- Who
- TCSExam research team
- What
- introduced a benchmark of 100 theoretical computer science research tasks for evaluating AI agents
- Quote
为回答这个问题，他们提出 TCSExam，用 100 项理论计算机科学科研任务，测试前沿 AI Agent 能否证明定理、开发 Solver、构造数学对象，甚至挑战长期未解决的猜想。
What happened
A joint research team introduced TCSExam, a benchmark of 100 theoretical computer science tasks that tests AI agents on theorem proving, solver development, mathematical constructions, and open conjectures. The framework uses Lean 4, program checkers, and checkpoints to verify results and distinguish partial progress from completed proofs. The article also reports that GPT-6 Astra extended a result for the Alon–Jaeger–Tarsi conjecture from 8 to 9 dimensions over a five-element finite field, while the general conjecture remains unresolved.
Why it matters
TCSExam separates partial progress from complete results through independently checkable certificates, while the experiments show where current agents still struggle with proof planning and generalization.
Read next
See the full storyResearchers Introduce TCSExam to Test AI on 100 Theoretical Computer Science TasksThe story so far
A joint research team introduced TCSExam, a benchmark of 100 theoretical computer science tasks for evaluating AI agents. It tests theorem proving, solver development, mathematical constructions, and attempts at open conjectures, using Lean 4, program checkers, and checkpoints to verify results and distinguish partial progress from completed proofs. The report says GPT-6 Astra extended a result for the Alon–Jaeger–Tarsi conjecture from 8 to 9 dimensions over a five-element finite field. The general conjecture remains unresolved.