arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2402.09345cs.LGcs.AI
收藏

InfoRM:通过信息论奖励建模缓解RLHF中的奖励破解问题

InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling

Wuhan University(武汉大学);The University of Sydney(悉尼大学);Fudan University(复旦大学);Nanyang Technological University(南洋理工大学)

Yuchun Miao, Sen Zhang, Liang Ding, Rong Bao, Lefei Zhang, Dacheng Tao

AI总结:

针对RLHF中因奖励模型依赖虚假特征导致的奖励破解问题,提出信息论奖励建模框架InfoRM,结合变分信息瓶颈过滤无关信息,并设计CSI指标检测过度优化,在多规模模型和多数据集上验证了有效性与鲁棒性。

摘要:

尽管基于人类反馈的强化学习(RLHF)在使语言模型对齐人类价值观方面已取得成功,但奖励破解(又称奖励过度优化)仍是一项关键挑战。该问题主要源于奖励泛化错误,即奖励模型(RM)利用与人类偏好无关的虚假特征计算奖励。本研究从信息论视角切入这一问题,提出了一种名为InfoRM的奖励建模框架,通过引入变分信息瓶颈目标来过滤无关信息。值得注意的是,我们进一步发现过度优化与InfoRM的IB潜在空间中的异常值存在相关性,这使得该框架成为检测奖励过度优化的有效工具。受这一发现启发,我们提出了聚类分离指数(CSI),用于量化IB潜在空间中的偏差,将其作为奖励过度优化的指标,以助力在线缓解策略的开发。我们在多种设置和不同RM规模(70M、440M、1.4B和7B)上开展了大量实验,证明了InfoRM的有效性。进一步分析表明,InfoRM的过度优化检测机制不仅有效,而且在广泛的数据集上均具有鲁棒性,标志着RLHF领域的一项显著进展。代码将在论文录用后发布。

英文摘要:

Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models with human values, reward hacking, also termed reward overoptimization, remains a critical challenge. This issue primarily arises from reward misgeneralization, where reward models (RMs) compute reward using spurious features that are irrelevant to human preferences. In this work, we tackle this problem from an information-theoretic perspective and propose a framework for reward modeling, namely InfoRM, by introducing a variational information bottleneck objective to filter out irrelevant information. Notably, we further identify a correlation between overoptimization and outliers in the IB latent space of InfoRM, establishing it as a promising tool for detecting reward overoptimization. Inspired by this finding, we propose the Cluster Separation Index (CSI), which quantifies deviations in the IB latent space, as an indicator of reward overoptimization to facilitate the development of online mitigation strategies. Extensive experiments on a wide range of settings and RM scales (70M, 440M, 1.4B, and 7B) demonstrate the effectiveness of InfoRM. Further analyses reveal that InfoRM's overoptimization detection mechanism is not only effective but also robust across a broad range of datasets, signifying a notable advancement in the field of RLHF. The code will be released upon acceptance.

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。

↑