Research map / Architectures and efficiency / Attention and transformers
Fast KV Compaction via Attention Matching
Adam Zweiger, Xinghong Fu, Han Guo, Yoon Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AM-HighestAttnKeys instead of KVzip at 1% cache | Qwen3-4B · QASPER | token-level F1: 0.127 -> 0.338 | Improved |
| use AM-HighestAttnKeys instead of H2O+ at 1% cache | Llama-3.1-8B-Instruct · QASPER | token-level F1: 0.199 -> 0.317 | Improved |
| use AM-HighestAttnKeys instead of KVzip at 10% cache | Llama-3.1-8B-Instruct · QASPER | token-level F1: 0.347 -> 0.452 | Improved |
| use AM-HighestAttnKeys instead of H2O+ at 1% cache | Qwen3-4B-Instruct-2507 · LongBench v2 subset | accuracy (%): 10.6 -> 61.7 | Improved |
| use AM-HighestAttnKeys instead of KVzip at 20% cache | Qwen3-4B · QASPER | token-level F1: 0.438 -> 0.435 | Mixed |
| apply AM-OMP on top of a summary (0.2x) | Summarize · QuALITY | accuracy (%): 55.2 -> 55.7 | No clear effect |
| apply AM-OMP on top of a summary (0.05x) | Summarize · QuALITY | accuracy (%): 55.2 -> 49.2 | Hurt |
| use OMP-fast (k=4, refit interval 2) instead of OMP (ablation) | AM-OMP · LongHealth Gemma-3-12B | key selection time (s): 565 -> 104 | Improved |
| use highest-attention key selection instead of OMP (ablation) | AM · LongHealth Gemma-3-12B | key selection time (s): 565 -> 3 | Improved |
| use context-prefill instead of repeat-prefill queries (ablation) | AM · reference query generation | time (s): 8 -> 7 | Improved |
| add L2 regularization when fitting beta and C_v (ablation) | AM · least squares fit | performance | Hurt |
| add post-compaction attention biases and fitted values (leave-one-out) (ablation) | AM-OMP · reference generations | log-perplexity | Improved |
Closest papers by meaning
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.90)
- SALS: Sparse Attention in Latent Space for KV Cache Compression (NeurIPS 2025 · similarity 0.90)
- Training Free Exponential Context Extension via Cascading KV Cache (ICLR 2025 · similarity 0.90)
- QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024 · similarity 0.89)
- Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs (ICLR 2024 · similarity 0.89)
- Homogeneous Keys, Heterogeneous Values: Exploiting Local KV Cache Asymmetry for Long-Context LLMs (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
