TCSExam: GPT-6 Astra proves 14 known theorems but independently solves no open conjecture

Tsinghua University, Moonshot AI and the Shanghai Qizhi Institute released TCSExam, 100 expert-reviewed tasks spanning Lean theorem proving, solver engineering, constructions and 30 open conjectures. GPT-6 Astra completed 14 of 20 research theorems without supplied lemmas and led the Full index at 71.9. No model independently resolved an open conjecture; the few accepted conjecture submissions adapted counterexamples or proofs OpenAI had already published. Lean-kernel and executable checks make outcomes reproducible, but a benchmark should not equate skillful reconstruction of known knowledge with an original research breakthrough.