method · first seen 27 Oct 2026
attention matching
Compacting a KV cache by selecting keys, refitting values, and adding logit biases to preserve attention outputs.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Folding Attention Biases into Keys Makes Compacted KV Caches Run on Standard KernelsPapers · 27 Oct 2026 · Utkarsh Kanwat
Latest sources
- Folding Attention Biases into Keys Makes Compacted KV Caches Run on Standard KernelsPapers · 27 Oct 2026 · Utkarsh Kanwat