Effectiveness of inference-time scaling with neural verifiers on advanced mathematics

Determine whether inference-time scaling of natural-language mathematical large language models by combining search with neural verifiers to mitigate hallucinated reasoning yields effective performance on advanced mathematical problems beyond the pre-college or AIME level.

Background

The paper contrasts the traditional informal approach to AI for mathematics with a complementary strategy that scales computation at inference time, potentially integrating search with neural verifiers to reduce hallucinations. While such methods have shown promise on high-school level tasks, the authors note that their efficacy on substantially more advanced problems remains unproven.

This open question frames the limits of inference-time scaling for informal math LLMs and motivates exploring formal mathematical reasoning as a more robust alternative for rigorous evaluation and feedback.