Project Ariadne investigates whether an embodied AI facilitator enhances collaborative problem-solving in virtual reality.

We designed a controlled study with two conditions, one with AI facilitation and one without, in which teams of 4 participants work together on an information-asymmetric escape-room task in a shared VR environment. Our evaluation metrics include:

• group cohesion (measured with GCS-R)
• memory pooling (measured via TMS coordination + specialization subscales)
• level of turn-taking
• cognitive load (measured via NASA-TLX)
• perceived agent utility (measured via post-questionnaire Likert scale rankings)

The implementation employs a dual-agent Talker-Thinker architecture that combines real-time conversational responsiveness with periodic strategic reasoning. Core technical components include streaming speech-to-text and synthesis, avatar animation control, WebSocket-based session management, and a persistent belief store for facilitation context. This work contributes evidence-based design guidance for human-AI collaboration in immersive educational and professional settings.