Parallel Test-Time Scaling with Multi-Sequence Verifiers

The article introduces the Multi-Sequence Verifier (MSV), a novel lightweight verifier designed to enhance parallel test-time scaling for large language models by predicting the correctness of candidates based on the entire set rather than in isolation. MSV demonstrates improved calibration, leading to a 6% relative increase in best-of-64 accuracy on mathematical reasoning benchmarks and enabling early-stopping strategies that achieve comparable accuracy with less than half the inference latency. This advancement is significant for practitioners aiming to optimize inference efficiency and selection accuracy in AI applications.