研究显示上下文压缩会削弱大语言模型的拒答对齐，并提出 Memory Steering 干预
Does Context Compression Preserve Refusal Alignment?
AI 导读
arXiv 论文 2611.06380（Canadian AI 2026）研究上下文压缩是否保留拒答对齐，发现基于编码器的压缩在保持高重建保真度的同时，系统性削弱指令微调大语言模型的拒答行为，且该效应跨模型家族和压缩架构持续存在。
正文 · 原文
Abstract:Context compression reduces inference cost by encoding inputs into compact representations while preserving semantic content. An open question is whether semantic preservation alone is sufficient to maintain downstream behaviours such as refusal alignment. We investigate this question and find that encoder-based compression systematically weakens refusal behaviour in instruction-tuned language models, despite high reconstruction fidelity. This effect persists across model families and compression architectures. Mechanistic analysis shows that compression attenuates activation along the decoder's learned refusal direction. We further explore Memory Steering, a lightweight inference-time intervention that restores refusal rates to near-baseline levels without retraining and operates entirely in compressed representation space. These results demonstrate that semantic preservation does not guarantee behavioural preservation under compression, highlighting the need to explicitly preserve alignment-relevant features in compression-aware systems.
| Comments: | Presented at Canadian AI 2026 |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2611.06380 [cs.AI] |
| (or arXiv:2611.06380v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.06380 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Anthony Nguyen [view email]
[v1]
Wed, 4 Nov 2026 11:42:42 UTC (583 KB)
来源：arXiv cs.AI · arxiv.org