Research map / Vision-language and multimodal / Human and embodied modeling
Evaluating the Interpretability of Sparse Autoencoders with Concept Annotations
Jonas Klotz, Cassio F. Dantas, Pallavi Jain, Diego Marcos, Begüm Demir
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 |
|---|---|---|---|
| use many-to-one FBMP coalitions (k=3) instead of one-to-one F1 matching (ablation) | SAEs (BatchTopK, Matryoshka, TopK, JumpReLU) on CLIP · CUB and COCO matching | Delta MATCHScore | Improved |
| use F_0.5 instead of F_1 or F_0.25 as the FBMP selection criterion (ablation) | FBMP · CUB and COCO matching | Delta MATCHScore | Improved |
| increase dictionary size (ablation) | TopK SAE (CLIP, CUB) · synCUB | TAPAScore | Hurt |
| increase dictionary size beyond 1024 (ablation) | BatchTopK, TopK, JumpReLU SAEs (CLIP, COCO) · synCOCO | TAPAScore | Hurt |
| increase dictionary size (ablation) | BatchTopK and Matryoshka SAEs (CLIP, CUB) · CUB | Delta MATCHScore | Mixed |
| increase dictionary size to 4096 (ablation) | TopK SAE (CLIP, CUB) · CUB | Delta MATCHScore | Mixed |
| use trained SAE instead of untrained or random activations, measured with FBMP matching and TAPAScore | sanity check on CUB and COCO · CLIP embeddings | metric separation | Improved |
| compare existing proxy metrics (MS, FMS, CKNNA) across trained vs untrained SAE | MS, FMS, CKNNA · CUB and COCO sanity check | metric sensitivity | No clear effect |
| matching score as predictor of TAPAScore (correlation) | FBMP F0.5 · synCUB | Pearson correlation | Improved |
| matching score as predictor of TAPAScore (correlation) | FBMP and F1 k=1 · synCOCO | Pearson correlation | Hurt |
| use Matryoshka nested dictionaries (ablation) | SAE variant · synCOCO | TAPAScore | Mixed |
Closest papers by meaning
- Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (ICML 2025 · similarity 0.92)
- Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models (ICML 2025 · similarity 0.91)
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (ICLR 2025 · similarity 0.90)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.90)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.90)
- Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models (NeurIPS 2025 · similarity 0.90)
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.
