LPBench: A Lean Benchmark for Research-Level Mathematical Theorem Proving
In October 2026, OpenAI released hundreds of manuscripts resolving open mathematical problems, without publishing the full set of problems attempted by its model. Assessing the broader capabilities and limitations of such systems calls for independent benchmarks. To this end, we introduce LPBench, a benchmark for research-level mathematics comprising 117 problems formalised in Lean 4, primarily in combinatorics and graph theory. As a first application, we evaluate Claude Opus 5.5 and Claude Fable 5.1, each with and without web search. Website: https://lp-bench.github.io/
Authors
- Sven Groppe (ORCID: https://orcid.org/0000-0001-5196-1117)
- Johannes Carmesin (ORCID: https://orcid.org/0000-0002-3026-6673)
- Jinghua Groppe (ORCID: https://orcid.org/0009-0009-0295-7029)
- Utku Okur (ORCID: https://orcid.org/0000-0002-9196-1127)
- Nathan Bowler
- Vaibhav Bajpai
Institutions
- Universität Hamburg (DE)
- TU Bergakademie Freiberg (DE)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24195211
- Primary Topic
- Mathematics, Computing, and Information Processing
- Type
- preprint