Research map / Probabilistic and scientific modeling / Molecules and proteins
Fast Decoding in Sequence Models Using Discrete Latent Variables
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 short discrete latent sequences and parallel output reconstruction | Latent Transformer · WMT English-German newstest2014 | latency, b=1: 408 ms -> 105 ms | Helped |
| rescore the top 10 latent-model outputs | Latent Transformer · WMT English-German newstest2014 | BLEU: 19.8 -> 21.0 | Helped |
| rescore the top 100 latent-model outputs | Latent Transformer · WMT English-German newstest2014 | BLEU: 19.8 -> 22.5 | Helped |
| replace improved semantic hashing with VQ-VAE (ablation) | Latent Transformer · WMT English-German newstest2014 | BLEU: 19.8 -> 2.78 | Hurt |
| replace improved semantic hashing with sliced DVQ (ablation) | Latent Transformer · WMT English-German newstest2014 | BLEU: 19.8 -> 19.7 | No clear effect |
| replace improved semantic hashing with projected DVQ (ablation) | Latent Transformer · WMT English-German newstest2014 | BLEU: 19.8 -> 19.8 | No clear effect |
| decompose vector quantization into 2 codebooks instead of using 1 (ablation) | DVQ · WMT English-German newstest2014 | percentage of latent codewords used: 5% -> 74.5% | Helped |
| increase decompositions from 2 to 4 (ablation) | DVQ · WMT English-German newstest2014 | percentage of latent codewords used: 74.5% -> 15.6% | Hurt |
| increase decompositions from 2 to 8 (ablation) | DVQ · WMT English-German newstest2014 | percentage of latent codewords used: 74.5% -> 31.2% | Hurt |
| increase latent bits from log2K=8 to log2K=14 at n/m=2 (ablation) | LT autoencoder · newstest2013 development set | log-perplexity: 1.33 -> 0.64 | Helped |
| increase compression ratio n/m from 2 to 8 with log2K=8 (ablation) | LT autoencoder · newstest2013 development set | log-perplexity: 1.33 -> 2.44 | Hurt |
| increase compression ratio n/m from 2 to 8 with log2K=14 (ablation) | LT autoencoder · newstest2013 development set | log-perplexity: 0.64 -> 1.77 | Hurt |
Closest papers by meaning
- Deep Encoder, Shallow Decoder: Reevaluating Non-autoregressive Machine Translation (ICLR 2021 · similarity 0.91)
- Fast Structured Decoding for Sequence Models (NeurIPS 2019 · similarity 0.90)
- Blockwise Parallel Decoding for Deep Autoregressive Models (NeurIPS 2018 · similarity 0.90)
- Non-autoregressive Sequence-to-Sequence Vision-Language Models (CVPR 2024 · similarity 0.86)
- Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows (NeurIPS 2025 · similarity 0.86)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · 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.