AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

FreqKV: Key-Value Compression in Frequency Domain for Context Window Extension

Jushi Kai, Yixuan Wang, Boyi Zeng, Haoli Bai, Bo Jiang, Ziwei He, Zhouhan Lin

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression

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
keep low-frequency instead of high-frequency components in KV compression (ablation)FreqKV · LongBench GovReportscore: 14.21 -> 25.51Improved
keep low-frequency instead of high-frequency components in KV compression (ablation)FreqKV · PG-19 validation PPL, 16384perplexity: 8.70 -> 7.73Improved
compress key states before RoPE rather than after (ablation)FreqKV · PG-19 test PPL, eval 8192perplexity: 7.13 -> 7.04Improved
add the sqrt(L/N) rescaling factor after IDCT (ablation)FreqKV · training losstraining lossImproved
compress in frequency domain instead of dropping old tokens (Dropping) (ablation)FreqKV · Needle-in-a-Haystack, LLaMA-2-chat-7Bretrieval accuracyImproved
increase sink tokens from 0 to 4 (ablation)FreqKV · PG-19 validation PPL, 8192perplexity: 7.84 -> 7.76Improved
increase sink tokens from 4 to 8 (ablation)FreqKV · PG-19 validation PPL, 8192perplexity: 7.76 -> 7.75No clear effect
retain more frequency components (larger retaining ratio) (ablation)FreqKV · PG-19 validation PPLperplexity and latencyMixed
use larger retaining ratio at evaluation than in training (ablation)FreqKV · PG-19 validation PPL, 8Kperplexity: 7.76 -> 7.77 (train 0.5, eval 0.5 vs 0.75)Improved
apply FreqKV compression instead of eviction for inputs beyond the native windowLLaMA-3-instruct-8B · RULER 16K averageaverage score: 0.10 -> 47.30Improved
use FreqKV instead of full FT/LongLoRA at long evaluation lengths on Proof-pileFreqKV · Proof-pile PPL, 7B trained 32K, eval 32768perplexity: 2.50 -> 2.72Hurt
use FreqKV instead of full-attention fine-tuning for trainingLLaMA-2-7B · training cost at 8Ktraining time (hours): 28.88 -> 17.07Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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