Research map / Architectures and efficiency / Attention and transformers
Multipole Semantic Attention: A Fast Approximation of Softmax Attention for Pretraining
Rupert Mitchell, Kristian Kersting
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add query clustering to the monopole approximation (ablation) | MuSe monopole K128 · 1B attention outputs | relative squared error: 0.1016 -> 0.0227 | Improved |
| add query clustering to retrieval selection (ablation) | MuSe retrieval K128R8 · 1B attention outputs | relative squared error: 0.2475 -> 0.2429 | No clear effect |
| add retrieval instead of dipole correction to the monopole (ablation) | MuSe monopole K128 · 1B attention outputs | relative squared error: 0.0227 -> 0.00622 | Improved |
| add untilted dipole corrections to the monopole (ablation) | MuSe monopole K128 · 1B attention outputs | relative squared error: 0.0227 -> 0.0150 | Improved |
| use MuSe instead of MoBA at matched sparsity | MuSe · Attention microbenchmark (1B) | relative squared error: 0.248 -> 0.006 | Improved |
| use MuSe instead of MoBA at matched sparsity | MuSe · Code pretraining 185M | cross-entropy loss: 0.9714 -> 0.9384 | Improved |
| train with MuSe instead of exact attention (test with MuSe) | MuSe · 1B code pretraining | cross-entropy loss: 0.7026 -> 0.7001 | Improved |
| 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 -> 0.6994 | Improved |
| make tilted aggregation weights non-differentiable (ablation) | MuSe · 1B code pretraining (test with CUDNN) | cross-entropy loss: 0.7108 -> 0.6991 | Improved |
| remove exponential tilting by query centroids on pretrained Llama (ablation) | MuSe · Llama 3.2-1B and 3.1-8B on PG-19 | cumulative loss | Hurt |
| train with MuSe instead of CUDNN Flash Attention | MuSe · 1B pretraining throughput (8 A100, 64k) | throughput (KTok/s): 65.3 -> 88.9 | Improved |
| train with MuSe instead of exact attention | MuSe · Code NIAH multi-key 4x | uid exact match (%): 81.50 -> 97.36 | Improved |
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.
- MoBA — MoBA: Mixture of Block Attention for Long-Context LLMs (NeurIPS 2025) · better on 3 of 4 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
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (ICLR 2025 · similarity 0.88)
- Multipole Attention for Efficient Long Context Reasoning (NeurIPS 2025 · similarity 0.87)
- A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts (ICLR 2025 · similarity 0.87)
- Training-Free Long-Context Scaling of Large Language Models (ICML 2024 · similarity 0.87)
- Virtual Head Attention (NeurIPS 2026 · similarity 0.87)
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · similarity 0.86)
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.
