On November 5, 2026, researchers including Andrew Chi-Chih Yao posted TCSExam to arXiv, a benchmark covering the full range of theoretical computer science research: theorem proving, solver engineering, constructions and attempts at open conjectures. TCSExam-Full has 100 expert-reviewed tasks, each checked by the Lean kernel or by an executable task checker, so grading does not depend on a model judge.
The authors also built TCSExam-Lite, 20 tasks with five per track, which they report cuts task attempts by 80% at the same per-attempt budget. Eight agents were evaluated on Full and nine on Lite. GPT-6 Astra leads Full with 71.9, and Opus 5.5 leads Lite with 72.3. Rankings of the eight shared agents correlate strongly between the two versions (Spearman’s rho 0.93), which the authors use to support Lite as a cheaper proxy.
Why it matters: Machine-checkable grading is the most reliable way to evaluate research-level math and theory, where plausible-looking but wrong proofs are the main failure. A benchmark that spans proofs, engineering and open problems gives a broader read on AI as a theory research assistant.
What it does not show: The top model differs between Full and Lite, so the cheap version does not settle close rankings. Attempts at open conjectures are only part of the score, so a high number does not mean agents are resolving open problems.