Self-Consistency Preference Optimization
This addresses the challenge of assigning correct rewards in self-alignment for AI models, particularly in reasoning tasks, with incremental advancements over existing methods.
The paper tackled the problem of improving complex reasoning tasks in self-alignment by introducing self-consistency preference optimization (ScPO), which trains models to prefer consistent answers over inconsistent ones on unsupervised problems, resulting in large improvements on benchmarks like GSM8K and MATH, and enabling an 8B model to outperform larger models on ZebraLogic.
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.