Research map / Architectures and efficiency / Attention and transformers
Understanding axial attention in TSFMs through single location regression
Andrei Pantea, Aymeric Dieuleveut
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Extends the Single Location Regression framework to 2D structured data to theoretically study axial attention in time series foundation models.
Problem
Theoretical properties of axial attention, used in time series foundation models, remain largely unexplored; the paper studies sparse information settings where a target depends on a single location.
Approach
Extends the Single Location Regression framework (Marion et al., 2025) from 1D to 2D structured data to analyze axial attention under sparse information settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| keep the inner non-linearity instead of a semi-linear (linear row mixing) predictor (ablation) | axial attention SLR predictor · Single location regression (theory) | risk vs Bayes risk | Improved |
| use small constant inverse temperature during PGD training (ablation) | projected gradient descent on erf predictor · M=N=5, d=20, ell=100 | alignment kappa convergence | Improved |
| add a warm-up phase with larger inverse temperature (ablation) | projected gradient descent on erf predictor · M=N=5, d=20, ell=100 | escape from kappa = 0 | Improved |
| swap erf for softmax in the predictor (ablation) | axial attention SLR predictor · penalized risk, random initialization | convergence to manifold | No clear effect |
Experimental setup
Builds on: Single Location Regression
Closest papers by meaning
- Attention layers provably solve single-location regression (ICLR 2025 · similarity 0.88)
- Calibrating Scientific Foundation Models with Inference-Time Stochastic Attention (NeurIPS 2026 · similarity 0.86)
- In-Context Linear Regression Demystified: Training Dynamics and Mechanistic Interpretability of Multi-Head Softmax Attention (ICML 2025 · similarity 0.85)
- Towards Neural Scaling Laws for Time Series Foundation Models (ICLR 2025 · similarity 0.85)
- Ordinary Least Squares as an Attention Mechanism (NeurIPS 2026 · similarity 0.85)
- Align-RAG: Alignment Is All You Need for TSFM In-Context Learning (NeurIPS 2026 · similarity 0.84)
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.
