YusongZhao666/csae-ckpt

SAE++ Checkpoint

This repository contains a trained checkpoint of the SAE++ model introduced in the paper SAE++: Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs.

SAE++ is a cascaded sparse autoencoder architecture that learns multi-level visual concepts from MLLM activations. The Level-1 SAE recovers fine-grained concepts, and a Level-2 SAE is trained directly on the Level-1 decoder weights to learn higher-level "concepts of concepts." This improves hierarchical interpretability and enables group-level steering of MLLM outputs.

For usage, training scripts, evaluation, and steering demos, please refer to the official code repository: https://github.com/Wang-ML-Lab/sae-plus-plus.