New method finds cleaner mental models inside AI language systems

Original: Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability

Source: arxiv.org ↗