Xubin Wang

Selection-Centric AI for Reliable, Efficient, and Deployable Intelligence

My research asks one central question: what should an AI system select before it learns, reasons, and acts? I frame this through Selection-Centric AI, where reliability and efficiency are determined not only by model architecture, but by principled choices over data, context, tools, and execution policy under constraints of accuracy, latency, memory, privacy, energy, and cost. This perspective treats selection as a first-class scientific variable rather than a purely engineering heuristic.

Methodologically, I integrate reinforcement learning, evolutionary optimization, and robust representation learning across three layers: data-centric selection (feature/sample quality control for high-dimensional and imbalanced data), reasoning-context optimization (demonstration/context selection for LLMs), and system-level deployment (quality-latency-resource trade-offs for edge and on-device intelligence). This connects statistical learning, foundation-model reasoning, and deployable AI systems in a unified research program. Across these layers, my goal is to develop methods that are both theoretically grounded and practically reproducible.

Representative results include RL-based demonstration selection for in-context learning (ICML 2025), multi-task evolutionary feature selection (IEEE TKDE 2024), and a comprehensive survey of on-device intelligence (ACM Computing Surveys 2025, ESI Hot Paper). My recent and ongoing work appears in ICML, TKDE, TCBB, CSUR, KBS, and related venues, with applications in biomedical data analysis, RAG and knowledge-base construction, reasoning/agent workflows, and resource-constrained deployment. I am especially interested in research that bridges algorithmic advances with measurable impact in real-world AI systems.