Research map / Architectures and efficiency / Attention and transformers
Affine-Scaled Attention: Towards Flexible and Stable Transformer Attention
Jeongin Bae, Baeseong park, Gunho Park, Minsub Kim, Joonhyung Lee, Junhee Yoo, Sunghyeon Woo, Jiwon Ryu, Se Jung Kwon, Dongsoo Lee
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 Affine-Scaled Attention | 3B student · C4 validation | perplexity: 44.9 -> 40.8 | Improved |
| add Affine-Scaled Attention | 1B student · C4 validation | perplexity: 61.9 -> 58.9 | Improved |
| add Affine-Scaled Attention | 0.5B student · C4 validation | perplexity: 71.7 -> 70.6 | Improved |
| add Affine-Scaled Attention | 3B student · commonsense reasoning average accuracy | accuracy: 60.8 -> 62.4 | Improved |
| add Affine-Scaled Attention | 1B student · commonsense reasoning average accuracy | accuracy: 56.7 -> 57.2 | Improved |
| add Affine-Scaled Attention | 0.5B student · commonsense reasoning average accuracy | accuracy: 50.6 -> 51.3 | Improved |
| add Affine-Scaled Attention | 3B student · ARC-Challenge and BoolQ | accuracy: ARC-c improves from 38.4 to 41.5 and BoolQ from 63.6 to 65.1 | Improved |
| use attention sink instead of baseline softmax | 0.5B student · C4 validation | perplexity: 71.7 -> 81.5 | Hurt |
| use attention sink instead of baseline softmax | 3B student · C4 validation | perplexity: 44.9 -> 48.5 | Hurt |
| add Affine-Scaled Attention on top of Gated Attention | 3B student with Gated Attention · commonsense reasoning average accuracy | accuracy: 61.6 -> 63.1 | Improved |
| add Affine-Scaled Attention on top of Gated Attention | 1B student with Gated Attention · commonsense reasoning average accuracy | accuracy: 57.1 -> 58.0 | Improved |
| add Affine-Scaled Attention on top of Gated Attention | 3B student with Gated Attention · C4 validation | perplexity: 44.6 -> 42.7 | Improved |
Closest papers by meaning
- Theory, Analysis, and Best Practices for Sigmoid Self-Attention (ICLR 2025 · similarity 0.90)
- The Hedgehog & the Porcupine: Expressive Linear Attentions with Softmax Mimicry (ICLR 2024 · similarity 0.90)
- Hierarchical Self-Attention: Generalizing Neural Attention Mechanics to Multi-Scale Problems (NeurIPS 2025 · similarity 0.89)
- Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free (NeurIPS 2025 · similarity 0.89)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.89)
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · similarity 0.89)
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.
