Research map / Architectures and efficiency / Attention and transformers
Theory of Scaling Laws for In-Context Regression: Depth, Width, Context and Time
Blake Bordelon, Mary Letey, Cengiz Pehlevan
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 |
|---|---|---|---|
| increase depth L at infinite context length (ablation) | deep linear attention · ISO (isotropic covariates) | ICL loss | No clear effect |
| increase depth L at finite context length alpha near 1 (ablation) | deep linear attention · ISO (isotropic covariates) | ICL loss | Improved |
| increase depth L beyond 1 at long context (ablation) | deep linear attention · FS (fixed structured covariance) | ICL loss | No clear effect |
| increase depth L (ablation) | deep linear attention · RRS (randomly rotated covariances), infinite context | ICL loss | Improved |
| scale only width or only depth (ablation) | deep linear attention · RRS powerlaw data | ICL loss | Hurt |
| increase width and depth together with compute (ablation) | deep linear attention · RRS powerlaw data | ICL loss | Improved |
| evaluate FS-pretrained model under covariance shift (ablation) | deep linear attention · FS, shifted covariance | OOD loss | Hurt |
| use softmax attention trained with Adam instead of linear attention (ablation) | softmax transformer · ICL regression | ICL loss | Improved |
Closest papers by meaning
- Training Dynamics of In-Context Learning in Linear Attention (ICML 2025 · similarity 0.90)
- How Many Pretraining Tasks Are Needed for In-Context Learning of Linear Regression? (ICLR 2024 · similarity 0.89)
- Fine-grained Analysis of In-context Linear Estimation: Data, Architecture, and Beyond (NeurIPS 2024 · similarity 0.89)
- In-Context Linear Regression Demystified: Training Dynamics and Mechanistic Interpretability of Multi-Head Softmax Attention (ICML 2025 · similarity 0.89)
- On the Invariance and Generality of Neural Scaling Laws (NeurIPS 2026 · similarity 0.88)
- Scaling Laws in Linear Regression: Compute, Parameters, and Data (NeurIPS 2024 · 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.
