Researchers Diallo et al. (2025) identified specific activation axes within the Llama-3-8B model corresponding to six Ekman emotions: anger, disgust, fear, joy, sadness, and surprise. By applying a shift vector multiplied by a coefficient lambda to each transformer layer, they successfully steered the emotional tone of the model's responses across 19 different task types.
In a human evaluation involving 190 participants, the lambda coefficient was adjusted from 0 to 0.35. Results showed that disgust and fear ratings increased significantly as the shift intensified. While anger, joy, and sadness peaked at a lambda of 0.30, surprise ratings decreased. The study also noted that emotional axes often overlapped, with shifts in one emotion triggering changes in others.
The study assessed response clarity on a scale of 0 to 7. At a lambda of 0.15, clarity remained comparable to the baseline. However, at higher values, particularly at 0.35, clarity dropped significantly, especially for disgust and surprise. Beyond a lambda of 0.35, the generated texts frequently lost coherence, indicating a limit to effective emotional steering.
The findings suggest that these emotional forms belong to both affective and stylistic-rhetorical classes within the model's activation space. This research provides a framework for understanding how internal model states can be manipulated to produce specific emotional outputs without compromising the underlying architecture of the large language model.
