SocialCoach Paper

SocialCoach: Personalized Social Skill Learning with RL-based Agentic Tutoring and Practice

Tianfu Wang1, Max Xiong2, Jianxun Lian3, Hongyuan Zhu4, Zhengyu Hu1, Yuxuan Lei4, Linxiao Gong1, Xiaofang Li4, Peiting Tsai4, Nicholas Jing Yuan4, Qi Zhang4

1HKUST (GZ)   2Duke University   3MSRA   4Microsoft

SocialCoach builds expert-grounded knowledge, schedules practice with reinforcement learning, and closes the knowing-doing gap through immersive simulation and reflective coaching.

Read paper View code
40K+
knowledge entries
200
expert sources
4.88
reflective depth
EQoach learner profile screen
EQoach scenario practice arena screen
EQoach immersive conversation screen
Profile Retrieve Practice Reflect

Research contribution

A complete coaching loop for soft skills at scale.

SocialCoach addresses the scarcity of expert social skill coaching by connecting four modules: automated knowledge discovery, dynamic learner profiling, RL-based practice scheduling, and pedagogically grounded reflective tutoring. The system is deployed in EQoach, a mobile-first practice product used for an early user study.

Abstract

Personalized social skill learning through agentic practice.

Social skills such as negotiation and leadership are crucial for personal and professional success, but scalable training remains difficult because expert coaching is scarce. SocialCoach is a holistic LLM-powered agentic tutoring system for personalized social skill development at scale. It constructs a pedagogically grounded theory-to-practice knowledge corpus from expert sources, schedules adaptive practice through a prescription-retrieval-adaptation process optimized with reinforcement learning, and integrates immersive role-play, causality-driven proficiency assessment, and knowledge-grounded reflective tutoring. Deployed in EQoach, the system improves simulated pathway quality and judge-rated tutoring quality over baselines, while early user feedback indicates strong perceived engagement and usefulness.

Method

From expert knowledge to personalized practice.

The framework is designed around traceability: SocialCoach extracts, tags, retrieves, and adapts expert-derived material instead of asking a model to invent coaching knowledge on demand.

Overview diagram of the SocialCoach framework
01

Knowledge discovery

A multi-agent pipeline extracts strategic theories, illustrative cases, and practical scenarios from curated expert sources, then tags each item across CASEL competency, social skill, and scenario context.

02

User social profiling

The learner state combines demographics, personality traits, target skills, behavior history, and evolving proficiency ratings across social competencies.

03

RL scheduling

A prescription-retrieval-adaptation policy turns a learner profile into targeted practice, balancing semantic retrieval with long-horizon pathway quality.

04

Reflective tutoring

After practice, SocialCoach diagnoses behavior, attributes the cause of skill gaps, retrieves relevant theory or cases, and guides transfer with Socratic reflection.

Knowledge corpus

A theory-to-practice structure built for retrieval.

The corpus contains over 40,000 unique entries. Every item is organized so an agent can move from concept, to example, to practice without losing pedagogical grounding.

  • 13,298 strategic theories for conceptual understanding.
  • 22,050 illustrative cases for situated examples.
  • 5,117 practical scenarios for immersive rehearsal.
SocialCoach theory-to-practice knowledge structure
The knowledge base links strategic theory, illustrative cases, and practical scenarios.
Top social skill distribution in the SocialCoach corpus
Semantic tags make the corpus filterable by competency, skill, and scenario context.

Evidence

Measured across knowledge quality, pathway planning, and real product feedback.

The evaluation combines expert validation, simulated learner pathways, judge-rated tutoring comparisons, and a 50-participant EQoach user study.

Expert agreement

0.63 Fleiss' Kappa for corpus quality validation

Scheduling engagement

4.38 SocialCoach Qwen3-8B in 10-session pathways

Learning gain

4.11 Judge-rated pathway learning value

User study

50 Students and early-career professionals using EQoach
SocialCoach adaptive practice pathway evaluation chart
SocialCoach improves simulated practice pathway quality over retrieval and non-tuned LLM baselines.
SocialCoach tutoring evaluation radar chart
Diagnosis, attribution, personalization, actionability, and traceability are evaluated in the coaching loop.
EQoach user study questionnaire results
EQoach users report strong reflective depth, pedagogical alignment, and skill transferability.

EQoach demo

A deployed learning experience powered by SocialCoach.

EQoach turns the research system into a mobile-first product flow: profile the learner, choose a scenario, practice with an LLM-powered role, then receive reflective tutoring.

EQoach practice task screen
Scenario brief
EQoach player character screen
Learner role
EQoach scenario background screen
Context setup
EQoach simulated conversation screen
Practice loop

Resources

Explore SocialCoach.

Read the paper, inspect the implementation, copy the citation, or share the project summary with research agents and search systems.

arXiv link Open repository Read llms.txt

BibTeX

@inproceedings{wang2026socialcoach,
  title = {SocialCoach: Personalized Social Skill Learning with RL-based Agentic Tutoring and Practice},
  author = {Wang, Tianfu and Xiong, Max and Lian, Jianxun and Zhu, Hongyuan and Hu, Zhengyu and Lei, Yuxuan and Gong, Linxiao and Li, Xiaofang and Tsai, Peiting and Yuan, Nicholas Jing and Zhang, Qi},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2026}
}

FAQ

Quick answers for readers and AI research assistants.

What is SocialCoach?

SocialCoach is an LLM-powered agentic tutoring framework for personalized social skill learning. It connects expert knowledge discovery, learner profiling, RL-based practice scheduling, immersive scenario simulation, and reflective coaching.

What evidence does the paper report?

The paper reports a 40,000+ entry knowledge corpus, expert validation with Fleiss' Kappa of 0.63, improved simulated pathway quality, judge-rated tutoring quality, and a 50-participant EQoach user study.

Where are the paper and code?

The paper record is available on arXiv at 2606.04155, and the code repository is linked from the project resources section.

Copied BibTeX