AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

ManifoldKV: Training-Free KV Cache Compression via Euclidean Outlier Detection

Debajyoti Datta, Trishala Neeraj, Bibek Paudel, Vyom Sharma, Subhabrata Mukherjee

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 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
replace cosine scoring with L2 distance from centroid (ablation)KeyDiff-style scorer · RULER standaloneaccuracy (%): 52.8% -> 92.7%Improved
replace L2 with L1 distance (ablation)ManifoldKV · RULER standaloneaccuracy (%): 92.7% -> 78.5%Hurt
replace L2 with L-infinity distance (ablation)ManifoldKV · RULER standaloneaccuracy (%): 92.7% -> 71.2%Hurt
use local sliding-window centroids (4K) instead of global centroid (ablation)ManifoldKV · RULER 64Kaccuracy (%): 35.2% -> 84.3%Improved
shrink window from 16K to 4K (ablation)WindowedManifoldKV · RULER 64Kaccuracy (%): 82.4% -> 84.3%Improved
shrink window from 4K to 2K (ablation)WindowedManifoldKV · RULER 64Kaccuracy (%): 84.3% -> 83.8%Hurt
L2 scoring instead of cosine (KeyDiff) on 3-key retrievalManifoldKV · RULER multikey_3 (8K, 50% compression)accuracy (%): 77.0 -> 92.4Improved
L2 scoring instead of cosine (KeyDiff) on 2-key retrievalManifoldKV · RULER multikey_2 (8K, 50% compression)accuracy (%): 92.6 -> 99.8Improved
L2 scoring instead of cosine under AdaKV on overall RULERManifoldKV · RULER (4K-16K, AdaKV)accuracy (%): 95.66% -> 95.73%No clear effect
geometric scorer instead of attention-based SnapKV under AdaKVManifoldKV · RULER (4K-16K, AdaKV)accuracy (%): 83.97% -> 95.73%Improved
blend L2 with cosine (Hybrid 0.5) instead of windowingManifoldKV · RULER 64Kaccuracy (%): 81.09% -> 79.14%Hurt
increase context length from 4K to 64K with global centroidGlobal ManifoldKV · RULER (Llama-3.1-8B)accuracy (%): 95.7% -> 35.2%Hurt

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