Improving Neuron-level Interpretability with White-box Language Models

Hao Bai, Yi Ma

Neurons in auto-regressive language models like GPT-2 can be interpreted by analyzing their activation patterns. Recent studies have shown that techniques such as dictionary learni…