Research map / Optimization / Implicit bias and generalization
Optimizing Attention with Mirror Descent: Generalized Max-Margin Token Selection
Addison Kristanto Julistiono, Davoud Ataee Tarzanagh, Navid Azizan
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 |
|---|---|---|---|
| use l1.1 potential instead of l2 (GD) in mirror descent | MD for attention · IMDb, model size 3 | test accuracy: 81.66 -> 83.47 | Improved |
| use l1.1 potential instead of l2 (GD) in mirror descent | MD for attention · IMDb, model size 4 | test accuracy: 81.05 -> 83.36 | Improved |
| use l1.1 potential instead of l2 (GD) in mirror descent | MD for attention · IMDb, model size 6 | test accuracy: 82.22 -> 83.65 | Improved |
| use l1.1 potential instead of l3 | MD for attention · IMDb, model size 6 | test accuracy: 81.97 -> 83.65 | Improved |
| use l3 potential instead of l2 (GD) | MD for attention · IMDb, model sizes 3, 4, 6 | test accuracy: 81.66 -> 82.57 (size 3); 82.22 -> 81.97 (size 6) | Mixed |
| train with l1.1-MD instead of GD | transformer classifier · IMDb attention weights | number of weights >= 0.06 in W_K: 27,224 -> 18,206 | Improved |
| train with l1.1-MD instead of GD | transformer classifier · IMDb attention weights | number of weights >= 0.06 in W_V: 10,127 -> 7,643 | Improved |
| train with l1.1-MD instead of GD | transformer classifier · IMDb token selection on GPT-4o reviews | optimal token selection: 3 out of 5 sample sentences | Improved |
| train with l1.1-MD instead of Adam | ViT · CIFAR-10 | test accuracy: approximately 82% | No clear effect |
| train with lp-MD instead of Adam/SGD | ViT · CIFAR-100 | test accuracy: around 55% | No clear effect |
| train with l1.1-MD instead of Adam | ViT · CIFAR-10 and CIFAR-100 attention weights | sparsity of attention weights | Improved |
Closest papers by meaning
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · similarity 0.88)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.88)
- Bayes optimal learning of attention-indexed models (NeurIPS 2025 · similarity 0.87)
- Asymptotics of SGD in Sequence-Single Index Models and Single-Layer Attention Networks (NeurIPS 2025 · similarity 0.87)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.87)
- Attention with Trained Embeddings Provably Selects Important Tokens (NeurIPS 2025 · similarity 0.87)
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.
