Research map / Large language models / Language, safety and interpretability
Selective Annotation Makes Language Models Better Few-Shot Learners
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Select 100 annotations with vote-k rather than randomly | GPT-J/Codex in-context learning · 10 NLP datasets | task metric: +5.2% absolute gain and 11.4% relative gain on average | Helped |
| Select 18 annotations with vote-k rather than randomly | GPT-J/Codex in-context learning · 10 NLP datasets | task metric: 5.8% absolute gain and 12.9% relative gain on average | Helped |
| Use vote-k rather than random annotation (ablation) | GPT-J · HellaSwag, SST-5, MWoZ | task metric: 62.5 vs. 63.2 on HellaSwag; 35.7 vs. 43.8 on MWoZ | Hurt |
| Replace vote-k with MFL, diversity, or least-confidence selection (ablation) | GPT-J/Codex in-context learning · HellaSwag, SST-5, MWoZ | task metric: vote-k: 70.7, 53.0, 51.4; MFL: 66.5, 45.6, 48.3; Diversity: 68.2, 48.5, 49.2; Least-confidence: 68.4, 46.2, 49.4 | Hurt |
| Remove confidence-score computations from vote-k (ablation) | vote-k · HellaSwag, SST-5, MWoZ | task metric: vote-k: 70.7, 53.0, 51.4; Fast vote-k: 69.5, 51.9, 50.2 | Mixed |
| Apply vote-k selection rather than random selection | GPT-J in-context learning · CivilComments random/domain splits | accuracy: +5.5 random; +9.9 domain | Helped |
| Use vote-k rather than random selection for finetuning without outlier removal (ablation) | Finetuning · HellaSwag and MRPC | task performance: HellaSwag: 55.6 vs 53.5; MRPC: 56.3 vs 55.6 | Hurt |
| Remove 10% outliers, then use vote-k rather than random selection (ablation) | Finetuning · HellaSwag and MRPC | task performance: HellaSwag: 56.8 vs 59.6; MRPC: 57.9 vs 60.4 | Helped |
| Increase model size while retaining vote-k selection | In-context learning models · HellaSwag | accuracy gain over random: 7.5% using GPT-Neo; 2.6% using GPT-3 | Mixed |
Closest papers by meaning
- IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models (ICLR 2024 · similarity 0.92)
- GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition (AAAI 2025 · similarity 0.87)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.87)
- Where does In-context Learning Happen in Large Language Models? (NeurIPS 2024 · similarity 0.87)
- Large Language Models are Demonstration Pre-Selectors for Themselves (ICML 2025 · similarity 0.87)
- GistScore: Learning Better Representations for In-Context Example Selection with Gist Bottlenecks (ICML 2024 · similarity 0.87)
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.