arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2404.04656cs.LGcs.AIcs.CL
收藏

面向大语言模型对齐的二元分类器优化

Binary Classifier Optimization for Large Language Model Alignment

Kakao Corp(Kakao公司);LBOX

Seungjae Jung, Gunsoo Han, Daniel Wontae Nam, Kyoung-Woon On

AI总结:

本文提出 BCO,仅利用点赞/点踩等二元反馈训练分类器并以 logit 作为隐式奖励对齐 LLM,通过奖励偏移逼近 DPO 损失,在多模型多数据集上取得稳健且可比的效果。

摘要:

在 ChatGPT 等真实服务中,基于用户反馈对齐模型对于提升模型性能至关重要。然而,由于提供反馈简单方便,用户通常只给出基本的二元信号,例如“点赞”或“点踩”。另一方面,大多数现有对齐研究依赖基于偏好的方法,这类方法需要将正面和负面回复组成一对。我们提出 Binary Classifier Optimization(BCO,二元分类器优化),这是一种仅使用二元反馈即可有效对齐 LLM 的技术。BCO 训练一个二元分类器,并将其 logit 作为隐式奖励,从而有效最小化 Direct Preference Optimization(DPO,直接偏好优化)损失。我们证明,分类器训练中采用的二元交叉熵损失是 DPO 损失的上界。此外,一种新颖的奖励偏移技术进一步缩小了两种损失之间的差距。我们在两种设置下验证了该方法:首先,在成对偏好数据集上,我们的方法表现与 DPO 相当;其次,在源自真实用户查询的 Likert-5 量表标注数据集上进行验证。我们的模型在四个基础 LLM 和三个不同数据集上始终展现出有效且稳健的对齐效果,显示出该方法从二元信号中学习的优势。

英文摘要:

In real-world services such as ChatGPT, aligning models based on user feedback is crucial for improving model performance. However, due to the simplicity and convenience of providing feedback, users typically offer only basic binary signals, such as 'thumbs-up' or 'thumbs-down'. Most existing alignment research, on the other hand, relies on preference-based approaches that require both positive and negative responses as a pair. We propose Binary Classifier Optimization (BCO), a technique that effectively aligns LLMs using only binary feedback. BCO trains a binary classifier, where the logit serves as an implicit reward, effectively minimizing the Direct Preference Optimization (DPO) loss. We demonstrate that the binary cross-entropy loss employed in classifier training acts as an upper bound for the DPO loss. Additionally, a novel reward shift technique further minimizes the gap between the losses. We validate our methodology in two settings: first, on a paired preference dataset, where our method performs on par with DPO; and second, on a Likert-5 scale annotation dataset which stems from real users' queries. Our model consistently demonstrates effective and robust alignment across four base LLMs and three different datasets, showcasing the strength of our approach to learning from binary signals.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑