Whether, Not Which: Mechanistic Interpretability Reveals Dissociable Affect Reception and Emotion Categorization in LLMs

LLMs can detect emotional meaning beyond keywords: The research found that large language models can identify emotionally significant content with near-perfect accuracy (AUROC 1.000) even when explicit emotion words like "devastated" are removed, suggesting they process genuine emotional meaning from situational and behavioral cues.

Two separate emotion processing systems exist in LLMs: The study identified distinct mechanisms—"affect reception" (detecting that something is emotionally significant) and "emotion categorization" (labeling which specific emotion it is)—with affect reception working reliably across all tested models while emotion categorization relies more heavily on emotion keywords.

Previous emotion research may have been testing keyword detection, not emotion understanding: Prior studies claiming LLMs have "emotion circuits" used stimuli with explicit emotion keywords, leaving open the question of whether models were actually understanding emotions or just recognizing words. This study provides the first rigorous test using clinical vignettes without emotion keywords.

Findings are consistent across different model sizes and types: Results replicated across six different models (including Llama and Gemma variants ranging from 1B to 9B parameters), suggesting these emotion processing mechanisms are fundamental features of how LLMs work rather than quirks of specific models.

Larger models show better emotion categorization but keyword-dependence persists: While bigger models improved at mapping emotions to specific labels, all models still showed 1-7% accuracy drops without keywords, indicating that even advanced LLMs partially rely on explicit emotional language rather than pure semantic understanding.