Research map / Multi-agent systems and games / LLM agents
Achieving Expert-Level Agent from Foundation Model via Complexity Curriculum Reinforcement Learning with Synthetic Data
Haiteng Zhao, Junhao Shen, Yiming Zhang, Songyang Gao, Kuikun Liu, Tianyou Ma, Fan Zheng, Dahua Lin, Wenwei Zhang, Kai Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a complexity-curriculum reinforcement learning approach with synthetic data to train LLM agents for olympiad geometry problem solving, aiming to match expert systems like AlphaGeometry 2.
Problem
LLM agents struggle with geometry problem solving because of weak heuristics for auxiliary constructions, so the field relies on expert models such as AlphaGeometry 2 that depend on large-scale data synthesis and search.
Approach
Uses a complexity curriculum within reinforcement learning on synthetic geometry data to train a foundation-model agent to produce auxiliary constructions, instead of relying on large-scale search-based expert systems.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use InternGeometry LLM agent instead of expert models | AlphaGeometry 2 · IMO 50 pass@256 | problems solved: 42 -> 44 | Improved |
| add proposition proving steps (ablation) | InternGeometry · IMO 50 pass@256 | problems solved: 35 -> 44 | Improved |
| add slow thinking (ablation) | InternGeometry · IMO 50 pass@256 | problems solved: 23 -> 35 | Improved |
| add context compression (dynamic memory) (ablation) | InternGeometry · IMO 50 pass@256 | problems solved: 20 -> 44 | Improved |
| add rejection sampling against repeated actions (ablation) | InternGeometry · IMO 50 pass@256 | problems solved: 38 -> 44 | Improved |
| add CBRL after SFT cold start (ablation) | InternGeometry · IMO 50 pass@256 | problems solved: 22 -> 44 | Improved |
| use only easy synthetic data in RL (ablation) | InternGeometry · IMO 50 pass@256 | problems solved: 44 -> 29 | Hurt |
| use only challenging data in RL (ablation) | InternGeometry · IMO 50 pass@256 | problems solved: 44 -> 24 | Hurt |
| add complexity schedule over synthesized data (ablation) | InternGeometry · IMO 50 pass@256 | problems solved: 38 -> 44 | Improved |
| lengthen agent trajectory instead of more sampling (ablation) | InternGeometry · IMO 50 pass@256 | pass@K | Improved |
Experimental setup
Baselines: AlphaGeometry 2
Builds on: AlphaGeometry 2
Closest papers by meaning
- Position: We Need An Algorithmic Understanding of Generative AI (ICML 2025 · similarity 0.86)
- MathSmith: Towards Extremely Hard Mathematical Reasoning by Forging Synthetic Problems with a Reinforced Policy (AAAI 2026 · similarity 0.86)
- Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles (NeurIPS 2025 · similarity 0.85)
- MM-Agent: LLM as Agents for Real-world Mathematical Modeling Problem (NeurIPS 2025 · similarity 0.85)
- Verifier-Backed Hard Problem Generation for Mathematical Reasoning (NeurIPS 2026 · similarity 0.85)
- Assessing the Creativity of LLMs in Proposing Novel Solutions to Mathematical Problems (AAAI 2025 · similarity 0.85)
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.
