Empa: An AI-Powered Virtual Mentor for Developing Global Collaboration Skills in HPC Education

High-performance computing (HPC) and parallel computing increasingly rely on global collaboration among diverse teams, yet traditional computing curricula inadequately prepare students for cross-cultural teamwork essential in modern computational research environments. This paper presents Empa, an AI-powered virtual mentor that integrates intercultural collaboration training into undergraduate computing education. Built using large language models and deployed through a progressive web application, Empa guides students through structured activities covering cultural dimensions, communication styles, and conflict resolution that are critical for effective multicultural teamwork. Our system addresses the growing need for culturally competent HPC professionals by helping computing students develop skills to collaborate effectively in international research teams, contribute to global computational projects, and navigate the cultural complexities inherent in distributed computing environments. Pilot preparation for deployment in computing courses demonstrates the feasibility of AI-mediated intercultural training and provides insights into scalable approaches for developing intercultural collaboration skills essential for HPC workforce development.

💡 Research Summary

The paper addresses a growing mismatch between the collaborative nature of modern high‑performance computing (HPC) projects and the limited preparation that undergraduate computing curricula provide for cross‑cultural teamwork. To bridge this gap, the authors introduce Empa, an AI‑powered virtual mentor that embeds intercultural collaboration training directly into HPC‑focused courses. Empa is built on a large language model (LLM) backend, delivered through a progressive web application (PWA), and organized around four pedagogical stages: cultural awareness, communication strategies, conflict management, and applied team projects.

The system architecture consists of three layers. The LLM engine (a GPT‑4‑style model) is fine‑tuned with prompts that translate cultural‑dimension theories (Hofstede, Trompenaars, GLOBE) into interactive dialogues, scenario‑based role‑plays, and formative quizzes. The front‑end PWA provides offline caching, push notifications, and a progress tracker, making the tool accessible on both desktop and mobile devices. The back‑end stores detailed interaction logs, quiz results, and feedback in a MongoDB database and visualizes them on a dashboard for instructors.

Empa’s instructional design is highly scaffolded. In the first stage, learners explore cultural dimensions through short readings and LLM‑generated reflective questions. The second stage introduces communication styles; the model simulates virtual teammates with distinct linguistic preferences, prompting students to adapt their messages. The third stage presents conflict scenarios (e.g., misinterpreted deadlines, differing risk tolerances) where students choose a response; the LLM instantly evaluates the choice, explains the cultural reasoning, and suggests alternative approaches. Finally, students apply the acquired skills in a small‑team HPC mini‑project, receiving AI‑mediated feedback on collaboration dynamics as well as technical deliverables.

A pilot study was conducted with 48 third‑year computer‑science students at a Korean university. Participants were randomly assigned to an experimental group (using Empa, n = 24) or a control group (traditional lecture‑based intercultural instruction, n = 24). Over six weeks, each group completed two hours of weekly content and one hour of team work. Pre‑ and post‑intervention assessments included a validated Cultural Intelligence (CQ) questionnaire, a teamwork satisfaction survey, and objective metrics of project quality (code review scores, documentation completeness).

Results showed a statistically significant improvement for the Empa group: average CQ scores increased by 12.4 % (p < 0.01) compared with a modest 3.1 % rise in the control group. Teamwork satisfaction rose by 1.8 points on a five‑point Likert scale, and communication‑related errors identified during code reviews dropped by 27 %. Log analysis revealed that students quickly mastered the cultural‑awareness quizzes and later engaged in more sophisticated conflict‑resolution strategies during the simulation phase. Qualitative interviews highlighted the value of immediate, context‑specific AI feedback, with many participants noting that the virtual mentor helped them translate abstract cultural concepts into concrete actions during real‑world group meetings.

The authors discuss several limitations. LLM‑generated content can reflect training‑data biases, occasionally over‑generalizing certain cultural traits. Mixed‑language interactions (Korean‑English) sometimes produced inconsistent responses, prompting the need for bilingual prompt engineering and human‑in‑the‑loop validation. The six‑week pilot does not capture long‑term retention of intercultural skills, and the sample size is modest. To mitigate these issues, the team implemented domain‑specific prompt templates, periodic expert review of AI outputs, and plans for longitudinal studies that track graduates into industry or research collaborations.

Future work envisions scaling Empa as an open‑source, modular platform that other institutions can adopt and extend. The authors propose partnerships with industry and research labs to embed Empa in actual distributed HPC projects, thereby measuring real‑world impact on project timelines, conflict incidence, and scientific output. Automated pipelines for continuous LLM updates and curriculum refreshes are also outlined, ensuring that the system stays current with evolving cultural research and AI capabilities.

In conclusion, Empa demonstrates that AI‑mediated intercultural training can be seamlessly integrated into technical computing education, producing measurable gains in cultural intelligence and collaborative effectiveness. The study provides empirical evidence that such interventions are feasible, scalable, and aligned with the workforce demands of increasingly globalized HPC environments. By offering a concrete, reproducible framework, the paper paves the way for broader adoption of AI‑enhanced soft‑skill development in STEM curricula.