GitHub repository: https://github.com/mzhaoshuai/RefAlign
This repository contains the SFT (Supervised Fine-Tuning) model mzhaoshuai/zephyr-7b-alpha-conf-sft, which is an integral part of the RefAlign framework. This model serves as an initial SFT step for Confidence Alignment experiments, trained with shuchangtao/CONQORD_dataset (specifically, conqord_step1_data), as described in the accompanying research.
Large language models~(LLMs) are expected to be helpful, harmless, and honest. In different alignment scenarios, such as safety, confidence, and general preference alignment, binary preference data collection and reward modeling are resource-intensive but play a central role in transferring human preferences. In this work, we explore using the similarity between sampled generations and reference answers as a supplementary reward function for alignment. When unary reference answers are available, such similarity-based rewards can circumvent the need for binary preference data and explicit reward modeling. We introduce \textit{RefAlign}, a versatile REINFORCE-style alignment algorithm that does not rely on reward or reference models. RefAlign utilizes language generation evaluation metrics, such as BERTScore, between sampled generations and reference answers as surrogate rewards. Beyond general preference optimization, RefAlign can be naturally extended to diverse scenarios, including safety and confidence alignment, by combining similarity-based rewards with task-specific objectives. Across multiple scenarios, RefAlign achieves performance comparable to prior alignment methods while operating without binary preference data or reward models.
@article{zhao2025learning,
title={Learning from reference answers: Versatile language model alignment without binary human preference data},
author={Zhao, Shuai and Xu, Yunqiu and Zhu, Linchao and Yang, Yi},
journal={arXiv preprint arXiv:2504.09895},
year={2025}
}