ChampionKV：面向 KV Cache 压缩的按头自适应预算分配

View PDF HTML (experimental)

Abstract:Some observation-window-based KV cache compressors use prompt attention to identify important tokens but allocate uniform or predefined budgets across heads, overlooking input-dependent differences in budget demand. We propose ChampionKV, an input-adaptive head-wise budget allocator for long-context KV cache compression. It uses the maximum aggregated observation-window attention score as a lightweight empirical proxy for attention concentration, assigning larger budgets to heads with dispersed attention and smaller budgets to concentrated heads. An abnormal-head mask and a floor-budget mechanism improve robustness under weak attention responses. ChampionKV plugs into compressors such as SnapKV and PyramidKV without changing their token-selection procedures. On LongBench and RULER, both Champion variants consistently improve over their SnapKV and PyramidKV base compressors; the improvements are largest and statistically significant on Llama3.1-70B-Instruct, especially under tight budgets, whereas results on the tested 7B/8B models remain competitive but mixed. ChampionKV does not uniformly outperform Ada-KV, whose global cross-head Top-K selection retains an advantage in several settings. Controlled analysis further shows improved retained attention mass over the corresponding base compressors at negligible overhead.

Subjects:	Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
Cite as:	arXiv:2610.18026 [cs.LG]
	(or arXiv:2610.18026v1 [cs.LG] for this version)
	https://doi.org/10.48550/arXiv.2610.18026 arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Yuechuan Chen [view email]
[v1] Tue, 13 Oct 2026 09:19:24 UTC (4,017 KB)