PaperAgent
Nov 9, 2026 · Artificial Intelligence
TCSExam Reveals LLMs' Differentiated Research Skills: Theorem Proving Strong, Autonomy Weak
The TCSExam benchmark evaluates frontier LLMs on four theoretical computer science research tasks—theorem proving, solver engineering, mathematical construction, and open conjectures—using verifiable Lean 4 proofs and executable checkers; GPT-6 Astra leads at 70% theorem-proving success, Kimi K3.1 Pro follows at 38%, but models show divergent strengths across tasks and still require human-provided lemmas for complex proofs.
AI researchLLM benchmarkLean 4
0 likes · 10 min read