arXiv：cs.AI· Vedant Jumle·· 16 小时前
激活引导作为记忆时能留下什么：Qwen3.5 上的预注册研究
What Survives When Activation Steering Is Used as Memory
AI 导读
研究在冻结 LLM 旁以无训练激活记忆测试技能记忆假设，在 Qwen3.5-2B 和 9B 上开展预注册实验，由四个模型家族担任评判。结果显示记忆具选择性（约 98% 无关回复不变），但事实仅在最后几层以“该说的词”存储，干净使用率至多 15%，而上下文内为 96-98%；偏好只在狭窄强度窗口内迁移，且写入偏好时减去的参照决定其概念还是方向存活，除非减去相反项，存储的厌恶会被回忆成喜欢。
正文
Abstract: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.
| Comments: | 12 pages, 7 figures, 2 tables. Code: this https URL |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.14958 [cs.AI] |
| (or arXiv:2610.14958v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.14958 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Vedant Jumle [view email]
[v1]
Sat, 10 Oct 2026 14:31:47 UTC (115 KB)
来源：arXiv：cs.AI · arxiv.org