Research map / Architectures and efficiency / RNNs and normalization
Model-Level Dual Learning
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 |
|---|---|---|---|
| tie parameters of primal and dual models (symmetric model-level dual learning) | Transformer for NMT · IWSLT De→En | BLEU: +1.85 over Transformer baseline (32.86 to 34.71) | Helped |
| tie parameters of primal and dual models (symmetric model-level dual learning) | Transformer for NMT · IWSLT En→De | BLEU: +0.90 (27.74 to 28.64) | Helped |
| model-level dual learning instead of data-level dual supervised learning | Transformer for NMT · IWSLT De↔En | BLEU: +1.13 (De→En) and +0.73 (En→De) over DSL | Helped |
| tie parameters of primal and dual models | Transformer big for WMT14 · WMT14 En↔De | BLEU: +0.5 on En→De (28.4 to 28.9) and +0.5 on De→En (31.4 to 31.9) | Helped |
| tie parameters of primal and dual models | Transformer big for NIST Zh↔En · NIST Zh→En | BLEU: about +1.21 average over Transformer | Helped |
| tie parameters of primal and dual models | Transformer big for NIST En→Zh · NIST En→Zh | BLEU: +0.69 BLEU on average | Helped |
| asymmetric model-level dual learning (sharing LSTM encoder and label embedding) for sentiment analysis | LSTM classifier on IMDB · IMDB sentiment (supervised) | error rate (%): 10.10 to 7.41 (-2.69); 9.20 to 7.41 vs DSL | Helped |
| asymmetric model-level dual learning | LSTM on IMDB dual task · IMDB conditional perplexity | perplexity: 58.78 (DSL) to 55.59 | Helped |
| add dual inference at test time (ablation) | model-level dual learning on Transformer · IWSLT De↔En | BLEU: +0.48 De→En (34.71 to 35.19); +0.19 En→De (28.64 to 28.83) | Helped |
| add dual inference at test time (ablation) | model-level dual learning LSTM classifier · IMDB sentiment | error rate (%): 7.41 to 6.96 | Helped |
| model-level parameter sharing vs baseline, varying depth from 2 to 8 blocks (ablation) | Transformer on IWSLT De↔En · IWSLT De↔En block sweep | BLEU: gain consistent across 2/4/6/8 blocks, larger for 2 to 6 blocks | Helped |
| tie parameters of primal and dual models | Transformer for NMT · model size | parameter count: 38M to 20M parameters (about 47% reduction) on IWSLT | Helped |
Closest papers by meaning
- Dual Supervised Learning (ICML 2017 · similarity 0.90)
- Dual Learning for Machine Translation (NeurIPS 2016 · similarity 0.90)
- TANDEM: Bi-Level Data Mixture Optimization with Twin Networks (NeurIPS 2025 · similarity 0.84)
- Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality (ICML 2024 · similarity 0.84)
- Alignment of Large Language Models with Constrained Learning (NeurIPS 2025 · similarity 0.83)
- Dual Process Learning: Controlling Use of In-Context vs. In-Weights Strategies with Weight Forgetting (ICLR 2025 · similarity 0.83)
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.