Xubin Wang's Site

These two surveys establish complementary foundations for data-centric, resource-constrained intelligent systems. Cognitive Edge Computing systematizes a closed-loop deployment paradigm for reasoning-capable large language models (LLMs) and autonomous agents across cloud–edge–device hierarchies. It organizes optimization layers (quantization, sparsity, low-rank adaptation, distillation), elastic collaboration (selective offloading, hierarchical routing, privacy-preserving personalization), and adaptive intelligence (context compression, dynamic agent/tool orchestration) into a cognition-preserving framework with standardized multi-axis evaluation (latency, energy/token, robustness, privacy, sustainability). In parallel, the On-Device AI, CSUR'25 survey characterizes design and evolution of lightweight yet capable models under stringent memory, power, and real-time constraints—covering data preprocessing pipelines, feature selection and curation, compression (quantization/pruning/distillation), hardware-aware co-design, and emerging foundation model influences.

We tackle the challenges associated with text classification tasks, particularly in few-shot prompting scenarios, by introducing the Reinforced Diverse Example Selector [RDES, ICML' 25]. RDES employs a reinforcement learning framework, specifically Q-learning and PPO, to optimize the selection of diverse reference examples, ensuring a balanced representation of data that enhances classification accuracy. Additionally, we explore the integration of Chain-of-Thought reasoning into the selection process, which further boosts the model's predictive performance. In parallel, we present the RAG-QA-Generator, an automated tool designed for the construction and management of knowledge bases within Retrieval-Augmented Generation (RAG) systems. This tool processes document data and utilizes large language models to generate high-quality question-answer pairs, facilitating the automated development of RAG system knowledge bases. Together, these contributions highlight the potential of advanced methodologies in addressing the complexities of text classification and knowledge management.

Our research at the intersection of Evolutionary Machine Learning and its applications has focused on addressing prominent challenges in complex domains like feature selection, biomarker identification and cancer classification. We have proposed novel algorithms and frameworks such as [SaWDE, KBS'22] for large-scale feature selection using a self-adaptive differential evolution approach, [MEL, TKDE'24] for multi-task evolutionary learning through information sharing, [FWPSO, BIBM'22] for efficient biomarker gene identification from microarray data via feature weighting particle swarm optimization, [EODE, TCBB'24] for ensemble-based improved cancer screening through optimized feature selection, modeling, and classification, and [HSNOE, ESWA'24] which leverages a hybrid sampling technique and ant colony-based feature selection within an ensemble to enhance identification of hidden responders in imbalanced biological data. Extensive experimentation demonstrates the superior and robust performance of our proposed approaches, validating their ability to provide effective solutions for challenging machine learning problems across domains while overcoming issues like local optima, dimensionality, and generalization across datasets.