Research map / Architectures and efficiency / RNNs and normalization
Mixtape: Breaking the Softmax Bottleneck Efficiently
Zhilin Yang, Thang Luong, Ruslan Salakhutdinov, Quoc V Le
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 gate sharing for infrequent tokens (ablation) | Mixtape · Penn Treebank | perplexity: 56.33 -> 56.37 | No clear effect |
| add gate sharing for infrequent tokens (ablation) | Mixtape · One Billion Word | perplexity: 36.77 -> 36.52 | No clear effect |
| add gate sharing for infrequent tokens (ablation) | Mixtape · Penn Treebank | training time per instance, all layers, same memory (s): 10.44 -> 3.66 | Improved |
| add gate sharing for infrequent tokens (ablation) | Mixtape · One Billion Word | training time per instance, all layers, same memory (s): 48.6 -> 11.3 | Improved |
| replace MoS-15 with Mixtape output layer | Penn Treebank LSTM · Penn Treebank | perplexity: 56.14 -> 56.37 | No clear effect |
| replace MoS-7 with Mixtape output layer | One Billion Word LSTM · One Billion Word | perplexity: 37.10 -> 36.52 | Improved |
| replace softmax with Mixtape output layer | LSTM · Penn Treebank | perplexity: 59.19 -> 56.37 | Improved |
| replace softmax with Mixtape output layer | LSTM · One Billion Word | perplexity: 42.77 -> 36.52 | Improved |
| replace softmax with Mixtape output layer | Transformer · WMT14 En-De | BLEU: 29.0 -> 29.3 | Improved |
| replace softmax with Mixtape output layer | Transformer · WMT14 En-Fr | BLEU: 43.0 -> 43.9 | Improved |
| replace MoS-9 with Mixtape output layer | Transformer · WMT14 En-De | BLEU: 29.5 -> 29.3 | Hurt |
| replace MoS-9 with Mixtape output layer | Transformer · WMT14 En-Fr | BLEU: 42.1 -> 43.9 | Improved |
Closest papers by meaning
- Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free (NeurIPS 2025 · similarity 0.84)
- Streamlining Redundant Layers to Compress Large Language Models (ICLR 2025 · similarity 0.83)
- SkipGPT: Each Token is One of a Kind (ICML 2025 · similarity 0.83)
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.83)
- The Sparsity Whisperer (NeurIPS 2026 · similarity 0.83)
- Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization (NeurIPS 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.
