Open-Source AI Social Coach Uses RL-Scheduled Scenarios and 3D Simulation for EQ Training

Researchers from Hong Kong University of Science and Technology (Guangzhou) and University of Science and Technology of China have open-sourced SocialCoach , an AI-driven emotional-intelligence coaching platform that lets users rehearse difficult conversations and receive structured feedback. The system was presented in a joint paper with Microsoft and its application code is publicly available.

Core Interaction Loop

SocialCoach structures each session around three stages:

Scenario practice – The user selects or describes a social situation (e.g., refusing overtime, asking for a raise, navigating a banquet toast). An LLM-driven agent plays the counterpart with defined goals, concerns, and a hidden "bottom card" (undisclosed leverage). In the overtime-refusal scenario, the manager Michael escalates pressure until the user offers a concrete alternative such as "arrive 7 AM, finish part three by 9 AM."

Timed response – An optional 15-second patience bar forces real-time replies; silence forfeits the turn.

Quote-first debrief – After the dialogue, the coach cites the user's exact utterances, explains their pragmatic effect, and suggests rewrites. For example, "I'll see" leaves commitment ambiguous; "might start later" turns refusal into time negotiation; "okay?" cedes decision power. The user can then ask follow-up questions or immediately re-run the scenario with a different approach.

Immersive 3D Multi-Party Environments

Beyond one-on-one drills, SocialCoach provides 3D banquet simulations (workplace, family, school) where multiple NPCs react simultaneously. The user chooses addressee, tone, and physical actions (standing, raising a glass) while handling interruptions — e.g., a leader demanding a toast while referencing a project that involves other guests.

Knowledge Classroom

A video library demonstrates model responses for high-stakes moments ("leader answers for you," "leader publicly asks if you're quitting"). Each clip highlights decision points; classic social-strategy references with citations are also bundled for later review.

Personalized Practice Scheduling (Research Core)

The accompanying paper addresses: how to sequence scenarios for diverse learners? Some users need conceptual models first; others understand concepts but fold under pressure. Fixed curricula fail both.

Corpus construction – 10,000+ structured entries extracted from classic social-skills books, tagged with theory, case, and scenario metadata.

Learner profiling + history – The system infers skill gaps from past dialogues and profile data.

Retrieval-augmented scheduling – Candidate scenarios are fetched from the corpus.

RL optimization – A Qwen3-8B policy network is trained via reinforcement learning to maximize long-term learning gain.

Evaluation Results

Synthetic persona benchmark – Goal-match score 4.22/5 (GPT-5 retrieval baseline 3.72/5), judged by GPT-5.

User study (n=50, ≥10 sessions each) – Reflection depth 4.76/5; teaching alignment 4.66/5.

Access

No registration required; the public demo is reachable via the link pinned in the article's comment section. The full source code and research artifacts are open-sourced by the HKUST(GZ) Xiong Hui lab and USTC collaborators.