Research map / Architectures and efficiency / Attention and transformers
Robust Filter Attention: Self-Attention as Precision-Weighted State Estimation
Peter Racioppo
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 precision-weighted RFA attention instead of RoPE | RoPE · WikiText-103 (eval L=4096) | test perplexity: 72.69 -> 38.46 | Improved |
| add spectral coupling of decay to head frequency (SC-RFA) | RFA (M1) · WikiText-103 (eval L=4096) | test perplexity: 38.46 -> 37.19 | Improved |
| add DLE-derived precision kernel (ablation) | SC-RFA (M2) · WikiText-103 (eval L=4096) | test perplexity: 62.83 -> 37.19 | Improved |
| add multiplicative precision term P_dt (ablation) | SC-RFA (M2) · WikiText-103 (eval L=512) | test perplexity: 27.65 -> 27.54 | Improved |
| add value rotation and counter-rotation (ablation) | SC-RFA (M2) · WikiText-103 (eval L=4096) | test perplexity: 463.29 -> 37.19 | Improved |
| add rotations (ablation) | SC-RFA (M2) · WikiText-103 (eval L=512) | test perplexity: 28.58 -> 27.54 | Improved |
| add rotations (ablation) | SC-RFA (M2) · WikiText-103 (eval L=4096) | test perplexity: 26.83 -> 37.19 | Hurt |
| use power-law robust weight instead of exponential weight (ablation) | SC-RFA (M2) · WikiText-103 (eval L=4096) | test perplexity: 33.51 -> 37.19 | Hurt |
| use power-law robust weight instead of exponential weight (ablation) | SC-RFA (M2) · WikiText-103 (eval L=512) | test perplexity: 27.98 -> 27.54 | Improved |
| add decay and uncertainty to pure rotation (ablation) | Pure rotation (M2.6) · WikiText-103 (eval L=4096) | test perplexity: 131.29 -> 37.19 | Improved |
| increase damping b from 0.05 to 5.0 (ablation) | SC-RFA · WikiText-103 (eval L=4096) | test perplexity: 37.19 -> 28.16 | Improved |
| increase damping b from 0.05 to 5.0 (ablation) | SC-RFA · WikiText-103 (eval L=512) | test perplexity: 27.54 -> 27.91 | Hurt |
Closest papers by meaning
- Transformers Learn Nonlinear Features In Context: Nonconvex Mean-field Dynamics on the Attention Landscape (ICML 2024 · similarity 0.87)
- Selective Attention: Enhancing Transformer through Principled Context Control (NeurIPS 2024 · similarity 0.86)
- Attention layers provably solve single-location regression (ICLR 2025 · similarity 0.86)
- One Filters All: A Generalist Filter For State Estimation (NeurIPS 2025 · similarity 0.86)
- Theory, Analysis, and Best Practices for Sigmoid Self-Attention (ICLR 2025 · similarity 0.85)
- From Attention to Activation: Unraveling the Enigmas of Large Language Models (ICLR 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.
