Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Pretraining task diversity and the emergence of non-Bayesian in-context learning for regression
Allan Raventos, Mansheej Paul, Feng Chen, Surya Ganguli
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 pretraining task diversity beyond about 2^6 tasks | Pretrained transformer · Linear regression, new tasks | MSE relative to dMMSE and Ridge | Improved |
| Keep diversity low and increase the number of sequences per task | Pretrained transformer · Linear regression, new tasks | alignment with Ridge | Hurt |
| Increase sequences per task while keeping diversity below the threshold | Pretrained transformer · Linear regression, tasks drawn from the pretraining distribution | prediction similarity to dMMSE | Improved |
| Increase sequences per task while keeping diversity above the threshold | Pretrained transformer · Linear regression, new tasks | prediction similarity to Ridge | Improved |
| Increase training steps from 500K to 1M at fixed batch size | Pretrained transformer · Linear regression, new tasks | task diversity threshold | No clear effect |
| Increase task dimension at fixed SNR | Pretrained transformer · Linear regression, new tasks | task diversity threshold | Hurt |
| Increase task dimension at M=2^20 | Pretrained transformer · Linear regression, new tasks | prediction difference from Ridge | No clear effect |
| Increase task dimension at M=2^20 | dMMSE · Linear regression, new tasks | prediction difference from Ridge | Hurt |
| Add weight decay (ablation) | Pretrained transformer · Linear regression, new tasks | task diversity threshold | Improved |
| Add weight decay (ablation) | Pretrained transformer · Linear regression, new tasks | performance | Hurt |
| Increase embedding dimension in the small transformer (ablation) | Small transformer · Linear regression, new tasks | task diversity threshold | Hurt |
| Increase depth in the small transformer (ablation) | Small transformer · Linear regression, new tasks | task diversity threshold | Hurt |
Closest papers by meaning
- When can in-context learning generalize out of task distribution? (ICML 2025 · similarity 0.95)
- How Many Pretraining Tasks Are Needed for In-Context Learning of Linear Regression? (ICLR 2024 · similarity 0.93)
- In-Context Learning through the Bayesian Prism (ICLR 2024 · similarity 0.92)
- Technical Debt in In-Context Learning: Diminishing Efficiency in Long Context (NeurIPS 2025 · similarity 0.91)
- Emergence and Effectiveness of Task Vectors in In-Context Learning: An Encoder Decoder Perspective (ICML 2025 · similarity 0.91)
- Differential learning kinetics govern the transition from memorization to generalization during in-context learning (ICLR 2025 · similarity 0.91)
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.
