Research map / Architectures and efficiency / RNNs and normalization
Navigating with Graph Representations for Fast and Scalable Decoding of Neural Language Models
Minjia Zhang, Wenhan Wang, Xiaodong Liu, Jianfeng Gao, Yuxiong He
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 |
|---|---|---|---|
| Replace full softmax with FGD at efSearch 50 | NMT decoder · IWSLT’14 German-English | softmax execution time: 6.3ms -> 0.43ms | Improved |
| Replace full softmax with FGD at efSearch 20 | NMT decoder · IWSLT’14 German-English | BLEU: losing 0.67 BLEU score | Hurt |
| Increase efSearch from 20 to 50 (ablation) | FGD · IWSLT’14 German-English | BLEU: 26.69 -> 29.06 | Improved |
| Increase efSearch from 20 to 50 (ablation) | FGD · IWSLT’14 German-English | softmax execution time: 0.21ms -> 0.43ms | Hurt |
| Increase efSearch from 20 to 200 (ablation) | FGD · IWSLT’14 German-English | precision@1: 0.939 -> 0.992 | Improved |
| Increase efSearch from 20 to 200 (ablation) | FGD · IWSLT’14 German-English | distance calculations: 981 / 50K -> 5785 / 50K | Hurt |
| Increase beam size from 1 to 10 (ablation) | FGD · IWSLT’14 German-English | BLEU: 26.69 -> 27.62 (efSearch 20); 27.55 -> 28.9 (efSearch 50) | Mixed |
| Increase vocabulary size from 10K to 80K (ablation) | FGD language model · WikiText-2 | precision@1: 0.938 -> 0.933 (efSearch 50); 1.000 -> 0.982 (efSearch 200) | Mixed |
| Increase vocabulary size (ablation) | FGD language model · WikiText-2 | decoding time | Hurt |
| Use SVD-softmax instead of FGD | NMT decoder · IWSLT’14 German-English | softmax execution time: 0.43ms vs 5.53ms | Hurt |
Closest papers by meaning
- Towards Interpretability Without Sacrifice: Faithful Dense Layer Decomposition with Mixture of Decoders (NeurIPS 2025 · similarity 0.88)
- MBR and QE Finetuning: Training-time Distillation of the Best and Most Expensive Decoding Methods (ICLR 2024 · similarity 0.87)
- Foundations of Top-$k$ Decoding for Language Models (NeurIPS 2025 · similarity 0.87)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.87)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.86)
- SparseLLM: Towards Global Pruning of Pre-trained Language Models (NeurIPS 2024 · 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.
