IMProofBench

Informal Mathematical Proof Benchmark

IMProofBench evaluates the ability of AI systems to create research-level mathematical proofs. We maintain a curated, private repository of PhD-level problems across pure mathematics to measure genuine mathematical reasoning capabilities while preventing data contamination and benchmark overfitting.

Model Performance

Top models by percentage of benchmark questions with complete and correct solutions.

Benchmark Results
1
GPT-5.4 46.4%
2
Gemini 3.1 Pro Preview 45.7%
3
GPT-5.2 Pro (web search) 39.0%
4
Claude Opus 4.6 33.6%
46.4%
GPT-5.4
45.7%
Gemini 3.1 Pro Preview
39.0%
GPT-5.2 Pro (web search)
33.6%
Claude Opus 4.6

Become a Contributor!

Create a question and see what state-of-the-art models can do in your field of mathematics. If your question is included in the benchmark, receive co-authorship on future papers. You retain full rights to your question and can retract it at any time.

Create a Question

Community

Connect with the project team and mathematical researchers.

About

Learn about the benchmark's goals and methodology.

Key Features

AI Model Testing
Test problems against frontier models with immediate feedback
Peer Review System
Expert review ensures problem quality and appropriate difficulty
Automated Grading
Subquestions enable objective evaluation alongside proof assessment
Privacy Preservation
Majority private dataset prevents overfitting and gaming