Math Olympiad Agent: Self-Verification for IMO Problems

Sep 10, 2025

This project implements an AI agent for solving International Mathematical Olympiad (IMO)-level problems. It adapts the self-verification pipeline from the paper “Gemini 2.5 Pro Capable of Winning Gold at IMO 2025” (Yichen Huang and Lin F. Yang) and uses a multi-stage iterative process with automated verification. Built with LangGraph for workflow orchestration, the agent supports the A2A (Agent-to-Agent) protocol for multi-agent collaboration and is extensible to different LLM backends.

Highlights

Cutting‑Edge Inspiration: Direct adaptation of the Gemini 2.5 Pro IMO pipeline (5/6 problems solved in the 2025 evaluation).
Rigorous Proofs: Produces logically sound solutions rather than just final answers.
Extensible Design: Compatible with multiple LLMs (Gemini, OpenAI-compatible, etc.) and supports the A2A protocol.
Open-Source Reimplementation: Faithfully recreates the paper’s method while leaving room for experimentation.

Math Olympiad Agent: Pipeline

You can see the full project on my GitHub: here.