Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Embroid: Unsupervised Prediction Smoothing Can Improve Few-Shot Classification
Neel Guha, Mayee F Chen, Kush Bhatia, Azalia Mirhoseini, Frederic Sala, Christopher Re
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 |
|---|---|---|---|
| Add neighborhood-smoothed predictions from embeddings and combine them with the original prompt prediction | Prompt-based classifiers · 95 tasks, GPT-JT | F1 improvement: 7.3 | Improved |
| Use Embroid prediction correction | GPT-3.5 prompts · 12 representative tasks | F1 improvement: 4.9 | Improved |
| Use Embroid instead of three-prompt aggregation | J1-Jumbo predictions · All tasks (average macro-F1) | macro-F1: 60.7 -> 64.5 | Improved |
| Use Embroid instead of three-prompt aggregation | GPT-3.5 predictions · All tasks (average macro-F1) | macro-F1: 84.7 -> 86.0 | Improved |
| Apply Embroid to base prompt predictions | GPT-3.5 · 10 tasks | average F1: 76.3 -> 81.9 | Improved |
| Apply Embroid to chain-of-thought prompt predictions | GPT-3.5 · 10 tasks | average F1: 80.1 -> 85.4 | Improved |
| Use two legal embedding spaces rather than general BERT and RoBERTa (ablation) | Embroid with GPT-JT · Law tasks | task win rate: 77\% of tasks; up to 6 points F1 | Improved |
| Use two science embedding spaces rather than general BERT and RoBERTa (ablation) | Embroid with GPT-JT · Science tasks | task win rate: 92\% of tasks; up to 4.3 points F1 | Improved |
| Add text-embedding-ada-002 embeddings (ablation) | Embroid · 19 tasks | F1 improvement: up to 13 points F1; average of 2 points | Improved |
| Increase the number of prompts used by Embroid from one to three (ablation) | Embroid · All tasks (average macro-F1), J1-Jumbo | macro-F1: 60.4 -> 64.5 | Improved |
| Increase the number of prompts used by Embroid from one to three (ablation) | Embroid · All tasks (average macro-F1), GPT-3.5 | macro-F1: 83.9 -> 86.0 | Improved |
| Use selective annotation with 100 labeled examples to construct the prompt, then apply Embroid | GPT-JT · All tasks | F1 improvement: 88\% of tasks; average of 4.3 points | Improved |
Closest papers by meaning
- Gen-Z: Generative Zero-Shot Text Classification with Contextualized Label Descriptions (ICLR 2024 · similarity 0.86)
- Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.86)
- SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.86)
- ProAPO: Progressively Automatic Prompt Optimization for Visual Classification (CVPR 2025 · similarity 0.86)
- Large Language Model Unlearning via Embedding-Corrupted Prompts (NeurIPS 2024 · similarity 0.86)
- PromptCoT: Align Prompt Distribution via Adapted Chain-of-Thought (CVPR 2024 · similarity 0.86)
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.
