---
title: "CASK"
description: "KV cache compression for reasoning LLMs · arXiv"
canonical: https://portfolio.skyline23.com/en/projects/cask
language: en
content-signal: ai-train=no, search=yes, ai-input=yes
---

# CASK

> KV cache compression for reasoning LLMs · arXiv

## Overview

CASK studies KV cache compression during long LLM reasoning. It keeps answer-critical states and merges repeated intermediate states. Published on arXiv on April 13, 2026 (arXiv:2604.10900).

## Highlights

- Preserves a protected core and merges scratch states; prefix eviction makes room for decode-stage compression.
- Compares Qwen3-8B with TriAttention on H100 PCIe using six examples each from AIME24 and AIME25; records next-token agreement and mean NLL.
- Records gains and limits separately: multi_news and vcsum exercise decode-stage merging, while qmsum and gov_report exhaust the prefix budget. The paper and implementation are public.

## Metadata

- Type: Project
- Category: ML
- Years: 2026
- Stack: LLM, KV cache, Reasoning

## Links

- Paper: [arXiv](https://arxiv.org/abs/2604.10900)
- Source: [GitHub](https://github.com/Skyline-23/CASK)

## Language versions

- [EN](https://portfolio.skyline23.com/en/projects/cask/index.md) · [KO](https://portfolio.skyline23.com/ko/projects/cask/index.md) · [JA](https://portfolio.skyline23.com/ja/projects/cask/index.md)

- [Portfolio index](https://portfolio.skyline23.com/en/index.md)