Research map / Large language models / Language, safety and interpretability
Sequence Generation with Mixed Representations
Lijun Wu, Shufang Xie, Yingce Xia, Yang Fan, Jian-Huang Lai, Tao Qin, Tieyan Liu
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 |
|---|---|---|---|
| mixed representations from BPE and SP tokenizers (two encoders with mix-attention, two decoders) | Transformer 256 backbone · IWSLT14 translation (12 directions) | BLEU: more than 1.0 BLEU on De→En, En→Ro, Pt-br→En vs Transformer 256 | Improved |
| mixed representations (no co-teaching) vs Transformer 512 baseline | Transformer · IWSLT14 En→De and De→En | BLEU: about 0.5 points for most tasks | Improved |
| co-teaching algorithm added on top of mixed representations | mixed-representation model · IWSLT14 translation (12 directions) | BLEU: more than 1.0 on 11 of 12 tasks; over 1.5 on En→Tr | Improved |
| stacked representations instead of additive mixing of self-attention and mix-attention (ablation) | mixed-representation encoder · IWSLT14 En↔De | BLEU: worse than additive mixed representations on all four En↔De entries (e.g., En→De BPE 28.53 vs 28.96) | Hurt |
| co-teaching with a single tokenizer (forward-translated data, BPE only) (ablation) | Transformer 512 · IWSLT14 En↔De | BLEU: En→De BPE 28.80 to 29.67; De→En BPE 34.84 to 36.10 | Improved |
| two-tokenizer co-teaching vs single-tokenizer co-teaching (ablation) | mixed-representation model · IWSLT14 En→De SP | BLEU: 29.93 (two tokenizers) vs 29.28 (one tokenizer) | Improved |
| replace SentencePiece with WordPiece in the mixed model (BPE+WP) (ablation) | mixed-representation model · IWSLT14 En↔De | BLEU: En→De 28.98 (BPE) / 29.15 (WP) vs 28.80 baseline; De→En 35.23/35.24 vs 34.84 | Improved |
| mixed representations + co-teaching on abstractive summarization | Transformer 256 backbone · Gigaword | ROUGE-1: 36.0 vs 35.5 (Transformer 512) | Improved |
Closest papers by meaning
- The Foundations of Tokenization: Statistical and Computational Concerns (ICLR 2025 · similarity 0.90)
- An Analysis of Tokenization: Transformers under Markov Data (NeurIPS 2024 · similarity 0.86)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.86)
- Exact Byte-Level Probabilities from Tokenized Language Models for FIM-Tasks and Model Ensembles (ICLR 2025 · similarity 0.86)
- Scaffold-BPE: Enhancing Byte Pair Encoding for Large Language Models with Simple and Effective Scaffold Token Removal (AAAI 2025 · similarity 0.86)
- From Bytes to Ideas: Language Modeling with Autoregressive U-Nets (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.
