Andrew Yao, Yang Zhilin and others release the TCSExam benchmark: none of 30 open conjectures independently cracked by AI
姚期智杨植麟首次联手发论文!30道数学世界难题无一AI独立攻破
Authors including Andrew Yao and Yang Zhilin released the theoretical computer science benchmark TCSExam on arXiv; among 100 research-level problems, GPT-6 Astra ranks first with 71.9 points and Kimi K3.1 Pro second with 57.4, but none of the 30 open conjectures unsolved by humans was independently cracked by any model. The problems fall into four categories: formal theorem proving, solver engineering, mathematical construction and open conjectures, all scored by whether the code passes the Lean 4 formalization kernel and automated verification tools. The paper says two models scored sporadically on open conjectures, but in the process searched online for counterexamples and proofs OpenAI had previously made public, so this does not count as independent solving. On specific constructions the models still gave results beyond the reference: Kimi K3 connected 630 edges under the 124-vertex condition (the reference construction has 629), and raised the minimum Hamming distance of an error-correcting code from 5 to 6; GPT-6 Astra tightened a reference upper bound on the number of codewords of a binary code from 84 to 82.