Research map / Probabilistic and scientific modeling / Molecules and proteins
Levenshtein Transformer
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 LevT distillation instead of Transformer beam4 | Machine translation · Ro-En test | BLEU: 32.30 -> 33.26 | Helped |
| Use LevT distillation instead of Transformer beam4 | Machine translation · En-De test | BLEU: 27.17 -> 27.27 | Helped |
| Use LevT distillation instead of Transformer beam4 | Machine translation · En-Ja test | BLEU: 43.68 -> 43.17 | Hurt |
| Use LevT distillation instead of Transformer beam4 | Text summarization · Gigaword test | ROUGE-1: 37.87 -> 37.40 | Hurt |
| Use LevT distillation instead of Transformer beam4 | Text summarization · Gigaword test | ROUGE-2: 18.92 -> 18.33 | Hurt |
| Use LevT distillation instead of Transformer beam4 | Text summarization · Gigaword test | ROUGE-L: 35.13 -> 34.51 | Hurt |
| Use LevT distillation instead of Transformer beam4 | Sequence generation · Ro-En test | Latency (ms): 349 -> 90 | Helped |
| Use LevT distillation instead of Transformer beam4 | Sequence generation · En-De test | Latency (ms): 369 -> 92 | Helped |
| Apply MT-trained LevT directly to post-editing | APE refinement · Synthetic APE tasks | BLEU: 27.5 -> 30.1; 26.2 -> 28.0; 15.4 -> 16.5; 37.7 -> 39.4 | Mixed |
| Fine-tune MT-pretrained LevT on post-editing data | APE refinement · Real En-De APE | BLEU: 59.6 -> 70.1 | Helped |
| Do not share deletion weights with insertion operations (ablation) | LevT · En-De test | BLEU: 27.27 -> 27.73 | Helped |
| Share placeholder and token classifier weights (ablation) | LevT · En-De test | BLEU: 25.11 -> 27.73 | Helped |
Closest papers by meaning
- Insertion Transformer: Flexible Sequence Generation via Insertion Operations (ICML 2019 · similarity 0.88)
- Deliberation Networks: Sequence Generation Beyond One-Pass Decoding (NeurIPS 2017 · similarity 0.86)
- Sequence to Better Sequence: Continuous Revision of Combinatorial Structures (ICML 2017 · similarity 0.86)
- Target-Side Input Augmentation for Sequence to Sequence Generation (ICLR 2022 · similarity 0.85)
- Copy That! Editing Sequences by Copying Spans (AAAI 2021 · similarity 0.85)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.84)
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.