Research map / Large language models / Language, safety and interpretability
Is In-Context Learning Learning?
Adrian de Wynter
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 the number of in-context exemplars | ICL (all prompts) · Nine formal tasks | accuracy: slope 8.3 (modus ponens) to 3.3 (CoT) | Improved |
| Increase shots until models converge | ICL · Nine formal tasks | accuracy gap between models: -2.6 per shot slope of model gap | Improved |
| Replace natural-language system prompt with random words (word salad) (ablation) | ICL · Nine formal tasks | accuracy | No clear effect |
| Randomise the exemplars and the prompt (Salad-of-Thought) (ablation) | CoT · Nine formal tasks | accuracy: average 23 +/- 4 | Hurt |
| Shuffle position of the same exemplars (ablation) | ICL · Nine formal tasks | accuracy | No clear effect |
| Fully randomise exemplars drawn i.i.d. from the training set (ablation) | GPT-4o · Nine formal tasks | accuracy | Hurt |
| Use only negative labels in exemplars (imbalanced) (ablation) | GPT-4o · Eight formal tasks | accuracy | No clear effect |
| Use chain-of-thought prompting | ICL · Distribution shift, delta to 0.85 | accuracy: OOD slope -1.4 versus -0.4 for modus ponens | Mixed |
| Increase distributional shift between exemplars and test | ICL · Nine formal tasks | accuracy: average slope -0.6 +/- 0.3 | Hurt |
| Use classical baselines instead of ICL | kNN, DT, MLP · Nine formal tasks | accuracy | Mixed |
| Factor out parsing (compliance) errors (ablation) | ICL · Nine formal tasks | perceived accuracy | Mixed |
| Add chain-of-thought prompt to few-shot | CoT · Nine formal tasks | accuracy: best prompt in four tasks; worst in Pattern Matching and Hamiltonian | Mixed |
Closest papers by meaning
- Can In-context Learning Really Generalize to Out-of-distribution Tasks? (ICLR 2025 · similarity 0.91)
- Dual Operating Modes of In-Context Learning (ICML 2024 · similarity 0.91)
- Capturing In-Context Learning Dynamics with Task Operators (NeurIPS 2026 · similarity 0.91)
- In-Context Learning Strategies Emerge Rationally (NeurIPS 2025 · similarity 0.90)
- Investigating the Pre-Training Dynamics of In-Context Learning: Task Recognition vs. Task Learning (ICLR 2025 · similarity 0.90)
- Toward Understanding In-context vs. In-weight Learning (ICLR 2025 · similarity 0.90)
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.
