LLMs Overcompress Emotion in Internal Representations
November 9, 2026
Analysis of seven LLMs on the GoEmotions dataset shows models overcompress 28 fine-grained emotion categories into broader positive and negative groupings in their hidden representations.
HOW THIS AFFECTS YOU
●
researcherYou can expect LLM embeddings to lack fine-grained emotional nuance during clustering.