How google-deepmind/superhuman works
This repository is a collection of mathematical reasoning projects, benchmarks, and research artifacts rather than one application. The clearest documented workflow is Aletheia iteratively generating, verifying, and revising research-math solutions; the IMO benchmark datasets support evaluation of AI mathematical reasoning. The tree also contains Lean formalization and solution collections, but the available README excerpt does not describe their internal workflows, so no implementation-level links are inferred.
Main entry points: imobench IMO Bench datasets · aletheia Aletheia agent · answerbench.csv AnswerBench · proofbench.csv ProofBench
Components 10
Evaluation benchmarks
- IMO Bench datasets
imobench - AnswerBench
imobench/answerbench.csv - ProofBench
imobench/proofbench.csv - GradingBench
imobench/gradingbench.csv
Research agent
- Aletheia agent
aletheia - Research solutions
aletheia
Formal mathematics
- FutureMath formalization
FutureMath - Erdős–Hajnal proofs
FutureMath/Erdos-Hajnal/ErdosHajnal - LEAP solutions
leap/solutions
Other components
- Math researcher
Connections 7
- IMO Bench datasets → AnswerBench: includes
README.md - IMO Bench datasets → ProofBench: includes
README.md - IMO Bench datasets → GradingBench: includes
README.md - IMO Bench datasets → Math researcher: evaluates reasoning
README.md - Math researcher → Aletheia agent: uses
README.md - Aletheia agent → Research solutions: generates and revises
README.md - Aletheia agent → Aletheia agent: verifies solutions
README.md