Research map / Architectures and efficiency / RNNs and normalization
Improving Sparse Autoencoder with Dynamic Attention
Dongsheng Wang, Jinsen Zhang, Dawei Su, Hui Huang
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 sparsemax attention (replace ReLU) in transformer SAE (ablation) | Transformer SAE · ImageNet zero-shot, on_1 | accuracy: 3.86 -> 10.93 | Improved |
| add sparsemax attention (replace ReLU) in transformer SAE (ablation) | Transformer SAE · ImageNet zero-shot, on_50 | accuracy: 36.33 -> 59.95 | Improved |
| add transformer cross-attention architecture (replace MLP) with sparsemax (ablation) | Sparsemax SAE · ImageNet zero-shot, on_1 | accuracy: 7.91 -> 10.93 | Improved |
| add transformer cross-attention architecture (replace MLP) with sparsemax (ablation) | Sparsemax SAE · ImageNet zero-shot, on_50 | accuracy: 55.32 -> 59.95 | Improved |
| add transformer architecture with ReLU to ReLU SAE (ablation) | ReLU SAE · ImageNet zero-shot, on_10 | accuracy: 22.17 -> 24.08 | Improved |
| add sparsemax to MLP SAE (vs ReLU SAE) (ablation) | MLP SAE · ImageNet zero-shot, on_10 | accuracy: 22.17 -> 39.73 | Improved |
| set K by the average sparsemax activation count (K=24 instead of 32) (ablation) | BatchTopK · ImageNet zero-shot, on_10 | accuracy: 48.99 -> 49.52 | Improved |
| set K=24 instead of 32 (ablation) | TopK · ImageNet zero-shot, on_1 | accuracy: 8.64 -> 0.99 | Hurt |
| set K=24 instead of 32 (ablation) | TopK · ImageNet zero-shot, on_10 | accuracy: 30.21 -> 42.88 | Improved |
| use Sparsemax SAE instead of BatchTopK (K=24) | SAE · OpenWebText NMSE (M=3072) | NMSE: 0.014 -> 0.005 | Improved |
| use Sparsemax SAE instead of BatchTopK | SAE · OpenWebText CE degradation (M=3072) | CE degradation: 0.196 -> 0.031 | Improved |
Closest papers by meaning
- Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? (NeurIPS 2026 · similarity 0.91)
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.90)
- Mechanistic Interpretability with Sparse Autoencoder Neural Operators (NeurIPS 2026 · similarity 0.90)
- Learning Multi-Level Features with Matryoshka Sparse Autoencoders (ICML 2025 · similarity 0.90)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.90)
- Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders (ICML 2025 · 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.
