Research map / Architectures and efficiency / RNNs and normalization
RenewNAT: Renewing Potential Translation for Non-autoregressive 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 |
|---|---|---|---|
| Add the MLM renewal module to Vanilla NAT | Vanilla NAT · WMT14 DE→EN | BLEU: 24.42 -> 27.16 | Helped |
| Add the MLM renewal module to GLAT | GLAT · WMT14 EN→DE | BLEU: 25.21 -> 25.75 | Helped |
| Add the MLM renewal module to GLAT + DSLP | GLAT + DSLP · WMT16 RO→EN | BLEU: 33.03 -> 33.52 | Helped |
| Use K=2 MLM layers in the six-layer decoder (ablation) | RenewNAT · IWSLT14 DE→EN | BLEU: 32.49 -> 33.00 | Helped |
| Increase MLM layers from K=2 to K=4 (ablation) | RenewNAT (N=6) · IWSLT14 DE→EN | BLEU: 33.00 -> 32.03 | Hurt |
| Increase decoder depth to 12 layers and use K=6 MLM layers (ablation) | RenewNAT · IWSLT14 DE→EN | BLEU: 33.00 | No clear effect |
| Mask selected tokens and renew them instead of renewing every word completely (ablation) | RenewNAT · IWSLT14 DE→EN | BLEU: 31.14 -> 33.00 | Helped |
| Use mixed ground-truth and NAT output tokens as MLM input rather than output tokens (ablation) | RenewNAT · IWSLT14 DE→EN | BLEU: 32.73 -> 33.22 | Helped |
| Remove the MLM sub-module (ablation) | RenewNAT on GLAT · WMT14 DE→EN | BLEU: 29.66 -> 29.02 | Hurt |
| Remove the MLM sub-module (ablation) | RenewNAT on GLAT · WMT16 EN→RO | BLEU: 32.52 -> 32.10 | Hurt |
| Use RenewNAT without distillation | GLAT · IWSLT14 DE→EN raw data | BLEU: 28.55 -> 30.12 | Helped |
| Use a confidence threshold chosen for the base model (ablation) | RenewNAT · WMT14 DE→EN | BLEU | Helped |
Closest papers by meaning
- Non-Autoregressive Neural Machine Translation (ICLR 2018 · similarity 0.90)
- Non-autoregressive Machine Translation with Probabilistic Context-free Grammar (NeurIPS 2023 · similarity 0.90)
- Rephrasing the Reference for Non-autoregressive Machine Translation (AAAI 2023 · similarity 0.89)
- AEQA-NAT : Adaptive End-to-end Quantization Alignment Training Framework for Non-autoregressive Machine Translation (ICML 2025 · similarity 0.88)
- DiffNorm: Self-Supervised Normalization for Non-autoregressive Speech-to-speech Translation (NeurIPS 2024 · similarity 0.85)
- Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows (NeurIPS 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.