事實脈絡
arXiv:2608.07852v1 Announce Type: new Abstract: How a language model internally represents who is speaking, the Assistant, an assigned roleplay persona, or a narrated story character, remains underexplored. We study spe
解讀與影響
大型语言模型在对话中能自如切换身份，时而作为中立的AI助理，时而扮演特定的虚构角色，但其内部如何区分和表征这些“自我”与“他者”，一直是个黑箱。来自arXiv上的一项最新研究，通过稀疏自编码器（Sparse Autoencoders）技术，对这一问题进行了深入的解剖。
研究团队聚焦于模型激活空间中的“方向”（directions），即特定的神经元激活模式。他们发现，模型内部并非存在一个单一的“助理”概念，而是根据上下文动态地激活不同的子空间。例如，“通用AI助理”的身份表征与“扮演莎士比亚戏剧角色”时的身份表征，在几何结构上既有重叠，也存在显著分离。这些表征可以被稀疏自编码器分解为一组可解释的特征，每个特征对应着身份的一个侧面，如“正式的语气”、“第一人称视角”或“虚构世界观的知识”。
进一步的分析显示，这些身份表征具有跨层级的连贯性。在模型的浅层，表征更多与语言风格和代词使用相关；而在深层，则与角色一致性的逻辑推理和背景知识调用紧密关联。当模型被要求进行角色扮演时，相关的“角色特征”会被显著激活，同时抑制默认的“助理特征”，这种动态的激活与抑制机制确保了角色切换的稳定性和一致性。该研究为理解语言模型如何构建和操纵复杂的社会性概念提供了新的视角，也为后续提升模型的可控性与安全性奠定了理论基础。
參考來源
來源原文
arXiv:2608.07852v1 Announce Type: new Abstract: How a language model internally represents who is speaking, the Assistant, an assigned roleplay persona, or a narrated story character, remains underexplored. We study speaker representations using a dataset of user-expressed emotional text and corresponding model responses. We decompose three generation settings (Assistant, Roleplay, and Story) into sparse autoencoder features extracted at turn-boundary and pronoun-token positions and selected through a filtering pipeline for different depths. We characterize each surviving feature through its steering effects and activation distribution. Our main finding is that the Assistant and roleplay personas are not independent alternatives: personas retain the Assistant-associated feature core while progressively differentiating from it across layers, starting from operational machinery towards behavioral and stylistic features. Meanwhile, generated story characters lack the Assistant-associated core. Both Story and Roleplay can be distinguished from the Assistant with Immersive Simulation Mode. However, the Assistant can sometimes enter or slowly drift into it even in the default setting.