ARTFEED — Contemporary Art Intelligence

CAM-Bench: New Benchmark for Computational Math in Lean

other · 2026-05-20

CAM-Bench introduces a new set of 1,000 theorem-proving challenges in Lean 4, concentrating on applied mathematics and computational topics such as numerical analysis, optimization, and numerical linear algebra. This benchmark fills a void left by current assessments that primarily focus on Olympiad-style challenges and algebraic areas. The problems are derived from textbook exercises and rely on specific definitions, notation, algorithms, and fundamental results introduced locally. A dependency-recovery pipeline reconstructs the context of the textbook to accurately present each problem, subsequently normalizing it into an informal theorem and converting it into a Lean target. This benchmark facilitates the mechanical verification of mathematical reasoning within large language models.

Key facts

  • CAM-Bench contains 1,000 Lean 4 proof targets.
  • Covers optimization, numerical linear algebra, and numerical analysis.
  • Problems adapted from textbook exercises.
  • Uses a dependency-recovery pipeline to reconstruct local textbook context.
  • Each problem is normalized into a standalone informal theorem.
  • Translated into a Lean target for formal verification.
  • Addresses underrepresentation of computational and applied mathematics in existing benchmarks.
  • Enables mechanically verifiable evaluation of LLM mathematical reasoning.

Entities

Institutions

  • arXiv

Sources