AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Multipole Semantic Attention: A Fast Approximation of Softmax Attention for Pretraining

Rupert Mitchell, Kristian Kersting

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Linear attention approximations

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 query clustering to the monopole approximation (ablation)MuSe monopole K128 · 1B attention outputsrelative squared error: 0.1016 -> 0.0227Improved
add query clustering to retrieval selection (ablation)MuSe retrieval K128R8 · 1B attention outputsrelative squared error: 0.2475 -> 0.2429No clear effect
add retrieval instead of dipole correction to the monopole (ablation)MuSe monopole K128 · 1B attention outputsrelative squared error: 0.0227 -> 0.00622Improved
add untilted dipole corrections to the monopole (ablation)MuSe monopole K128 · 1B attention outputsrelative squared error: 0.0227 -> 0.0150Improved
use MuSe instead of MoBA at matched sparsityMuSe · Attention microbenchmark (1B)relative squared error: 0.248 -> 0.006Improved
use MuSe instead of MoBA at matched sparsityMuSe · Code pretraining 185Mcross-entropy loss: 0.9714 -> 0.9384Improved
train with MuSe instead of exact attention (test with MuSe)MuSe · 1B code pretrainingcross-entropy loss: 0.7026 -> 0.7001Improved
add brief CUDNN fine-tuning (0.1% tokens) to MuSe-trained model (ablation)MuSe · 1B code pretraining (test with CUDNN)cross-entropy loss: 0.7108 -> 0.6994Improved
make tilted aggregation weights non-differentiable (ablation)MuSe · 1B code pretraining (test with CUDNN)cross-entropy loss: 0.7108 -> 0.6991Improved
remove exponential tilting by query centroids on pretrained Llama (ablation)MuSe · Llama 3.2-1B and 3.1-8B on PG-19cumulative lossHurt
train with MuSe instead of CUDNN Flash AttentionMuSe · 1B pretraining throughput (8 A100, 64k)throughput (KTok/s): 65.3 -> 88.9Improved
train with MuSe instead of exact attentionMuSe · Code NIAH multi-key 4xuid exact match (%): 81.50 -> 97.36Improved

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