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 2510.07290cs.CLcs.LG
收藏

论大语言模型中道德自我纠正的收敛性

On the Convergence of Moral Self-Correction in Large Language Models

Michigan State University(密歇根州立大学);Amazon(亚马逊);Pennsylvania State University(宾夕法尼亚州立大学);University of California, Santa Barbara(加州大学圣塔芭芭拉分校)

Guangliang Liu, Haitao Mao, Bochuan Cao, Zhiyu Xue, Xitong Zhang, Rongrong Wang, Kristen Marie Johnson

AI总结:

本文聚焦大语言模型的道德自我纠正,揭示其通过多轮交互实现性能收敛的关键特征,并机制性分析表明持续指令激活道德概念降低不确定性导致收敛,展示其强大潜力。

摘要:

大语言模型(LLMs)在被指示时能够改进其响应,这种能力被称为自我纠正。当指令仅提供一般且抽象的目标,而未包含响应中潜在问题的具体细节时,LLMs必须依赖其内部知识来提高响应质量,这一过程被称为内在自我纠正。内在自我纠正的经验成功在各种应用中显而易见,但其如何有效以及为何有效仍属未知。聚焦于LLMs中的道德自我纠正,我们揭示了内在自我纠正的一个关键特征:通过多轮交互实现性能收敛;并对这一收敛行为提供了机制性分析。基于我们的实验结果和分析,我们揭示了收敛的潜在机制:持续注入的自我纠正指令激活了道德概念,这些概念降低了模型的不确定性,随着激活的道德概念在连续轮次中趋于稳定,性能随之收敛。本文通过展示道德自我纠正具有性能收敛这一理想属性,证明了其强大的潜力。

英文摘要:

Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only a general and abstract goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. Focusing on moral self-correction in LLMs, we reveal a key characteristic of intrinsic self-correction: performance convergence through multi-round interactions; and provide a mechanistic analysis of this convergence behavior. Based on our experimental results and analysis, we uncover the underlying mechanism of convergence: consistently injected self-correction instructions activate moral concepts that reduce model uncertainty, leading to converged performance as the activated moral concepts stabilize over successive rounds. This paper demonstrates the strong potential of moral self-correction by showing that it exhibits a desirable property of converged performance.

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。

↑