Research map / Architectures and efficiency / Attention and transformers
Deep Equilibrium Models
Shaojie Bai, J. Zico Kolter, Vladlen Koltun
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 recurrent sequence modeling with a DEQ self-attention transformer | Copy Memory T=400 · Copy Memory T=400 | loss: 3.5e-6 vs 0.0501; 3.5e-6 vs 0.0491 | Improved |
| Use a DEQ TrellisNet instead of a 60-layer TrellisNet | TrellisNet · Penn Treebank word-level test | perplexity: 57.0 -> 57.1 | No clear effect |
| Use a DEQ TrellisNet instead of a 70-layer TrellisNet | TrellisNet · WikiText-103 word-level test | perplexity: 29.2 -> 29.0 | Improved |
| Use a DEQ transformer instead of a layer-stacked Transformer-XL | Transformer-XL medium, 16 layers · WikiText-103 word-level test | perplexity: 24.3 -> 24.2 | Improved |
| Use a DEQ transformer instead of a layer-stacked adaptive-embedding Transformer-XL | Transformer-XL medium, adaptive embedding · WikiText-103 word-level test | perplexity: 23.6 -> 23.2 | Improved |
| Use a DEQ transformer instead of a weight-tied Transformer-XL | Transformer-XL small · WikiText-103 word-level test | perplexity: 34.9 -> 32.4 | Improved |
| Use the DEQ formulation instead of conventional layer stacking | DEQ models versus deep networks · Language modeling memory benchmarks | memory: up to 88% | Improved |
| Increase equilibrium tolerance or lower the Broyden iteration limit (ablation) | DEQ-Transformer · WikiText-103 | performance | Hurt |
| Set the equilibrium tolerance too high (ablation) | DEQ · Language modeling | accuracy | Hurt |
| Add dropout (ablation) | DEQ · Root-finding convergence | Broyden convergence speed | Hurt |
| Use weight normalization (ablation) | DEQ · Root-finding convergence | Broyden iteration growth | Improved |
| Pretrain a shallow weight-tied stack before DEQ training (ablation) | DEQ · DEQ training | training efficiency | Improved |
Closest papers by meaning
- Separation and Bias of Deep Equilibrium Models on Expressivity and Learning Dynamics (NeurIPS 2024 · similarity 0.90)
- Positive Concave Deep Equilibrium Models (ICML 2024 · similarity 0.89)
- FFN Fusion: Rethinking Sequential Computation in Large Language Models (NeurIPS 2025 · similarity 0.86)
- Search for Efficient Large Language Models (NeurIPS 2024 · similarity 0.86)
- Skip a Layer or Loop It? Learning Program-of-Layers in LLMs (ICML 2026 · similarity 0.86)
- Chain-of-Model Learning for Language Model (NeurIPS 2025 · 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.
