AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Multi-Head Low-Rank Attention

Songtao Liu, Hongwu Peng, Zhiwei Zhang, Zhengyu Chen, Yue Guo

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 3 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 MLRA-4 instead of MLAMLA · FineWeb-Edu-100B pretraining, 7-dataset avgperplexity: 13.727 -> 13.672Improved
use MLRA-4 instead of GQAGQA · 7-dataset avgperplexity: 14.139 -> 13.672Improved
use MLRA-4 instead of MLAMLA · zero-shot common-sense avgaccuracy: 58.75 -> 58.84Improved
use 4 branches (MLRA-4) instead of 2 (MLRA-2) (ablation)MLRA · 7-dataset avgperplexity: 13.804 -> 13.672Improved
add gating before attention output projection (ablation)MLRA-4 · 7-dataset avgperplexity: 13.672 -> 13.621Improved
add gating before attention output projection (ablation)GQA · 7-dataset avgperplexity: 14.139 -> 13.806Improved
use zero initialization for output projections instead of N(0, 0.02) (ablation)all attention variants · pretraining loss and perplexitylossImproved
add variance scaling of latent states (ablation)MLA, GLA-2, MLRA-2 · pretrainingperplexityImproved
double the number of attention heads at fixed KV cache (ablation)GQA, MLA, GLA-2 · pretrainingperplexityNo clear effect
use MLRA-4 with TP=4 kernel instead of MLA decodingMLA · long-context decoding latencyspeedup: about 2.8x over MLAImproved
use MLRA-4 kernel instead of GQAGQA · long-context decoding 128K-2Mspeedup: 1.05x-1.26xImproved

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