Research map / Image generation / Autoregressive and latent generation
Sequence Modeling with Unconstrained Generation Order
Dmitrii Emelianenko, Elena Voita, Pavel Serdyukov
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 unconstrained insertion order instead of fixed-order decoding | Transformer-base machine translation · WMT2017 En-Ru | BLEU: 31.6 -> 33.2 | Improved |
| Use unconstrained insertion order instead of fixed-order decoding | Transformer-base machine translation · WMT2017 Ru-En | BLEU: 35.3 -> 36.4 | Improved |
| Use unconstrained insertion order instead of the strongest fixed order | Transformer-base machine translation · ASPEC En-Ja test | BLEU: 48.6 -> 50.3 | Improved |
| Use unconstrained insertion order instead of fixed-order decoding | Transformer-base machine translation · IWSLT En-Ar | BLEU: 12.0 -> 12.2 | No clear effect |
| Use unconstrained insertion order instead of fixed-order decoding | Transformer-base machine translation · IWSLT En-De | BLEU: 28.04 -> 28.36 | Improved |
| Use unconstrained insertion order instead of fixed-order decoding | Transformer-base machine translation · IWSLT De-En | BLEU: 33.17 -> 33.08 | No clear effect |
| Use unconstrained insertion order instead of fixed-order decoding | Transformer-base Image-to-LaTeX · ImageToLatex-140K | BLEU: 89.5 -> 90.3 | Improved |
| Use unconstrained insertion order instead of fixed-order decoding | Transformer-base image captioning · MSCOCO | BLEU: 18.0 -> 25.6 | Improved |
| Use unconstrained insertion order instead of fixed-order decoding | Transformer-base image captioning · MSCOCO | CIDEr: 56.1 -> 81.0 | Improved |
| Add model-dependent trajectory sampling training after pretraining (ablation) | INTRUS · WMT En-Ru 2012-2013 test | BLEU: 24.6 -> 27.5 | Improved |
| Use sampled trajectory training instead of beam-search argmax trajectory training (ablation) | INTRUS · WMT En-Ru 2012-2013 test | BLEU: 26.6 -> 27.5 | Improved |
| Use uniform feasible-trajectory pretraining rather than left-to-right pretraining (ablation) | INTRUS · WMT En-Ru 2012-2013 test | BLEU: 26.3 -> 27.5 | Improved |
Closest papers by meaning
- Discovering Non-monotonic Autoregressive Orderings with Variational Inference (ICLR 2021 · similarity 0.87)
- A Chance-Constrained Generative Framework for Sequence Optimization (ICML 2020 · similarity 0.87)
- Insertion Transformer: Flexible Sequence Generation via Insertion Operations (ICML 2019 · similarity 0.86)
- Order-Independence Without Fine Tuning (NeurIPS 2024 · similarity 0.85)
- Grammar-Aligned Decoding (NeurIPS 2024 · similarity 0.85)
- TokenSwift: Lossless Acceleration of Ultra Long Sequence Generation (ICML 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.
