Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
The Unreasonable Effectiveness of Few-shot Learning for Machine Translation
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 MBR decoding (64 samples, BLEURT utility) instead of beam search (ablation) | Bilingual and trilingual 8B few-shot LMs · newstest21 WMT'21 en/zh/de | BLEURT: e.g. en→zh 62.6 beam vs 68.4 MBR for bilingual | Helped |
| Use few-shot demonstrations from the same regional variety as the reference (matched) instead of mismatched (ablation) | 8B bilingual LM on FRMT Mandarin · FRMT Mandarin | Lexical Accuracy: 86.2 matched vs 13.8 mismatched | Helped |
| Use demonstrations matching the target formality level instead of mismatched formality (ablation) | 8B bilingual LM on IWSLT22 En-De formality · IWSLT22 En-De formality | Formality accuracy: 84.9% matched vs 15.1% mismatched (formal); 85.5% vs 15.5% (informal) | Helped |
| Use high-quality demonstrations (low CDS score bucket) instead of lower-quality ones (higher CDS bucket) (ablation) | 8B bilingual LM few-shot translation · Paracrawl En-De demonstration buckets | BLEURT: not reported numerically (figure) | Helped |
| Multi-epoch training on Icelandic (fine-tuning up to ~6 epochs) instead of single pass (ablation) | En-Is 8B bilingual LM · Flores devtest / newstest21 En-Is | BLEURT: not reported numerically; single pass described as suboptimal, en→is benefits up to 6 epochs, is→en plateaus after 2 | Helped |
| Add trilingual training (third language with same per-language data) vs bilingual (ablation) | 8B decoder-only LM, en-zh-de · newstest21 en→zh | BLEURT: 68.9 trilingual vs 68.4 bilingual (MBR) | Helped |
| Use PaLM-style few-shot 540B model instead of 8B LM with 5 WMT-dev demos | WMT'21 en→de and de→en · newstest21 en→de | BLEURT: PaLM 75.9 vs bilingual MBR 75.5 (en→de); PaLM 74.8 vs 76.5 (de→en) | Mixed |
| Few-shot 5-demo MBR 8B LM vs supervised WMT'21 systems on en→zh | 8B trilingual LM · newstest21 en→zh | BLEURT: 68.9 vs WMT'21 1st 70.0 (not better); outperforms best WMT en→zh system in trilingual claim | Helped |
| Few-shot 8B LM vs commercial/supervised systems in XX→English direction | 8B bilingual and trilingual LMs · newstest21 zh→en, de→en | BLEURT: e.g. zh→en 68.3 trilingual vs 66.6 WMT 1st (better), but de→en 76.8 vs 76.9 close; authors state underperformance in XX→ | Hurt |
| Warm-start from English-German model when adapting to Icelandic | 8B LM for En-Is · newstest21 en→is | BLEURT: no ablation reported | No clear effect |
Closest papers by meaning
- Adaptive Few-shot Prompting for Machine Translation with Pre-trained Language Models (AAAI 2025 · similarity 0.88)
- Optimization Inspired Few-Shot Adaptation for Large Language Models (NeurIPS 2025 · similarity 0.88)
- Many-Shot In-Context Learning (NeurIPS 2024 · similarity 0.87)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.86)
- EBBS: An Ensemble with Bi-Level Beam Search for Zero-Shot Machine Translation (AAAI 2025 · similarity 0.86)
- The Surprising Effectiveness of Test-Time Training for Few-Shot Learning (ICML 2025 · similarity 0.85)
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.