Research map / Large language models / Language, safety and interpretability
Meta-in-context learning in large language models
Julian Coda-Forno, Marcel Binz, Zeynep Akata, Matthew Botvinick, Jane X Wang, Eric Schulz
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 |
|---|---|---|---|
| Present examples from previous tasks in the context before a new regression task | GPT-3 · 1D regression | MSE: -13.26 ± 3.08 task-number effect on MSE | Improved |
| Present examples from previous tasks in the context before a new bandit task | GPT-3 · Two-armed bandit | regret: -0.031 ± 0.012 task-number effect on regret | Improved |
| Use sequential task experience to adapt the model before real-world regression | GPT-3 · Real-world regression benchmark | RMSE | Improved |
| Apply meta-in-context learning to the real-world regression benchmark | GPT-4 · Real-world regression benchmark | RMSE | Improved |
| Use GPT-4 rather than GPT-3 for real-world regression, before or after meta-in-context learning | Real-world regression · Real-world regression benchmark | RMSE | Hurt |
| Encounter more similar previous tasks | GPT-3 after meta-in-context learning · Real-world regression benchmark | RMSE: -0.147 ± 0.020 | Improved |
| Encounter more similar previous tasks | GPT-4 after meta-in-context learning · Real-world regression benchmark | RMSE: -0.196 ± 0.024 | Improved |
| Accumulate additional observations within a task | GPT-3 · 1D regression | MSE: -30.614 ± 3.08 trial-number effect | Improved |
| Accumulate additional observations within a task | GPT-3 · Two-armed bandit | regret: -0.221 ± 0.012 trial-number effect | Improved |
| Accumulate additional observations within a task | GPT-3 · Real-world regression benchmark | RMSE: -0.089 ± 0.020 trial-number effect | Improved |
| Meta-in-context learning shifts prior expectations toward encountered task statistics | GPT-3 · Two-armed bandit | prior expected reward mean: 545.22 -> 6.49 | Improved |
| Meta-in-context learning constrains predictions to plausible target values | GPT-3 · Real-world regression benchmark | proportion of predictions outside target range | Improved |
Closest papers by meaning
- Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence (ICML 2025 · similarity 0.91)
- Context-Aware Meta-Learning (ICLR 2024 · similarity 0.90)
- In-context Learning in Presence of Spurious Correlations (NeurIPS 2026 · similarity 0.90)
- In-Context Language Learning: Architectures and Algorithms (ICML 2024 · similarity 0.89)
- Probing the Decision Boundaries of In-context Learning in Large Language Models (NeurIPS 2024 · similarity 0.89)
- An Information-Theoretic Analysis of In-Context Learning (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.
