Research map / Label-efficient and robust learning / Recognition and long-tail
Simple and Scalable Nearest Neighbor 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 |
|---|---|---|---|
| Add a distance-aware adapter to SK-MT (ablation) | SK-MT · Multi-domain test sets (Avg.) | BLEU: 39.6 -> 45.4; 42.3 -> 46.4 | Helped |
| Use m=16 reference samples instead of m=2 (ablation) | SK-MT · Multi-domain test sets (Avg.) | BLEU: 45.4 -> 46.4 | Helped |
| Use a dynamic small datastore and adaptive retrieval | kNN-MT · IT, Medical, Koran, Law test sets (Avg.) | BLEU: 45.5 -> 45.4 | No clear effect |
| Use m=16 dynamic reference samples rather than the full datastore | kNN-MT · IT, Medical, Koran, Law test sets (Avg.) | BLEU: 45.5 -> 46.4 | Helped |
| Increase reference sample count from m=8 to m=16 (ablation) | SK-MT · Law development set | BLEU | No clear effect |
| Increase k from 1 to 4 (ablation) | SK-MT · IT development set, m=2 | BLEU: 43.4 -> 41.6 | Hurt |
| Set temperature τ to 100 rather than τ=5 (ablation) | SK-MT · IT development set, m=16, k=2 | BLEU: 38.2 -> 43.9 | Helped |
| Apply sentence-level dynamic retrieval to WMT'14 | kNN-MT · WMT'14 De⇒En test | BLEU: 31.3 -> 31.4 | No clear effect |
| Use sentence-level dynamic retrieval instead of full-datastore search | kNN-MT · Law test set, batch=8 | Inference speed (ms/sentence): 467.4 -> 94.6 | Helped |
| Use sentence-level dynamic retrieval instead of full-datastore search | kNN-MT · Law datastore | Storage overhead: 37.0G -> 0.16M | Helped |
| Use SK-MT₂ rather than kNN-MT for online adaptation | NMT with online human-feedback adaptation · EMEA full | BLEU: 43.4 -> 49.7 | Helped |
| Use SK-MT₂ rather than kNN-MT for online adaptation | NMT with online human-feedback adaptation · JRC-Acquis full | BLEU: 48.1 -> 49.8 | Helped |
Closest papers by meaning
- Nearest Neighbor Machine Translation (ICLR 2021 · similarity 0.94)
- Generalization through Memorization: Nearest Neighbor Language Models (ICLR 2020 · similarity 0.89)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.88)
- Test-Time Training on Nearest Neighbors for Large Language Models (ICLR 2024 · similarity 0.87)
- Memory Decoder: A Pretrained, Plug-and-Play Memory for Large Language Models (NeurIPS 2025 · similarity 0.86)
- A Benchmark for Learning to Translate a New Language from One Grammar Book (ICLR 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.