ChampionKV: Head-Wise Adaptive Budget Allocation for KV Cache Compression
We propose ChampionKV, an input-adaptive head-wise budget allocator for long-context KV cache compression.
ProofPaper ↗
Key points
- 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.
- 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.
Sources (1)
- [1]ChampionKV: Head-Wise Adaptive Budget Allocation for KV Cache CompressionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 09:19 AM
We propose ChampionKV, an input-adaptive head-wise budget allocator for long-context KV cache compression.
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.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026BELT : BEhavior-Aligned Low-Rank KV Cache Compression with Learned Top-k Rank Allocation
- Oct 13, 2026Virtual Head Attention
- Oct 13, 2026Joint Subspace Sparse Attention for Efficient Long-Context Decoding
- Oct 12, 2026MLAfication: Efficiently Enabling High-Fidelity MLA for Modern LLMs
- Oct 12, 2026[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
- Oct 11, 2026E8 Lattice Key Quantization for Low-Bit KV-Cache Compression