Mathematical Reasoning

ICML 2025 Mathematical Reasoning Track

International Conference on Machine Learning 2025 featured a dedicated track on mathematical reasoning, with advances in neural theorem proving and problem solving.

ICML 2025 AI4Math Highlights

Dedicated Track: Mathematical Reasoning

First year with a dedicated track for mathematical reasoning research.

Key Papers

Neural Theorem Proving

  • Policy gradient methods for tactic prediction
  • Transformer architectures for proof representation
  • Graph neural networks for premise selection

Problem Solving

  • Olympiad-level geometry with neural-symbolic methods
  • Automated inequality proving
  • Combinatorial search with neural guidance

Training Methodologies

  • Curriculum learning for mathematical reasoning
  • Synthetic data generation for math problems
  • Reinforcement learning from human proofs

Tutorials

  • Introduction to Neural Theorem Proving: Lean and Coq
  • Mathematical Reasoning in LLMs: From GPT-4 to specialized models

Competitions

  • IMO-Grand Challenge: Progress report and new benchmarks
  • miniF2F: Neural theorem proving competition

Impact

ICML 2025 marked mainstream acceptance of AI4Math as a core ML research area.