Tag: Sparse Autoencoders
Browse our exclusive articles!
Visual Sparse Steering for Unsupervised Image Classification
Discover Visual Sparse Steering (VS2), an unsupervised method improving image classification accuracy using sparsity-guided vectors without labeled data.
Can Cross-Layer Transcoders Replace Vision Transformer Activations?
Explore how Cross-Layer Transcoders enhance interpretability and replace Vision Transformer activations for better AI model transparency.
Safe-SAIL: Fine-Grained Safety Analysis of Large Language Models
Discover Safe-SAIL, a framework using sparse autoencoders for detailed safety analysis and interpretability of large language models in critical domains.
DACO: Enhancing Safety in Multimodal Large Language Models
Discover DACO, a novel framework improving safety in multimodal large language models using dictionary-aligned concept control and sparse coding.
Boost Sparse Autoencoder Robustness with Masked Regularization
Enhance sparse autoencoder reliability using masked regularization to reduce feature absorption and improve out-of-distribution performance.