Research map / Architectures and efficiency / Attention and transformers
Statistical Advantage of Softmax Attention: Insights from Single-Location Regression
O Duranthon, Pierre Marion, Claire Boyer, Bruno Loureiro, Lenka Zdeborová
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 |
|---|---|---|---|
| use softmax instead of linear attention (ablation) | single-layer attention on single-location regression · Spiked-SLR and max-SLR, population risk | minimal population risk | Improved |
| use softmax instead of linear attention at finite samples (ablation) | single-layer attention on single-location regression · Spiked-SLR and max-SLR, finite sample ratio, L=3 | minimal test risk | Improved |
| vary sequence length (L uniform on {1,2,3}) instead of fixed L=2 (ablation) | linear attention · Single-location regression, population risk | minimal population risk | Hurt |
| use element-wise erf activation instead of a normalized activation (ablation) | attention activation · Single-location regression with variable L | minimal population risk | Hurt |
| use softplus kernelized attention instead of softmax (ablation) | attention activation · Single-location regression, larger L | minimal population risk | Hurt |
| use erf or softplus activation instead of linear (ablation) | attention activation · Spiked-SLR and max-SLR, population risk | minimal population risk | Improved |
| add a constant bias term to the activation (1 + chi instead of chi) (ablation) | linear and erf attention · Single-location regression, gradient at initialization | nonzero gradient on recovery order parameters | Improved |
Closest papers by meaning
- Attention layers provably solve single-location regression (ICLR 2025 · similarity 0.90)
- Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free (NeurIPS 2025 · similarity 0.89)
- From Attention to Activation: Unraveling the Enigmas of Large Language Models (ICLR 2025 · similarity 0.88)
- The Closeness of In-Context Learning and Weight Shifting for Softmax Regression (NeurIPS 2024 · similarity 0.88)
- Selective Attention: Enhancing Transformer through Principled Context Control (NeurIPS 2024 · similarity 0.88)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.88)
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.
