Beyond Prompting: Robust Personality Expression in LLM-Driven Social Robots

Sammanfattning

This thesis investigates how personality conditioning methods influence dialogue and generated behavior in language model-driven social robots. In human-robot interaction, personality is an important factor for user experience, yet it remains unclear how different conditioning methods affect personality expression in multi-turn dialogue and related behaviors such as gesture generation. To address this, three conditioning methods are compared: prompt-based conditioning, parameter-efficient fine-tuning, and activation steering. These methods are evaluated using small language models across simulated mulit-turn conversations. Personality expression is analyzed using a text-based classifier, while dialogue quality is assessed through the presence of artifacts such as meta-responses and empty outputs. Gesture generation is evaluated based on diversity, distribution and congruence between intended and produced gesture for the robot embodiment. The results show that personality alignment remains relatively weak across all methods, with more model-invasive methods showing no clear improvement over prompt-based conditioning in multi-turn dialogue. However, conditioning methods have a strong impact on system behavior. In particular, fine-tuning and steering introduce significant artifacts in dialogue, leading to degraded interaction quality. Conditioning methods are also found to strongly affect gesture generation, influencing both gesture diversity and consistency. Overall, the findings indicate that personality conditioning has limited effectiveness in controlling explicit personality traits outside controlled settings, but significantly shape both dialogue quality and nonverbal behavior. These insights contribute to the design of more consistent and reliable personality expression in language model-driven agents.