Title: SocialCoach: Enhancing Personalized Social Skill Acquisition Through RL-Driven Agentic Tutoring and Practice

In our increasingly interconnected society, mastering social competencies—such as negotiation and leadership—is vital for both career advancement and personal growth. Yet, achieving scalable and high-quality training remains difficult, largely due to a shortage of available expert coaches. To address this, we present SocialCoach, a comprehensive, large language model (LLM)-based agentic tutoring system designed to facilitate personalized social skill development on a large scale.

The system operates through three primary mechanisms. First, it utilizes a multi-agent pipeline to automatically generate a knowledge corpus that bridges theory and practice, grounded in pedagogical principles and drawn from various expert sources. Second, to tailor the learning experience, SocialCoach incorporates an adaptive practice scheduling module. This module follows a prescription-retrieval-adaptation workflow and is optimized via reinforcement learning within a learner simulation environment. This approach aims to enhance long-term learning outcomes while mitigating the cold-start problem. Finally, the platform closes the "knowing-doing" gap by combining immersive, goal-oriented practice sessions with causality-based proficiency assessments and reflective tutoring that is firmly rooted in the generated knowledge base.

We have integrated SocialCoach into our product, EQoach, and performed extensive evaluations. Our results demonstrate that SocialCoach outperforms baseline methods in both simulated pathway quality and tutor quality as rated by judges. Furthermore, initial feedback from users highlights high levels of perceived engagement and utility. These outcomes point to a viable architectural framework for developing gamified, personalized educational platforms focused on soft skills.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC