CompilerKV: Risk-Adaptive KV Compression via Offline Experience Compilation
Ning Yang, Chengzhi Wang, Yibo Liu, Baoliang Tian, Haijun Zhang · 2026
Read CompilerKV: Risk-Adaptive KV Compression via Offline Experience Compilation by Ning Yang, Chengzhi Wang, Yibo Liu, Baoliang Tian, Haijun Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-08686.txt.
Description
Prefill-only KV compression picks tokens once and never revisits the choice, yet it estimates its correction signals from a single noisy prompt. Ning Yang and four co-authors argue that is the wrong statistical unit: reliability and sensitivity patterns are far steadier across prompts than within one. Their arXiv paper introduces CompilerKV, which compiles correction tables offline from a calibration corpus.
Online correction then shrinks to constant-time lookups plus a budget clamp. Compiled tables transfer across disjoint corpora and even across models at small cost. At a 512-token budget the method leads compressed baselines on all four backbones tested, and it stays strongest under extreme cache pressure. Readers get a portable prior for KV retention that costs almost nothing at serving time.
Catalog record
| CREATOR | Ning Yang, Chengzhi Wang, Yibo Liu, Baoliang Tian, Haijun Zhang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 58 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-08686