Research map / Large language models / Language, safety and interpretability
Bidirectional Language Models Are Also 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 |
|---|---|---|---|
| Use SAP instead of one-shot mask infill (ablation) | mT5-3.7B on SQuAD v1.1 dev, 16-shot · SQuAD v1.1 dev (16-shot) | EM: 23.0 -> 35.4 (+12.5 EM) | Helped |
| Use SAP instead of one-shot mask infill (ablation) | mT5-3.7B on SQuAD v1.1 dev, 16-shot · SQuAD v1.1 dev (16-shot) | F1: 54.5 -> 60.0 (+5.5 F1) | Helped |
| Use SAP instead of mT5 without SAP (ablation) | mT5-3.7B on CNN / Daily Mail, 2-shot · CNN / Daily Mail v3.0.0 test (2-shot) | ROUGE-L: +5.3 | Helped |
| Concatenate only the first generated word at each step, rather than the full generation (ablation) | mT5-3.7B on FLORES-101 · FLORES-101 devtest English-Russian and Russian-English (2-shot) | spBLEU: 9.3 -> 20.1; 17.9 -> 26.9 | Helped |
| Iteratively prompt instead of using only the first-step generation (ablation) | mT5-3.7B on FLORES-101 · FLORES-101 devtest English-Russian and Russian-English (2-shot) | spBLEU: 1.9 -> 20.1; 5.6 -> 26.9 | Helped |
| Filter and select synthetic examples using mT5Score instead of random selection (ablation) | mT5-3.7B unsupervised zero-shot translation · FLORES-101 devtest English-Russian | spBLEU: 0.0 -> 20.0 | Helped |
| Filter and select synthetic examples using mT5Score instead of random selection (ablation) | mT5-3.7B unsupervised zero-shot translation · FLORES-101 devtest Russian-English | spBLEU: 25.5 -> 26.3 | Helped |
| Use a prompt ensemble rather than a single prompt (ablation) | mT5-3.7B unsupervised zero-shot translation · FLORES-101 devtest English-Russian | spBLEU: 20.0 -> 20.9 | Helped |
| Use a prompt ensemble rather than a single prompt (ablation) | mT5-3.7B unsupervised zero-shot translation · FLORES-101 devtest Russian-English | spBLEU: 26.3 -> 28.6 | Helped |
| Increase the prompt ensemble from 4 to 16 prompts (ablation) | mT5-3.7B unsupervised zero-shot translation · FLORES-101 devtest English-Russian | spBLEU: 20.9 -> 20.9 | No clear effect |
| Use English-centric rather than standard bootstrapping (ablation) | mT5-3.7B unsupervised zero-shot translation · FLORES-101 devtest English-Russian and Russian-Chinese | spBLEU: 20.9 -> 21.2; 5.8 -> 8.3 | Helped |
| Add bootstrapping and self-amplification (ablation) | mT5-3.7B + SAP · FLORES-101 devtest English-Russian and Russian-English | spBLEU: 0.4 -> 21.2; 4.6 -> 27.9 | Helped |
Closest papers by meaning
- Dialogue for Prompting: A Policy-Gradient-Based Discrete Prompt Generation for Few-Shot Learning (AAAI 2024 · similarity 0.89)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.86)
- BayesPrompt: Prompting Large-Scale Pre-Trained Language Models on Few-shot Inference via Debiased Domain Abstraction (ICLR 2024 · similarity 0.86)
- Concentrate Attention: Towards Domain-Generalizable Prompt Optimization for Language Models (NeurIPS 2024 · similarity 0.86)
- Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization (ICLR 2025 · similarity 0.86)
- SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning (CVPR 2025 · 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.