Large language models (LLMs) have evolved into LLM agents that can use the user conversation context and respond according to the roles of the LLM agents. Recent studies have suggested that LLM-based agents can be used as human-like partners in social interactions. However, the role of the environmental context, particularly spatial information of user space, in the interaction between humans and LLM agents has not been explored. In this study, participants engaged in counselling conversations under three different conditions based on their spatial awareness levels. The dependent measures included copresence, trust, therapist alliances, and self-disclosure. The results suggested that participants in the condition where the LLM actively reflected spatial information generally reported higher levels of user experience. Interestingly, when the LLM actively reflected the spatial context of the user, the participants tended to describe themselves and express their emotions more. These findings suggest that spatially aware LLM agents can contribute to better social interactions between humans and LLM agents. Our findings can be used to design future augmented reality applications in the counselling, education, and healthcare industries.

Emotion detection using features presented in the body has been comparatively understudied compared to other emotional modalities. This study investigated and compared how emotions are revealed through bodily sensations and body movement information. We propose a novel visualization method for addressing body part activation or deactivation associated with different emotions using motion capture data (bodily motion maps; BMMs) and aim to compare its emotional features with existing methods of addressing sensation towards body activation or deactivation using the computerised self-report method (bodily sensation maps; BSMs). A user study is conducted with twenty-nine participants to gather BSMs and BMMs under seven emotional statuses (happy, sad, surprise, angry, disgust, fearful, and neutral). The results indicate that people's BMMs as well as their BSMs contain shared features that can discriminate between emotions. The asynchronies between bodily sensation and body motion existed. Furthermore, convolutional neural network (CNN)-based automatic emotion recognition outperforms humans’ emotion recognition abilities. The implications of the current study's method of comparing emotional features between bodily sensation and body motion for future emotion-related studies for mental disorders are discussed.