Research map / Architectures and efficiency / Attention and transformers
Faster Query-Key Learning Sharpens Attention in Self-Attention Models
Rahul Vashisht, Harish G. Ramaswamy
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 faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · SQuAD QA (6-layer GPT) | test performance: 56.11 -> 56.86 | No clear effect |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · SQuAD QA (6-layer GPT) | attention confidence (AC): 12.65 -> 42.95 | Improved |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · SQuAD QA (6-layer GPT) | sufficiency (lower better): 0.69 -> 0.59 | Improved |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · SQuAD QA (6-layer GPT) | comprehensiveness: 0.38 -> 0.53 | Improved |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · Subject-verb agreement (1-layer) | test performance: 92.434 -> 93.412 | No clear effect |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · Subject-verb agreement (1-layer) | attention confidence (AC): 41.67 -> 76.85 | Improved |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · Subject-verb agreement (1-layer) | sufficiency (lower better): 0.24 -> 0.15 | Improved |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · Subject-verb agreement (1-layer) | comprehensiveness: 0.24 -> 0.30 | Improved |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · HateXplain (1-layer) | test performance: 55.74 -> 56.90 | No clear effect |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · HateXplain (1-layer) | attention confidence (AC): 0.00 -> 43.5 | Improved |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · HateXplain (1-layer) | sufficiency (lower better): 0.41 -> 0.32 | Improved |
| add faster query-key learning rate (ablation) | single/multi-layer self-attention, AdamW · HateXplain (1-layer) | comprehensiveness: 0.40 -> 0.48 | Improved |
Closest papers by meaning
- Selective Attention: Enhancing Transformer through Principled Context Control (NeurIPS 2024 · similarity 0.88)
- Bayes optimal learning of attention-indexed models (NeurIPS 2025 · similarity 0.88)
- High-Dimensional Learning Dynamics of Attention-Indexed Models (NeurIPS 2026 · similarity 0.87)
- Training Dynamics of In-Context Learning in Linear Attention (ICML 2025 · similarity 0.87)
- A Theoretical Study of (Hyper) Self-Attention through the Lens of Interactions: Representation, Training, Generalization (ICML 2025 · similarity 0.87)
- Self-attention Networks Localize When QK-eigenspectrum Concentrates (ICML 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.
