Tag
VerifierBench introduces a benchmark of 222 tasks and 4,993 hidden test cases to evaluate how well LLM agents generate executable verifiers, finding that verifiers tend to over-reject valid solutions due to hidden assumptions that survive self-testing.