Reproduce “SAE++: Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs” 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

Yusong Zhao, Hengyi Wang, Tanuja Ganu, Akshay Nambi, Hao Wang

SAE++ organizes visual features in multimodal models into hierarchies by learning higher-level groups over the first autoencoder’s feature directions rather than activation codes. It reports higher hierarchical mono-semanticity scores than baselines across three model families and four datasets. On ImageNet, the mean score is 0.7776, versus 0.5843 for Matryoshka SAE. This measures visual similarity, not general interpretability.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.