arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2610.14958cs.AIcs.CLcs.LG
收藏

当激活引导被用作记忆时,什么会留存下来

What Survives When Activation Steering Is Used as Memory

Delft University of Technology(代尔夫特理工大学)

Vedant Jumle

AI总结:

本文测试了激活引导作为记忆的假设,提出一种无需训练的激活记忆,在Qwen3.5模型上的预注册研究显示该记忆具有选择性,但事实存储使用率低,倾向仅在窄窗口转移,偏好参考决定留存内容。

摘要:

激活引导最近被用作技能的记忆:智能体的过往经验被存储为引导向量,后续可被检索。研究假设用户的事实和偏好需要文本形式存储,本文在冻结的大语言模型(LLM)之外,测试了一种无需训练的激活记忆。该记忆将经验在后期层状态中产生的变化存储为偏移量,以对话主旨的白化键归档,仅当主旨匹配时才在后续会话中重新加入。本文认为,这种偏移量是注意力针对每个问题在上下文中重新计算的效应的固定平均值,因此只有与问题无关的内容才能留存。在一项预注册研究中,使用Qwen3.5-2B和9B模型,评判者来自四个模型家族,该记忆具有选择性(约98%的无关回复未改变),但事实仅作为要说出的词存储在最后几层,几乎从未被使用(干净使用率最多15%,而上下文使用率为96-98%)。倾向仅在狭窄的强度窗口中转移。最引人注目的是,编写偏好时减去的参考决定了其概念或方向是否留存:除非减去相反内容,否则存储的厌恶会被回忆为喜欢。

英文摘要:

Activation steering has recently been used as a memory for skills: past experiences of an agent are stored as steering vectors and retrieved later. Facts and preferences about the user are assumed to need text instead. We test that assumption with a training-free activation memory beside a frozen LLM. The change an experience causes in late-layer states is stored as a shift, filed under a whitened key for the conversation's gist, and added back in a later session only when the gist matches. We argue that such a shift is a fixed average of an effect that attention, in context, recomputes for every question, so only question-independent content can survive. In a pre-registered study on Qwen3.5-2B and 9B, with judges from four model families, the memory is selective (about 98% of unrelated replies are unchanged), but facts are stored as the word to say, in the last layers only, and are almost never used (at most 15% clean use, against 96-98% in context). Leanings transfer only in a narrow strength window. Most strikingly, the reference subtracted when writing a preference decides whether its concept or its direction survives: unless the opposite is subtracted, a stored dislike is recalled as a like.

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。

↑