arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2502.06258cs.CLcs.LG
收藏

大语言模型中的涌现式响应规划

Emergent Response Planning in LLMs

Shanghai Artificial Intelligence Laboratory(上海人工智能实验室)

Zhichen Dong, Zhanhui Zhou, Zhixuan Liu, Chao Yang, Chaochao Lu

AI总结:

本文发现大语言模型虽仅训练预测下一词元,但其隐藏表示编码了未来响应的结构、内容与行为属性,表现出涌现式规划能力,并随模型规模与生成过程变化,可用于提升透明度和生成控制。

摘要:

在这项工作中,我们论证了大型语言模型(LLMs)虽然仅被训练用于预测下一个词元,却展现出涌现式的规划行为:$\textbf{它们的隐藏表示编码了超出下一个词元的未来输出}$。通过简单的探针实验,我们证明LLM的提示表示编码了其整个响应的全局属性,包括$\textit{结构属性}$(例如,响应长度、推理步骤)、$\textit{内容属性}$(例如,故事写作中的角色选择、响应末尾的多选题答案)以及$\textit{行为属性}$(例如,答案置信度、事实一致性)。除了识别响应规划之外,我们还探讨了它如何随模型大小在不同任务中扩展,以及它在生成过程中如何演变。LLMs在其隐藏表示中为未来进行前瞻性规划的发现,暗示了在提高透明度和生成控制方面的潜在应用。

英文摘要:

In this work, we argue that large language models (LLMs), though trained to predict only the next token, exhibit emergent planning behaviors: $\textbf{their hidden representations encode future outputs beyond the next token}$. Through simple probing, we demonstrate that LLM prompt representations encode global attributes of their entire responses, including $\textit{structure attributes}$ (e.g., response length, reasoning steps), $\textit{content attributes}$ (e.g., character choices in storywriting, multiple-choice answers at the end of response), and $\textit{behavior attributes}$ (e.g., answer confidence, factual consistency). In addition to identifying response planning, we explore how it scales with model size across tasks and how it evolves during generation. The findings that LLMs plan ahead for the future in their hidden representations suggest potential applications for improving transparency and generation control.

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。

↑