Pessimistic Verification for Open Ended Math Questions

The article introduces a new verification paradigm called pessimistic verification for math-solving agents, which enhances error detection by rejecting solutions flagged by any of multiple parallel verifiers. It also presents progressive pessimistic verification, utilizing fine-grained proof decomposition to improve verification accuracy and efficiency, outperforming existing methods like extended long chain-of-thought workflows. This approach demonstrates significant advancements in solving complex math problems, as validated on the IMO 2025 and MathArena Apex 2025 datasets, making it relevant for practitioners seeking robust verification techniques in AI-driven math solutions.