Robots with Attitudes: Influence of LLM-Driven Robot Personalities on Motivation and Performance

Dennis Becker, Kyra Ahrens, Connor Gäde, Erik Strahl, Stefan Wermter

Introduction

In the field of Human-Robot Interaction (HRI), a robot's personality is a fundamental component of successful collaboration. Traditionally, these personalities were "hard-coded," relying on rigid, pre-scripted behaviors that often felt artificial and limited. The advent of Large Language Models (LLMs) offers a transformative opportunity to create robots capable of fluid, unscripted conversation. However, it remains unclear whether an LLM can maintain a consistent persona over extended interactions and how such personas—specifically "agreeable" versus "non-agreeable" ones—impact human likability, motivation, and task performance.

Methodology

Researchers at the University of Hamburg investigated these questions using a two-part study design centered on the "Big Five" personality trait of Agreeableness.

Experimental design overview showing the LLM-driven personality assignment and the interactive drawing game setup.

The study first utilized an online pre-study to validate persona consistency. Using Vicuna (an open-source model based on Meta’s LLaMA), the team generated dialogues for two distinct personas: an encouraging "Art Teacher" and a snarky, non-agreeable variant.

In the main lab-based study, 63 participants interacted with NICO, a humanoid robot. The interaction revolved around a game of Quickdraw, where participants had 30 seconds to draw an object on a touchscreen while NICO attempted to recognize it in real-time using a neural network trained on the Quickdraw dataset. Between rounds, NICO—randomly assigned either an "Agreeable" (supportive, polite) or "Non-agreeable" (disinterested, snarky) persona—provided feedback and engaged in open-ended conversation. To enhance the persona, the robot's physical gestures (e.g., nodding, facepalming) were synchronized with its verbal output and the game's progress.

Results and Discussion

The study confirmed that LLMs are highly effective engines for maintaining consistent robotic personas. Participants accurately identified the intended traits, and Ten-Item Personality Inventory (TIPI) scores showed clear differentiation across Agreeableness, Conscientiousness, and Emotional Stability.

Comparison of participant perceptions for agreeable vs. non-agreeable robots across various measures, highlighting significant differences in likeability and safety.

The "Halo Effect" of Agreeableness

The agreeable robot was rated significantly higher on the Godspeed Likability scale ($W=837, p < .001$). Interestingly, it was also perceived as significantly safer ($p = .039$), despite the physical robot and environment being identical in both conditions. This suggests that "rude" or unpredictable verbal behavior in robots can be interpreted by humans as a lack of mechanical control or predictability.

Motivation and Performance

Contrary to expectations, the robot's assigned personality did not significantly alter participants' intrinsic motivation; many participants appeared to "brush off" the non-agreeable robot's snark. However, a significant correlation emerged between perceived personality and performance. A positive Kendall rank correlation was found between perceived robot agreeableness ($\tau = .248, p = .009$) and task scores. This indicates that the subjective "vibe" of the interaction may be more influential on performance than the objective prompt driving the robot.

Conclusion and Future Work

This research demonstrates that LLMs can successfully imbue robots with distinct, unscripted personalities that significantly influence human perception. While agreeableness enhances likability and perceived safety, the link to performance is mediated by individual perception.

Future research should explore the impact of these personalities over repeated interactions, as the "one-off" nature of this study might have masked long-term motivational effects. Additionally, integrating Vision-Language Models (VLMs) would allow robots to comment more naturally on the nuances of a user's work. Finally, exploring "personality matching"—pairing humans with robots of similar traits—could reveal more tailored pathways to optimizing collaborative performance.

Summary updated

BibTeX
@inproceedings{becker2025robots,
  author = {Dennis Becker and Kyra Ahrens and Connor Gäde and Erik Strahl and Stefan Wermter},
  title = {{Robots with Attitudes: Influence of LLM-Driven Robot Personalities on Motivation and Performance}},
  booktitle = {HAI},
  year = {2025},
  doi = {10.1145/3765766.3765780},
}
Figure