TCSExam tests frontier AI on 100 verifiable research tasks
The TCSExam paper evaluates frontier AI models on 100 theoretical computer science research tasks spanning theorem proving, solver engineering, mathematical construction, and open conjectures, with Lean 4 and program checkers used to verify submissions. Reported results include a program-checked 630-edge graph from Kimi K3 and a Lean-verified nine-dimensional partial proof from GPT-6 Astra, while several models also matched the graph result. No reference model independently solved an open conjecture; six accepted submissions covered two problems but relied on previously published counterexamples and proofs found online.
TCSExam provides a machine-checkable way to assess whether frontier models can perform research-level theoretical computer science work, while showing that verified partial progress still falls short of independently solving open problems.
导语:Kimi 刷新 629 条边图论参考纪录,只多加了1条边,为什么数学界等了这么多年?

Kimi 刷新 629 条边图论参考纪录,只多加了1条边,为什么数学界等了这么多年?
作者丨郑佳美
编辑丨岑 峰
给 124 个点连线,不能出现三角形,也不能出现四边形。此前的参考记录是 629 条——数学界认为这已经接近极限。而Kimi K3 构造出了 630 条边的图,并通过程序验证。
#TCSExam#Kimi K3#Kimi K3.1 Pro#GPT-6 Astra#DeepSeek V4.1 Flash#Opus 5.5#GLM-5.3#Qwen3.8 Max#Grok 4.6#Lean 4#Moonshot AI#Tsinghua University