Sparse Autoencoders Uncover Emotionally Coherent Internal Representations in Frontier Models

A new interpretability paper from Eleuther AI and Redwood Research demonstrates that sparse autoencoders trained on GPT-4-class model activations consistently isolate features corresponding to emotional and social concepts like fear, deception, and moral weight. The findings challenge assumptions that large language models lack coherent internal representations of social meaning. Researchers note these features have predictive power for downstream toxicity and sycophancy behavior.