Benchmark RadarAI BENCHMARK PROFILE
TheoremBench
Evaluates LLMs on theorem proving in Lean4 using classical theorems, with two versions: main and premised. Includes metrics for theorem-level coverage and token efficiency to assess partial progress and proof structure.
- Released
- 2026-06-08
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Provides a more realistic evaluation of provers beyond contest problems, revealing biases toward easy subtheorems and inefficient proof strategies. Supports finer-grained analysis of formal reasoning capabilities.
Motivation
LLMs have recently achieved strong results on formal proving benchmarks.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.