Researchers have developed a new framework called T to evaluate the semantic correctness of theorems generated by large language models in automated theorem proving. This approach, inspired by code generation testing, verifies theorems by checking if dependent successor theorems compile successfully. Experiments using T on real-world Lean 4 repositories revealed that while current models like Claude-Sonnet-4.5 can compile generated theorems, their semantic accuracy is significantly lower, highlighting a gap in their theorem generation capabilities. AI
IMPACT Introduces a novel semantic evaluation metric for LLM-generated theorems, revealing significant performance gaps in current models.
RANK_REASON The cluster describes a new evaluation framework and benchmark for AI in automated theorem proving, presented in an arXiv paper.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
