Research map / Architectures and efficiency / Attention and transformers
Judge Q: Trainable Queries for Optimized Information Retention in KV Cache Eviction
Yijun Liu, Yixuan Wang, Yuzhuang Xu, Shiyu Ji, Yang Xu, Qingfu Zhu, Wanxiang Che
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 trainable soft-token queries instead of last-window queries for KV importance | Llama-3.1-8B · LongBench avg, budget 128 | score: 34.31 -> 35.90 (vs SnapKV) | Improved |
| use soft-token queries instead of last-window queries | Mistral-7B · LongBench avg, budget 128 | score: 39.75 -> 42.38 (vs SnapKV) | Improved |
| use soft-token queries | Llama-3.1-8B · LongBench avg, budget 512 | score: 38.31 -> 39.17 (vs SnapKV) | Improved |
| use soft-token queries | Llama-3.1-8B · RULER 8192, budget 256 | accuracy: 57.83 -> 63.13 (vs SnapKV) | Improved |
| use soft-token queries | Mistral-7B · RULER 32768, budget 256 | accuracy: 26.25 -> 33.48 (vs SnapKV) | Improved |
| use soft-token queries | Llama-3.1-8B · Critical KV hit rate | hit rate (%): 53.44 -> 61.37 (budget 128) | Improved |
| use soft-token queries | DeepSeek-R1-Distill-Llama-8B · AIME24 (budget 3072) | accuracy: 31.1 -> 37.8 | Improved |
| use soft-token queries | DeepSeek-R1-Distill-Llama-8B · MATH-500 (budget 1024) | accuracy: 52.4 -> 55.0 | Improved |
| use soft-token queries when question is moved to prompt start | Llama-3.1-8B · LongBench Single-QA | performance loss: ~10% -> <7% loss | Improved |
| train soft tokens on model-generated responses instead of dataset responses (ablation) | Llama-3.2-1B-Instruct · LongBench | score | Improved |
| train on pretraining corpora (C4, Wikipedia, GitHub) instead of ShareGPT (ablation) | Judge Q · LongBench | score | Hurt |
| set number of soft tokens to about 32 (ablation) | Judge Q · validation loss | training cost vs generalization | Improved |
Closest papers by meaning
- RazorAttention: Efficient KV Cache Compression Through Retrieval Heads (ICLR 2025 · similarity 0.92)
- Randomization Boosts KV Caching, Learning Balances Query Load: A Joint Perspective (ICLR 2026 · similarity 0.92)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.92)
- Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference (NeurIPS 2025 · similarity 0.92)
- DropKV: Decoupling Residual-Output Perturbation for Near-Optimal KV-Cache Eviction (NeurIPS 2026 · similarity 0.92)
- KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction (NeurIPS 2025 · similarity 0.92)
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.
