---
id: 20261017-T0-07
title: "AbridgeKV：把长推理过程压缩进学习型暂停token的KV缓存"
title_en: "AbridgeKV: Compressing Long Reasoning into KV Cache of Learned Pause Tokens"
url: https://ai.daily.yangsir.net/daily/20261017-T0-07
issue_date: 2026-10-17
publish_date: 2026-10-16T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2610.20641
---

# AbridgeKV：把长推理过程压缩进学习型暂停token的KV缓存

长链思维推理的解码过程受内存限制：生成的KV缓存随测试时计算量线性增长，简单淘汰缓存条目又会导致模型表现崩溃，因为模型从未在缺失条目的缓存上训练过。AbridgeKV提出用学习型暂停token来压缩长推理过程，将推理信息编码进KV缓存中，从而控制内存占用。该方法针对的是推理效率与内存开销之间的核心矛盾，有望让长链推理在有限显存下运行。目前该研究以预印本形式发布在arXiv上。

## English Version

**AbridgeKV: Compressing Long Reasoning into KV Cache of Learned Pause Tokens**

Long chain-of-thought decoding is memory-bound: the KV cache grows linearly with test-time compute, and simply evicting entries breaks generation since models were never trained on pruned caches. AbridgeKV proposes compressing long reasoning into the KV cache via learned pause tokens, encoding reasoning information to control memory usage. The method targets the core tension between inference efficiency and memory overhead, potentially enabling long-chain reasoning under limited VRAM. The work is available as an arXiv preprint.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2610.20641)

**详情页**：https://ai.daily.yangsir.net/daily/20261017-T0-07

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*