Tianxin Wei’s Dissertation Defense
PhD candidate Tianxin Wei will present his dissertation titled, “Principled Frameworks toward Autonomous Personalized Intelligence.” His final examination committee includes Professor Jingrui He (Chair and Director of Research), Information Sciences; Professor Jiaqi Ma, Information Sciences; Professor Jiawei Han, Computer Science; and Dr. Wang-Cheng Kang, Google DeepMind.
Abstract
Personalized intelligent systems face challenges in learning from sparse and biased feedback, scaling to foundation models, and adapting as users and environments change. This dissertation develops principled frameworks toward autonomous personalized intelligence through four stages: modeling foundations, foundation-model scaling, adaptive and agentic systems, and autonomous self-evolving agents. The first stage addresses reliable user modeling through meta-learning, counterfactual reasoning, fairness-aware recommendation, and structured representation learning. The second extends personalization to foundation models through multimodal generative interfaces, model fusion and compression, and provenance mechanisms for generated content. The third develops systems that adapt during deployment through test-time personalization, self-evolving memory, adaptive model ensembles, and evaluation of agentic reasoning. The fourth investigates mechanisms supporting self-evolving agents, including context-to-harness skill compilation, collaboration across heterogeneous language models, and structured representations for generative recommendation. Across these stages, theoretical analyses, empirical evaluations, and controlled ablations examine how representations, update rules, feedback sources, and inference policies shape system behavior. A common design principle is to specify what state can change, which evidence guides updates, how computation is allocated, and how improvements are evaluated. Together, these contributions connect reliable user modeling with scalable foundation models and controlled continual adaptation, providing technical foundations for personalized systems that can adapt and evolve reliably.
Questions? Contact Tianxin Wei.