The Effectiveness of Style Vectors for Steering Large Language Models: A Human Evaluation

Published in IEEE Access, 2025

Recommended citation: Diallo, D., Dworatzyk, K., Jentzsch, S., Schütt, P., Theis, S., & Hecking, T. (2025). The Effectiveness of Style Vectors for Steering Large Language Models: A Human Evaluation. IEEE Access, 13, 191443-191457. https://doi.org/10.1109/ACCESS.2025.3628500

This work evaluates activation-based steering of emotional tone using more than 7,000 ratings from 190 participants. The results show that moderate steering can strengthen target emotions while preserving comprehensibility, and that automatic quality ratings broadly align with human judgments.

Read the paper

Recommended citation: Diallo, D., Dworatzyk, K., Jentzsch, S., Schütt, P., Theis, S., & Hecking, T. (2025). The Effectiveness of Style Vectors for Steering Large Language Models: A Human Evaluation. IEEE Access, 13, 191443-191457.