VirtualME personalizes code AI by profiling developer IDE behaviors

A research team led by Yuhong Liu has introduced VirtualME, an IDE-embedded infrastructure that models individual developers by continuously capturing their in-IDE behaviors. Unlike current one-size-fits-all code intelligence systems, VirtualME addresses the deeply personalized nature of programming—where developers vary in tool chains, domain expertise, and problem-solving strategies. The system comprises three components: log-level behavior extraction from the IDE, task-level behavior recognition via a multi-agent pipeline, and a rule-based engine that distills a four-dimensional developer persona (technology stack, ability, behavioral habits, and learning style). This persona is then fed into a personalized repository-level knowledge Q&A agent.

Evaluated on a multi-repository benchmark using real-world developer trajectories, VirtualME-enhanced answers outperformed generic baselines across five dimensions, achieving an average 33.80% improvement. The results demonstrate that rich, continuous behavior data can unlock adaptive, personalized code intelligence—moving beyond uniform AI assistance toward tools that truly understand and adapt to each developer's unique workflow. The paper has been accepted at FSE 2026.