---
title: "Does Context Compression Preserve Refusal Alignment?"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2611.06380v1"
permalink: "https://feed.astralor.com/items/itm_ae51207ad6744ca01d6bf03d"
published: "2026-11-04T11:42:42.000Z"
score: 57
channel: "tech"
tags: ["推理效率","安全","Does Context Compression","Preserve Refusal Alignment","Context","Mechanistic","Memory Steering","These"]
---

# Does Context Compression Preserve Refusal Alignment?

## 速览

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 …

## 关键点

- 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.

## 全文

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.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_ae51207ad6744ca01d6bf03d