Research map / Large language models / Agents, code and math
AgentMath: Empowering Mathematical Reasoning for Large Language Models via Tool-Augmented Agent
Haipeng Luo, Huawen Feng, Qingfeng Sun, Can Xu, Kai Zheng, Yufei Wang, TAO YANG, Han Hu, Yansong Tang
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 |
|---|---|---|---|
| use tool-augmented synthetic data instead of text-only data in SFT (20k) (ablation) | Qwen3-8B-Base SFT · AIME24 | accuracy: 57.1% -> 60.5% | Improved |
| use tool-augmented synthetic data instead of text-only data in SFT (20k) (ablation) | Qwen3-8B-Base SFT · AIME25 | accuracy: 49.2% -> 53.3% | Improved |
| add agentic RL with code tools instead of text-based RL (ablation) | Qwen3-8B-Base · AIME24 | accuracy: 68.7% -> 76.2% | Improved |
| add agentic RL with code tools instead of text-based RL (ablation) | Qwen3-8B-Base · AIME25 | accuracy: 57.5% -> 67.5% | Improved |
| add tool augmentation in RL for training efficiency (ablation) | Qwen3-8B-Base RL · RL steps to converge | steps: ~1600 -> ~400 steps | Improved |
| add format consistency correction to synthetic data (ablation) | AgentMath SFT data (20k) · AIME24 | accuracy: 35.3% -> 47.4% | Improved |
| add code executability verification (ablation) | AgentMath SFT data (20k) · AIME24 | accuracy: 47.4% -> 52.8% | Improved |
| add environmental feedback alignment (ablation) | AgentMath SFT data (20k) · AIME24 | accuracy: 52.8% -> 56.3% | Improved |
| add self-correction capability injection (ablation) | AgentMath SFT data (20k) · AIME24 | accuracy: 56.3% -> 58.6% | Improved |
| add SFT with selective feedback masking of interpreter outputs (ablation) | AgentMath SFT data (20k) · AIME24 | accuracy: 58.6% -> 60.5% | Improved |
| add SFT with selective feedback masking of interpreter outputs (ablation) | AgentMath SFT data (20k) · AIME25 | accuracy: 50.8% -> 53.3% | Improved |
| scale tool-augmented SFT data from 2k to 300k (ablation) | AgentMath SFT · AIME24 | accuracy: 27.2% -> 78.4% | Improved |
Closest papers by meaning
- Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.90)
- A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning (ICLR 2026 · similarity 0.90)
- AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.90)
- ToRA: A Tool-Integrated Reasoning Agent for Mathematical Problem Solving (ICLR 2024 · similarity 0.90)
- THOR: Tool-Integrated Hierarchical Optimization via RL for Mathematical Reasoning (ICLR 2026 · similarity 0.90)
- T1: Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling (ICML 2025 · 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.
