Bhardwaj Sangnie

Linear representations in language models can change dramatically within conversations, with factual information becoming non-factual and vice versa based on conversational context, posing challenges for interpretability and steering approaches.

“Google researcher who contributed to study on linear representations in language models.”