Research map / Architectures and efficiency / RNNs and normalization
Features Emerge as Discrete States: The First Application of SAEs to 3D Representations
Albert Miao, Chenliang Zhou, Jiawei Zhou, Cengiz Oztireli
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 |
|---|---|---|---|
| increase SAE codebook size 256 to 1024 (ablation) | BatchTopK SAE · Dora-VAE SAE variations | relative l2 reconstruction (k=8): 0.366 -> 0.356 (k=8, 256 -> 1024) | Improved |
| increase SAE codebook size 256 to 1024 (ablation) | BatchTopK SAE · Dora-VAE SAE variations | universality (k=8): 0.421 -> 0.208 (256 -> 1024) | Hurt |
| increase threshold k from 4 to 16 (ablation) | BatchTopK SAE (codebook 512) · Dora-VAE SAE variations | relative l2 reconstruction: 0.507 -> 0.187 | Improved |
| increase threshold k from 4 to 16 (ablation) | BatchTopK SAE (codebook 512) · Dora-VAE SAE variations | Dora-VAE loss: 0.427 -> 0.216 | Improved |
| increase codebook size 256 to 1024 at k=16 (ablation) | BatchTopK SAE · Dora-VAE SAE variations | Dora-VAE loss: 0.228 -> 0.239 | Mixed |
| include log loss difference as predictor of transition point (ablation) | OLS regression per feature · Transition point prediction | partial R^2: 0.188 vs 0.042 (avg value) and 0.025 (density) at n_abl>=500 | Improved |
Closest papers by meaning
- Sparse Autoencoders, Again? (ICML 2025 · similarity 0.89)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.89)
- The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws (NeurIPS 2026 · similarity 0.88)
- Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression (NeurIPS 2026 · similarity 0.87)
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (ICLR 2025 · similarity 0.87)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (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.
