Research map / Architectures and efficiency / RNNs and normalization
Dynamic Evaluation of Neural Sequence Models
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 |
|---|---|---|---|
| Increase segment length from BPTT 1 to BPTT 5 (ablation) | LSTM dynamic evaluation with SGD · Penn Treebank test | perplexity: 76.2 -> 75.6 | Helped |
| Add a global prior (ablation) | LSTM dynamic evaluation with SGD, BPTT 5 · Penn Treebank test | perplexity: 75.6 -> 74.8 | Helped |
| Replace SGD with RMS gradient scaling (ablation) | LSTM dynamic evaluation with global prior, BPTT 5 · Penn Treebank test | perplexity: 74.8 -> 72.2 | Helped |
| Scale global-prior decay by normalized RMS gradients (ablation) | LSTM dynamic evaluation with RMS scaling and BPTT 5 · Penn Treebank test | perplexity: 72.2 -> 71.7 | Helped |
| Apply dynamic evaluation | AWD-LSTM · Penn Treebank test | perplexity: 57.7 -> 51.1 | Helped |
| Apply dynamic evaluation | LSTM · WikiText-2 test | perplexity: 103.4 -> 59.8 | Helped |
| Apply dynamic evaluation | AWD-LSTM · WikiText-2 test | perplexity: 66.1 -> 44.3 | Helped |
| Apply sparse dynamic evaluation using a 250k-parameter adaptation matrix | mLSTM · Hutter Prize test | cross-entropy (bits/char): 1.24 -> 1.13 | Helped |
| Apply full dynamic evaluation | mLSTM · Hutter Prize test | cross-entropy (bits/char): 1.24 -> 1.08 | Helped |
| Apply dynamic evaluation | mLSTM · text8 test | cross-entropy (bits/char): 1.27 -> 1.19 | Helped |
Closest papers by meaning
- Minimum Divergence vs. Maximum Margin: an Empirical Comparison on Seq2Seq Models (ICLR 2019 · similarity 0.86)
- Asymptotics of SGD in Sequence-Single Index Models and Single-Layer Attention Networks (NeurIPS 2025 · similarity 0.86)
- Trainable Decoding of Sets of Sequences for Neural Sequence Models (ICML 2019 · similarity 0.86)
- Parallelizing non-linear sequential models over the sequence length (ICLR 2024 · similarity 0.85)
- Incremental Sequence Classification with Temporal Consistency (NeurIPS 2025 · similarity 0.85)
- Understanding and Improving Length Generalization in Recurrent Models (ICML 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.