SqueezeBench: Every Math Pass Raises the Level of the Next Test

Ten combinatorics problems are used to improve verified Lean bounds. A solution becomes a new record to be broken instead of removing the problem.

SqueezeBench seeks to avoid a mathematical benchmark becoming too easy when models progress. Each problem retains a lower terminal and an upper terminal of an entire function; candidates must provide improvement and verifiable evidence.

Two nuclei Independent Lean controls the evidence. The authors publish ten problems with 20 initial records. The agents evaluated improve these records, but often remain behind the best asymptotic limits of the literature. The device thus measures a certified progression without confounding success to the test and novel mathematical discovery.

Pre-publication filed on arXiv on 9 October 2026; The results are those reported by the authors and remain to be confirmed independently.