Research map / Large language models / Language, safety and interpretability
How Does the Pretraining Distribution Shape In-Context Learning? A Fundamental Trade-Off
Waïss Azizian, Ali Hasan
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 heavier-tailed task prior (lower Student-t degrees of freedom) in pretraining (ablation) | transformer ICL, linear regression, N=infinity · linear regression under task shift | log MSE | Mixed |
| use heavier-tailed prior with small number of pretraining tasks (N=100) (ablation) | transformer ICL, linear regression · linear regression under task shift, finite N | log MSE | Hurt |
| use heavier-tailed prior (lower nu or beta) in pretraining (ablation) | transformer ICL, Ornstein-Uhlenbeck next-step prediction · OU process under task shift | MSE | Improved |
| stronger temporal dependence (smaller Volterra kernel exponent) (ablation) | transformer ICL, Volterra process · Volterra equations, N from 5000 to 500 | MSE: MSE at shift 0 multiplied by 3 for exponent 1.0; barely changes for exponent 2.0 | Hurt |
| importance-reweight pretraining tasks toward uniform over observed range (ablation) | Student-t pretraining prior, linear regression and OU · task shift evaluation | MSE: small improvement under large shifts | Improved |
| importance-reweight pretraining tasks (ablation) | generalized normal prior, OU process · OU process task shift | MSE | No clear effect |
Closest papers by meaning
- Dual Operating Modes of In-Context Learning (ICML 2024 · similarity 0.90)
- In-Context Learning through the Bayesian Prism (ICLR 2024 · similarity 0.90)
- When can in-context learning generalize out of task distribution? (ICML 2025 · similarity 0.90)
- In-Context Learning Strategies Emerge Rationally (NeurIPS 2025 · similarity 0.90)
- Toward Understanding In-context vs. In-weight Learning (ICLR 2025 · similarity 0.90)
- Position: Do pretrained Transformers Learn In-Context by Gradient Descent? (ICML 2024 · similarity 0.89)
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.
