(LLMs), the systems behind modern AI chatbots, were put to a new kind of test: can they advance theoretical computer science, the branch of math that defines what computers can and cannot do? Standard often miss the creative and formal work this requires.
The TCSExam benchmark evaluates models across four tracks: writing machine-checkable proofs, building efficient solver programs, constructing explicit mathematical objects, and directly attacking open research conjectures. A compact twenty-task version, TCSExam-Lite, was shown to give almost identical rankings for a fraction of the cost.
The top-scoring models produced verified proofs and programs. Several LLMs independently improved on existing mathematical bounds and provided verified special cases of open problems. This shows that LLMs can now perform verifiable research tasks in theoretical computer science, giving researchers a structured way to track genuine scientific progress.