AI Paper Atlas

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

AAAI 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
use trainable soft-token queries instead of last-window queries for KV importanceLlama-3.1-8B · LongBench avg, budget 128score: 34.31 -> 35.90 (vs SnapKV)Improved
use soft-token queries instead of last-window queriesMistral-7B · LongBench avg, budget 128score: 39.75 -> 42.38 (vs SnapKV)Improved
use soft-token queriesLlama-3.1-8B · LongBench avg, budget 512score: 38.31 -> 39.17 (vs SnapKV)Improved
use soft-token queriesLlama-3.1-8B · RULER 8192, budget 256accuracy: 57.83 -> 63.13 (vs SnapKV)Improved
use soft-token queriesMistral-7B · RULER 32768, budget 256accuracy: 26.25 -> 33.48 (vs SnapKV)Improved
use soft-token queriesLlama-3.1-8B · Critical KV hit ratehit rate (%): 53.44 -> 61.37 (budget 128)Improved
use soft-token queriesDeepSeek-R1-Distill-Llama-8B · AIME24 (budget 3072)accuracy: 31.1 -> 37.8Improved
use soft-token queriesDeepSeek-R1-Distill-Llama-8B · MATH-500 (budget 1024)accuracy: 52.4 -> 55.0Improved
use soft-token queries when question is moved to prompt startLlama-3.1-8B · LongBench Single-QAperformance loss: ~10% -> <7% lossImproved
train soft tokens on model-generated responses instead of dataset responses (ablation)Llama-3.2-1B-Instruct · LongBenchscoreImproved
train on pretraining corpora (C4, Wikipedia, GitHub) instead of ShareGPT (ablation)Judge Q · LongBenchscoreHurt
set number of soft tokens to about 32 (ablation)Judge Q · validation losstraining cost vs generalizationImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy