What is Sophia?

Sophia is a multimodal AI tutoring system that helps students identify and remediate conceptual misunderstandings in data structures and algorithms. Deployed in a junior-level data structures course at Virginia Tech, Sophia achieved a satisfaction rating of 3.95/5.0 across 70+ students while addressing fundamental challenges in computer science education.

Unlike traditional LLM-based tutoring systems that often provide direct answers, Sophia uses dynamic concept mapping to understand each student's knowledge state and guides them toward self-discovery of their misconceptions through targeted pedagogical interventions.

Dynamic Concept Mapping

At its foundation, Sophia generates and refines a dynamic concept map based on all available student context, including the student's current task, code submissions, terminal output, IDE interactions, and conversation history. This concept map represents each student's current knowledge state, tracking their understanding level, the system's confidence in its assessment, evidence-based reasoning, and timestamps.

The knowledge states that comprise the student's mental model are maintained and refined over time—the more the student uses the system, the more accurate and personalized the system's understanding of their mental model becomes. This continuous contextual awareness allows Sophia to precisely target areas of uncertainty through appropriate remediation tools.

Multi-Agent Architecture

Sophia uses a sophisticated multi-agent architecture where different hierarchical levels focus on key roles in identifying and remediating student misconceptions.

This separation of cognitive tasks into dedicated agents improves transparency throughout the diagnostic process and provides better reasoning and documentation along the workflow.

Multimodal Interaction Capabilities

Sophia supports multiple interaction modalities to accommodate diverse learning styles. Students can engage through voice-based natural language interaction, which frees cognitive resources and allows them to focus on supporting materials like visualizations. The platform also supports sketch-based input, enabling students to express their understanding through drawings and diagrams.

Research shows that expressing thoughts through speech has a much smaller gap than between thought and writing, potentially promoting better identification of student misconceptions through their spoken thought processes. The sketching capability allows students to externalize their mental models in ways that text alone cannot capture.

Pedagogical Approach

Sophia's system prompt emphasizes guiding users toward self-discovery while remaining empathetic and conversational. Rather than providing direct solutions, the system leverages its understanding of student knowledge states to provide precise, targeted remediation using visualizations and interactive tools.

This approach addresses a critical limitation of traditional LLM-based tools, which often provide excessive hints or reveal significant portions of the solution, diminishing learners' motivation to engage in independent problem-solving and reducing critical thinking capabilities.

Comprehensive Assessment Reports

At the end of each lesson, Sophia visualizes the newly refined concept map for students via a radar chart, allowing them to intuitively understand their current conceptual strengths and weaknesses. The report includes a written overview outlining what the student should focus on next, as well as a detailed table presenting the knowledge state for each concept.

"I can tailor my approach... it tells you what you're good at which is encouraging."

Study Participant

Students valued these non-intrusive assessment reports, which provide actionable feedback without feeling like the AI is intruding on their self-directed learning process.

Deployment & Results

Deployed in an authentic classroom environment, Sophia was used by 37 participants who completed 40 lessons across three core concepts: Binary Search Trees, sorting algorithms, and linked lists.

Follow-up interviews revealed that students valued maintaining control over when and how they received assistance, preferring manual control over automatic AI intervention. The research provides valuable insights into barriers to AI adoption in educational settings.