AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

ParisKV: Fast and Drift-Robust KV-Cache Retrieval for Long-Context LLMs

Yanlin Qi, Xinhang Chen, Huiqiang Jiang, Qitong Wang, Botao Peng, Themis Palpanas

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Long-context KV and memory attention · 0 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
add normalization, rotation and theoretical spherical centroids (ablation)ParisKV coarse stage · AIME long decodingcoarse Recall@100 (%): 6 -> 16.1Improved
add normalization, rotation and theoretical spherical centroids (ablation)ParisKV · AIME long decodingfinal Recall@100 after reranking (%): 36.5 -> 64.3Improved
use ParisKV instead of PQCacheQwen3-4B · GPQA-Diamondaccuracy: 38.38 -> 72.22Improved
use ParisKV instead of PQCacheQwen3-4B · MATH500accuracy: 58.80 -> 92.80Improved
use ParisKV instead of PQCacheQwen3-4B · AIME25pass@8: 3.33 -> 80.00Improved
use ParisKV instead of full attentionQwen3-8B · MATH500accuracy: 87.40 -> 93.00Improved
use ParisKV instead of full attentionQwen3-8B · AIME25pass@8: 83.33 -> 73.33Hurt
use ParisKV instead of full attentionQwen3-4B · AIME25pass@8: 86.67 -> 80.00Hurt
use ParisKV instead of PQCache and MagicPIGDeepSeek-R1-Llama-8B · LongBench-V2 overallscore: 19.90 -> 28.43Improved
use ParisKV instead of ShadowKVSparse retrieval · RULER 128Kaverage score: 80.29 -> 83.49Improved
use ParisKV instead of full attention for decodingLlama3.1-8B and Qwen3-8B · 64K-256K decoding throughputpeak throughput: 2.1x to 2.8xImproved
use ParisKV instead of PQCacheLlama3.1-8B · 1024K decode latencyms/step: 2179 -> 49Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 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