Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P. F., Leike, J., and Lowe, R. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, 2022.Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, 2023.Gao, S., Ge, Q., Shen, W., Dou, S., Ye, J., Wang, X., Zheng, R., Zou, Y., Chen, Z., Yan, H., Zhang, Q., and Lin, D. Linear alignment: A closed-form solution for aligning human preferences without tuning and feedback. In International Conference on Machine Learning, 2024.Zhang, Z., Bai, F., Chen, Q., Ma, C., Wang, M., Sun, H., Zheng, Z., and Yang, Y. Amulet: ReAlignment during test time for personalized preference adaptation of LLMs. In International Conference on Learning Representations, 2025c.Liu, C. Y., Zeng, L., Liu, J., Yan, R., He, J., Wang, C., Yan, S., Liu, Y., and Zhou, Y. Skywork-reward: Bag of tricks for reward modeling in LLMs. CoRR, abs/2410.18451, 2024a.Beeching, E., Tunstall, L., and Rush, S. Scaling test-time compute with open models, 2025. URL https://huggingface.co/spaces/HuggingFaceH4/blogpost-scaling-test-time-compute.Zhao, S., Hong, M., Liu, Y., Hazarika, D., and Lin, K. Do LLMs recognize your preferences? evaluating personalized preference following in LLMs. In International Conference on Learning Representations, 2025.
从预训练到推理时,LLM偏好对齐的新挑战
大模型对齐这件事,最尴尬的地方在于:训练时看起来挺听话,一到真实用户面前就开始“自由发挥”。有的回答不够贴合偏好,有的明明看懂了上下文,却还是偏向默认口味。REAR 这篇论文盯住的,就是这个很现实的问题:偏好对齐不一定非要靠重新训练,推理时也能把回答重新排一遍。这事之所以重要,是因为传统对齐方法成本不低。无论是强化学习类方法,还是后续的偏好微调,都要数据、算力、流程一起上。对很多业务场景来说,最缺的不是“再训一个模型”的勇气,而是预算和时间。于是,测试时扩展(Test-Time Scaling, TTS)就成了一个很有吸引力的方向:不改权重,只在生成时多花一点计算,看看能不能把更好的回答挑出来。但问题也随之来了。TTS 在数学、代码这类“有标准答案”的任务上很好使,因为对错容易判;可一旦换到偏好对齐,情况就变复杂了。回答不是“对/错”两格按钮,而是“更符合谁的口味、更符合什么场景、更像不像人话”的综合判断。REAR 的思路,就是把这团乱麻拆开:既然模型本来就同时学了“回答问题”和“响应偏好”两部分,那测试时就重新调一调它们的比例。图1:REAR 的动机示例。模型会同时看到问题和偏好信息,再对多个候选回答进行重打分,最后把原本“不合口味”的回答换成更符合偏好的那个。先把几个缩写说清楚。RLHF 是 Reinforcement Learning from Human Feedback,中文是“基于人类反馈的强化学习”;TTS 是 Test-Time Scaling,中文是“测试时扩展”。REAR 则是 REAlignment Reward,中文可理解为“重对齐奖励”。