When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities
Sparse autoencoders (SAEs) have emerged as a promising technique for mechanistic interpretability by learning a set of sparse latent features in large models, each of which encodes a distinct concept. However, in vision-language models (VLMs), vanilla SAEs struggle to learn modality-consistent concepts, with concepts often exhibiting fragmented coverage (i.e., disjoint regions) in the visual modality. To address this challenge, we propose a Structured Sparse AutoEncoder ($S^2AE$) that enforces concept consistency from both semantic and spatial perspectives in the visual modality. Specifically,
Lineage graph
Paper → model → repo connections mined from source citations (Tier-1 exact match).
Why these links exist
Every edge carries a method, confidence, and the source snippet that justified it — so bad links are debuggable.
- PossiblePossibly related (embedding) · 52%vlm-starter →
- LinkedLinked via arxiv author · 85%Weiduo Liao →
“When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities”
- LinkedLinked via arxiv author · 85%Yunqiao Yang →
“When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities”
- LinkedLinked via arxiv author · 85%Ying Wei →
“When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities”
- PossiblePossibly related (embedding) · 49%A unifying framework from neural superposition to sparse interpretable codes →
- PossiblePossibly related (embedding) · 52%A unifying framework from neural superposition to sparse interpretable codes - Nature →
- PossiblePossibly related (embedding) · 52%Deep-Spark/DeepSparkInference →
- FuzzySimilar title/name (fuzzy) · 59%scikit-learn/scikit-learn →
“Fuzzy title match (0.73): “When Structured Sparse Autoencoders Learn Consistent Concept” ≈ “scikit-learn/scikit-learn””