM math capabilities with evolved problems.

Topics

• Code Agents
• Mathematical Problem Generation
• Large Language Models
• Multi-Agent Systems
• Problem Evolution

Code references

• TarferSoul/Code2Math

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• ParEVO: Synthesizing Code for Irregular Data: High-Performance Parallelism through Agentic Evolution — ParEVO synthesizes high-performance parallel code for irregular data using an evolutionary agent and specialized LLMs.
• LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks — Agentic frameworks like LEAP enable LLMs to achieve state-of-the-art formal theorem proving by bridging informal reasoning with verifiable proofs.
• Solvita: Enhancing Large Language Models for Competitive Programming via Agentic Evolution — Solvita enables LLMs to continuously learn and accumulate problem-solving experience in competitive programming via agentic evolution.
• Self-Evolving Deep Research via Joint Generation and Evaluation — Co-evolving LLM generation and evaluation within a shared-parameter framework overcomes static evaluator limitations in deep research.
• Programming with LLM Agents in 2025 — LLM agents, like Open Hands, significantly amplify developer productivity by automating iterative coding and visualization tasks.
• Controlled Self-Evolution for Algorithmic Code Optimization — Controlled Self-Evolution significantly boosts code optimization efficiency by guiding LLM-based iterative refinement.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.