AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

DesireKV: Decoupling Sensitivity and Importance for Reasoning-Aware KV Cache Compression

Pengyu Cheng, Jiacheng Wang, Tianle Chen, Bei Liu, Xiaofeng Hou, Jiacheng Liu

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: KV cache compression · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DesireKV, a KV cache compression method for chain-of-thought reasoning that decouples token sensitivity from importance to account for repetitive, low-density reasoning sequences.

Problem

KV cache memory grows large when LLMs produce long chain-of-thought reasoning, and existing KV cache compression methods are suboptimal because reasoning sequences have repetitive logic and low information density.

Approach

Decouples sensitivity and importance of tokens to perform reasoning-aware KV cache compression, rather than applying conventional text-generation compression criteria to reasoning traces.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add outlier-based sensitivity scoring for quantization (vs random 4/8-bit) (ablation)DesireKV · AIME2024Pass@1: 0.4292 -> 0.4625Improved
add attention-based importance for eviction (vs random eviction) (ablation)DesireKV · AIME2024Pass@1: 0.4375 -> 0.4625Improved
add confidence-based reasoning protection (ablation)DesireKV · AIME2024Pass@1: 0.4583 -> 0.4625Improved
add confidence-based reasoning protection (ablation)DesireKV · GPQA-DiamondPass@1: 0.4595 -> 0.4684Improved
add sensitivity scoring (ablation)DesireKV · GSM8KPass@1: 0.8614 -> 0.8811Improved
protect top outlier-score tokens instead of random tokens (20% BF16, rest 4-bit) (ablation)KV quantization · AIME2024Pass@1 (%): 37.9 -> 44.6Improved
replace lowest-attention protected tokens with low-confidence tokens (ablation)attention-driven protection · AIME2024Pass@1 (%): 42.9 -> 44.2Improved
quantize low-importance tokens (DesireKV) instead of evicting them (RPC)RPC · AIME2024Pass@1: 0.4250 -> 0.5343Improved
dual importance-sensitivity criteria vs uniform 2-bit quantization of pruned tokensDDKS · AIME2024Pass@1: 0.3375 -> 0.4625Improved
DesireKV vs K8V4 on average bit widthKIVI K8V4 · four benchmarks, Llama-8BAve Bit: 6.0 -> 2.9Improved
DesireKV vs K8V4 on MATH-500 accuracyKIVI K8V4 · MATH-500, Llama-8BPass@1: 0.8595 -> 0.8595No clear effect
DesireKV vs K4V4 on GSM8K, Qwen-7BKIVI K4V4 · GSM8K, Qwen-7BPass@1: 0.8795 -> 0.8797No clear effect

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