superhuman

Tech stack & purpose

This repository, maintained by Google DeepMind's Superhuman Reasoning team led by Thang Luong, collects projects and datasets for advanced mathematical reasoning in AI. The main components include IMO Bench, a suite of benchmarks with short-answer problems, proof-based problems, and human grading datasets designed to evaluate robust mathematical reasoning; Aletheia, a math research agent powered by Gemini Deep Think that generates, verifies, and revises solutions on research-level mathematics problems; and LEAP, an agentic framework built in Lean 4 that decomposes complex mathematical problems into subgoals and refines formal proofs using compiler feedback and LLM reviews. All software is licensed under Apache 2.0, while other materials are licensed under Creative Commons Attribution 4.0 International.

Community & reference links

• aclanthology.org/2025.emnlp-main.1794/
• Chen-Lodha
• full paper
• full paper
• full paper
• full paper
• full paper
• full paper
• arxiv.org/abs/2606.03303
• full paper
• Kirby3 problem list
• 2024 IMO-silver achievement
• 2025 IMO-gold achievement
• AxiomMath/putnam2025
• github.com/google-deepmind/alphageometry
• github.com/google-deepmind/alphageometry2
• github.com/google-deepmind/superhuman/blob/main/imobench/lean_proof_bench.csv
• imobench.github.io
• paper

Languages

Contributors

• 
  Yuri Chervonyi12 commits · 24 followers
• 
  Dawsen Hwang12 commits · 22 followers
• DeepMind2 commits
• Junsu Kim2 commits
• Po-Nien Kung1 commits
• Thang Luong1 commits