Emotion is Not Just a Label: Latent Emotional Factors in LLM Processing

Researchers introduce a new framework showing that emotional tone in text systematically affects how large language models process and reason over information. They developed AURA-QA, an emotionally balanced dataset, and proposed emotional regularization techniques that improve reading comprehension performance across multiple benchmarks.

• →Emotional tone in text systematically alters attention patterns and reasoning behavior in transformer models.
• →Researchers created AURA-QA, a new question-answering dataset with emotionally balanced human-authored passages.
• →Emotional regularization framework constrains emotion-driven representational drift during model training.
• →The approach improves reading comprehension performance on both emotionally-varying and standard datasets.
• →Attention metrics like locality and entropy vary across emotions and correlate with downstream task performance.