Research map / Architectures and efficiency / RNNs and normalization
Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Universal Sparse Autoencoders (USAEs), a single overcomplete TopK SAE with a shared concept code and per-model decoders trained jointly on DinoV2, SigLIP, and ViT final-layer activations to find cross-model visual concepts, and introduces coordinated activation maximization.
Problem
Existing concept-based interpretability methods analyze one model at a time, and post-hoc methods for finding universal features across models are costly and do not allow translation between models in a shared concept space.
Approach
Train one shared sparse code that is encoded from any one model's activations and decoded through model-specific dictionaries to reconstruct all models' activations, so concept alignment is imposed end-to-end; add coordinated activation maximization that maximizes the same concept index across models.
Key results
- Self-reconstruction R^2 averages about 0.8 for DinoV2 versus 0.7 for SigLIP and ViT, with positive off-diagonal cross-model R^2.
- Roughly 23-37% of universal concepts have cosine similarity above 0.5 with concepts from independently trained SAEs (AUC 0.13 for the random baseline).
- Energy and co-firing are strongly correlated for frequently co-firing concepts (r=0.89 vs r=0.63 overall).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply the ImageNet-trained USAE to out-of-distribution datasets | USAE · DTD and CelebA | activation reconstruction R²: 0.77 -> 0.69 (DinoV2, DTD); 0.77 -> 0.75 (DinoV2, CelebA); 0.61 -> 0.54 (SigLIP, DTD); 0.61 -> 0.52 (SigLIP, CelebA); 0. | Mixed |
| Use universal training instead of independent single-model SAE training | USAE concept learning · Vision-model concept consistency | concept overlap at cosine similarity >0.5: 23% and 26% for SigLIP and DinoV2; 38% for ViT | Mixed |
| Focus on the top 1,000 co-firing concepts | USAE versus independent SAEs · Vision-model concept consistency | average cosine similarity: 0.65 (DinoV2); 0.52 (ViT); 0.40 (SigLIP) | Helped |
| Consider concepts with more than 1000 co-fires | USAE concept universality and importance · Vision-model concepts | correlation between co-firing frequency and energy: r=0.63 vs r=0.89 | Helped |
| Replace the ℓ2 reconstruction loss with an ℓ1 loss (ablation) | TopK SAE used for USAE training · SAE training and concept interpretability | training dynamics and interpretability | Helped |
Surprising findings
- Universal training recovers only a fraction of independent SAE concepts while discovering shared representations not found in model-specific training.
- DinoV2 has the lowest co-firing proportion (0.266) versus SigLIP (0.344) and ViT (0.326), indicating it encodes more unique features.
Limitations
- Sensitivity to hyperparameters when increasing the number of models in universal training.
- Only final-layer activations are analyzed.
- A small fraction of concepts are uninterpretable.
Experimental setup
Datasets: ImageNet DTD CelebA
Benchmarks: ImageNet validation set DTD CelebA
Metrics: R^2 MSE firing entropy co-fire proportion cosine similarity concept consistency AUC
Baselines: independently trained SAEs
Builds on: TopK sparse autoencoder Maco Rosetta Neurons
Closest papers by meaning
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.91)
- Cascaded Sparse Autoencoders LearnMulti-Level Visual Concepts in Multimodal LLMs (NeurIPS 2026 · similarity 0.90)
- Mechanistic Interpretability with Sparse Autoencoder Neural Operators (NeurIPS 2026 · similarity 0.90)
- Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression (NeurIPS 2026 · similarity 0.90)
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (ICLR 2025 · similarity 0.89)
- Learning Multi-Level Features with Matryoshka 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.