Research map / Architectures and efficiency / RNNs and normalization
On the Limits of Sparse Autoencoders: A Theoretical Framework and Reweighted Remedy
Jingyi Cui, Qi Zhang, Yifei Wang, Yisen Wang
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 |
|---|---|---|---|
| reweight SAE reconstruction loss toward high-variance (monosemantic) dimensions, alpha=1 | TopK SAE on Pythia-160M · Auto-interpretability, layers 0-11 | auto-interpretability score (%): average +3.8% gain; layer 3: 77.8 -> 84.6 | Improved |
| use alpha=1 instead of alpha=0.5 in the weighting exponent (ablation) | Weighted SAE on Pythia-160M · Auto-interpretability, layers 9-11 | auto-interpretability score (%): layer 10: 76.1 (alpha=0.5) vs 72.4 (alpha=1); layer 3: 77.9 vs 84.6 | Mixed |
| add reweighted loss to SAE with grouped Top-1 TopK variant | SAE on Pythia-160M · Auto-interpretability, layers 9-11 | auto-interpretability score (%): +2.5, +1.9, +2.7 (layers 9, 10, 11) | Improved |
| add reweighted loss to SAE | SAE on Llama-3-8B · Auto-interpretability, layers 8-32 | auto-interpretability score (%): +4.5, +2.8, +3.8, +2.2 (layers 8, 16, 24, 32) | Improved |
| weight SAE loss by semantic consistency, alpha>0 | TopK SAE on NCL ResNet-18 (ImageNet-100) · Semantic consistency | semantic consistency (%) | Improved |
| vary alpha in {0.5,...,2.5} versus alpha=0 (plain SAE) (ablation) | WSAE on NCL ResNet-18 · Semantic consistency sensitivity | semantic consistency | Improved |
| increase ground-truth feature sparsity in synthetic data (ablation) | standard SAE · Synthetic toy superposition (n=200, n_p=20) | avg number of ground-truth features activated per latent | Improved |
| reweight loss in WSAE on synthetic data at low sparsity | SAE · Synthetic toy superposition | ground truth reconstruction error | Improved |
Closest papers by meaning
- Improving Sparse Decomposition of Language Model Activations with Gated Sparse Autoencoders (NeurIPS 2024 · similarity 0.93)
- Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words (ICLR 2025 · similarity 0.92)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.92)
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.91)
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.91)
- Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? (NeurIPS 2026 · similarity 0.91)
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.
