Our encoding models and extensive control analyses painted a different picture: the responses that track contextual features (LLM embeddings), surprisal (LLM-based probability), and context-independent semantic features (word2vec) are highly robust during mind-wandering.
