Research map / Architectures and efficiency / RNNs and normalization
Understanding Knowledge Distillation in Non-autoregressive 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 beam search instead of sampling to generate teacher targets (ablation) | Vanilla NAT · WMT14 En-De newstest2014 | BLEU: 6.6 -> 19.5 | Helped |
| Use beam search instead of top-10 sampling to generate teacher targets (ablation) | Vanilla NAT · WMT14 En-De newstest2014 | BLEU: 14.6 -> 19.5 | Helped |
| Use beam search instead of greedy decoding to generate teacher targets (ablation) | Vanilla NAT · WMT14 En-De newstest2014 | BLEU: 18.9 -> 19.5 | Helped |
| Train FlowSeq with distilled targets from Transformer-small rather than Transformer-base (ablation) | FlowSeq · WMT14 En-De newstest2014 | BLEU: 18.6 -> 22 | Helped |
| Use the sixth born-again teacher's distilled data (ablation) | Vanilla NAT · WMT14 En-De newstest2014 | BLEU: +2 BLEU | Helped |
| Use distilled data from a 3-expert MoE teacher rather than a plain base Transformer teacher (ablation) | Vanilla NAT · WMT14 En-De newstest2014 | BLEU: +1.21 BLEU | Helped |
| Select teacher outputs by sentence-level BLEU interpolation rather than standard distillation (ablation) | LevT · WMT14 En-De newstest2014 | BLEU: 26.94 -> 27.32 | Helped |
| Match the distilled-data complexity to student capacity, using smaller-teacher data for weaker students (ablation) | NAT models · WMT14 En-De newstest2014 | BLEU | Helped |
| Replace real multi-mode targets with teacher-decoded targets (ablation) | NAT models · Europarl synthetic En-De/Es/Fr | mode consistency | Helped |
Closest papers by meaning
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.88)
- Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.88)
- DDK: Distilling Domain Knowledge for Efficient Large Language Models (NeurIPS 2024 · similarity 0.87)
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · similarity 0.87)
- AEQA-NAT : Adaptive End-to-end Quantization Alignment Training Framework for Non-autoregressive Machine Translation (ICML 2025 · similarity 0.86)
- On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes (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.