Understanding LLM Intervention Explanations in Multi-Party Human-Robot Interaction

LLMs in social robots lack transparency about intervention decisions: The study reveals that when LLM-powered robots decide to join group conversations, their explanations follow predictable patterns focused on facilitation (agreement, participation, and interaction flow) rather than providing diverse reasoning—raising questions about whether these systems truly understand context or simply follow learned patterns.

Robot roles significantly influence intervention behavior: The research found that robots assigned different conversational roles (a "mover" supporting coordination versus an "opposer" driving goal-oriented discussion) generated distinctly different types of intervention explanations, suggesting that how we program robot roles directly shapes their decision-making in group settings.

Current LLM explanations may oversimplify complex social dynamics: With only five recurring themes identified across 610 explanations in multi-party interactions, the study suggests that LLM-generated justifications for when robots should speak may be too narrow to capture the nuanced social reasoning required for genuine group participation.

Real-time explainability in group settings is still underdeveloped: This appears to be one of the first systematic studies examining how LLMs explain their intervention choices during actual multi-party conversations, indicating that responsible deployment of conversational robots requires more research into whether their real-time explanations are meaningful to human participants.

Consistency across conditions may mask important limitations: While intervention patterns remained stable between homogeneous and heterogeneous robot conditions, this stability could indicate that LLMs are applying generic rules rather than adapting to the specific social context of each group interaction.