Social assistive robots are potentially going to integrate with human daily life in the near future. This study presents Haru, an embodied social robot capable of engaging in multiparty conversations through both verbal and non-verbal behaviors. Unlike traditional systems that rely on rule-based control or manual scripting, Haru autonomously generates human-like gaze behaviors using Generative Adversarial Imitation Learning (GAIL) from real-world demonstrations, allowing it to coordinate attention and turn-taking in dynamic group settings. In addition, Haru integrates a large language model (i.e., ChatGPT) to support open-domain spoken dialogue, enabling flexible and context-sensitive interaction. Results of a user study show that participants preferred Haru when it maintained mutual gaze during conversation, and perceived it as more polite, socially aware, and engaging. These findings highlight the importance of combining adaptive non-verbal cues with natural language abilities to enhance human-robot interaction.