Research map / Architectures and efficiency / Attention and transformers
Cascaded Text Generation with Markov Transformers
Yuntian Deng, Alexander Rush
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 |
|---|---|---|---|
| add knowledge distillation (train on autoregressive teacher outputs) (ablation) | Cascaded Generation (K=64, iters=2) · WMT14 En-De test | BLEU: 23.09 -> 26.52 | Improved |
| relax length constraint to multiple lengths (DeltaL=3 vs 0) (ablation) | Cascade (K=64, iters=5) · IWSLT14 De-En val | BLEU: 33.45 -> 35.03 | Improved |
| use max-marginals instead of n-gram scores for pruning (ablation) | Cascaded decoding · WMT14 En-De val | BLEU: 28.42 -> 29.24 | Improved |
| use cascaded decoding instead of serial beam search without distillation and no length constraint (ablation) | Markov Transformer · IWSLT14 De-En val | BLEU: 35.07 -> 33.45 | Hurt |
| use cascaded decoding instead of beam search under hard length constraint (ablation) | Markov Transformer · IWSLT14 De-En val | BLEU: 33.33 -> 33.45 | Improved |
| increase iterations (higher Markov order) from 2 to 5 (ablation) | Cascaded Generation (K=64) · WMT14 En-De test | BLEU: 26.52 -> 26.92 | Improved |
| use 4-order Markov transformer vs fully autoregressive transformer with beam 5 | Markov Transformer · IWSLT14 De-En val | BLEU: 35.63 -> 35.07 | Hurt |
| add more iterations to reduce repetitions (ablation) | Cascaded decoding · WMT14 En-De val | ratio of unique n-grams | Improved |
| parallelize decoder across 4 GPUs instead of 1 | Cascaded decoding K=64 iters=5 · WMT14 En-De test | speedup: 1.68x -> 2.79x | Improved |
Closest papers by meaning
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.89)
- Promises and Pitfalls of Generative Masked Language Modeling: Theoretical Framework and Practical Guidelines (ICML 2024 · similarity 0.89)
- Tractable Transformers for Flexible Conditional Generation (ICML 2025 · similarity 0.87)
- Dynamically Structured Diffusion Language Model Decoding via Bayesian Inference (NeurIPS 2026 · similarity 0.87)
- Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows (NeurIPS 2025 · similarity 0.86)
- Energy-Based Diffusion Language Models for Text Generation (ICLR 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.
