Self-Consistency Preference Optimization

TL;DR

This paper introduces Self-Consistency Preference Optimization (ScPO), a novel training method that enhances model reasoning accuracy by iteratively training models to prefer consistent answers, significantly improving performance on reasoning benchmarks.

Contribution

The paper extends the self-consistency concept from inference to training, proposing ScPO to improve reasoning tasks without supervised labels, achieving results comparable to supervised methods.

Findings

ScPO improves reasoning accuracy on GSM8K and MATH benchmarks.

Combining ScPO with supervised learning yields further performance gains.

ScPO fine-tunes Llama-3 8B to outperform larger models like Llama-3 70B.

Abstract

Self-alignment, whereby models learn to improve themselves without human annotation, is a rapidly growing research area. However, existing techniques often fail to improve complex reasoning tasks due to the difficulty of assigning correct rewards. An orthogonal approach that is known to improve correctness is self-consistency, a method applied at inference time based on multiple sampling in order to find the most consistent answer. In this work, we extend the self-consistency concept to help train models. We thus introduce self-consistency preference optimization (ScPO), which iteratively trains consistent answers to be preferred over inconsistent ones on unsupervised new problems. We show ScPO leads to large improvements over conventional reward model training on reasoning tasks such as GSM8K and MATH, closing the gap with supervised training with gold answers or preferences, and that combining ScPO with standard supervised learning improves results even further. On ZebraLogic, ScPO finetunes Llama-3 8B to be superior to Llama-3 70B, Gemma-2 27B, and Claude-3 Haiku.