Reproduce “Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chungpa Lee, Jihoon Kwon, Kyle Min, Jy-yong Sohn
The authors align image and text sparse autoencoder features after training separate models for each modality, matching latent coordinates by activation correlation while preserving learned directions. On MS-COCO image-to-text retrieval with CLIP ViT-B/32 embeddings, the method reaches 16.0 ± 0.5 Recall@1 versus 7.1 ± 0.1 for group-sparse alignment, with equal reported reconstruction MSE; semantic matches remain unverified.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.