Research map / Architectures and efficiency / Attention and transformers
Insertion Transformer: Flexible Sequence Generation via Insertion Operations
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 |
|---|---|---|---|
| train with a balanced binary-tree ordering instead of left-to-right ordering | Insertion Transformer · WMT 2014 English-German newstest2014 | BLEU: 23.94 -> 27.29 | Helped |
| use parallel insertion decoding with the balanced binary-tree objective | Insertion Transformer + Binary Tree · WMT 2014 English-German newstest2014 | iterations: n -> ≈ log₂ n | Helped |
| train on teacher-distilled data | Insertion Transformer · WMT 2014 English-German newstest2013 | BLEU: The paper reports improvements of 3 to 4 BLEU points across the board. | Helped |
| apply an EOS penalty during decoding | Insertion Transformer · WMT 2014 English-German newstest2013 | BLEU: nearly 4 points | Helped |
| add joint normalization or contextualized vocabulary bias (ablation) | Insertion Transformer baseline · WMT 2014 English-German newstest2013, no EOS penalty | BLEU: 22.39 -> 22.92; 22.39 -> 23.00 | Helped |
| add joint normalization or contextualized vocabulary bias (ablation) | Insertion Transformer baseline · WMT 2014 English-German newstest2013, with EOS penalty | BLEU: 25.58 -> 25.14; 25.58 -> 25.41 | No clear effect |
| add mixture-of-softmaxes (ablation) | Insertion Transformer baseline · WMT 2014 English-German newstest2013, uniform loss, slot finalization, distilled data | BLEU (+EOS): 25.58 -> 25.58 | No clear effect |
| use a uniform loss instead of a binary-tree loss | Insertion Transformer · WMT 2014 English-German newstest2014 | BLEU: 27.29 -> 27.12 | Hurt |
Closest papers by meaning
- InsNet: An Efficient, Flexible, and Performant Insertion-based Text Generation Model (NeurIPS 2022 · similarity 0.89)
- MARS: Enabling Autoregressive Models Multi-Token Generation (NeurIPS 2026 · similarity 0.87)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.87)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.87)
- Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion (NeurIPS 2026 · similarity 0.86)
- Multi-Branch Self-Drafting for LLM Inference Acceleration (AAAI 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.