Pessimistic Verification: Helping LLMs Check Mathematical Proofs
For mathematical agents, solving problems is only half of the story. They also need to check whether a proof is actually correct. Verification is central to iterative problem-solving workflows and to reinforcement learning for open-ended mathematical reasoning. Our ICML 2026 paper, Pessimistic Verification for Open-Ended Math Questions, studies a simple principle: when verifying a mathematical proof, the bottleneck is usually error detection. Instead of asking multiple reviewers to vote, we ask them to look for flaws. If any reviewer finds a critical error, the proof should be rejected. ...