Research & Papers

SocialCoach uses RL-based agentic tutoring to teach social skills at scale

LLM-powered system personalizes negotiation and leadership training with adaptive practice scheduling.

Deep Dive

SocialCoach tackles the scarcity of expert coaching for soft skills like negotiation and leadership. Built with a multi-agent pipeline, it automatically constructs a pedagogically-grounded knowledge corpus from diverse expert sources, ensuring a theory-to-practice foundation. To personalize learning, it employs an adaptive practice scheduling module using a prescription-retrieval-adaptation process, optimized via reinforcement learning in a learner simulation environment to maximize long-term learning and overcome cold-start problems.

The system then integrates immersive, goal-driven practice with causality-driven proficiency assessment and knowledge-grounded reflective tutoring to bridge the knowing-doing gap. Deployed in the product EQoach, extensive experiments show SocialCoach improves simulated pathway quality and judge-rated tutoring quality over baseline approaches. Early user feedback indicates strong perceived engagement and usefulness, suggesting a practical architecture for scalable, gamified soft skill learning platforms.

Key Points
  • SocialCoach automatically constructs a pedagogically-grounded knowledge corpus from expert sources using a multi-agent LLM pipeline.
  • Reinforcement learning optimizes an adaptive practice scheduling module for personalized learning journeys, handling cold-start issues.
  • Deployed in product EQoach, it improves simulated pathway quality and judge-rated tutoring quality, with early users reporting high engagement.

Why It Matters

Scalable AI tutoring for soft skills could transform professional development, reducing reliance on rare human coaches.