Computer Science Department , PhD Defense Yang Wu " Advancing the Proactive Capacity of LLMs in Cost-Sensitive Domains"
10:00 a.m. to 11:00 a.m.
Yang Wu
PhD Candidate
WPI – Computer Science Department
Thursday, April 9th, 2026
Time: 10:00 AM-11:30 AM
Location: Innovation Studio 105
Zoom Link: https://wpi.zoom.us/j/6065536440
Committee members :
Prof. Xiaozhong Liu (Advisor), WPI-Computer Science
Prof. Kyumin Lee, WPI-Computer Science
Prof. Frank Zou, WPI- Mathematical Science
Prof. Haixu Tang, Indiana University Bloomington-Data Science
Abstract:
Recent advances in large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks; however, their effectiveness remains limited in cost-sensitive domains such as biomedical research, legal reasoning, and personalized education, where high-quality data is scarce, expert knowledge is expensive, and real-time interaction constraints are critical.
This dissertation addresses how to advance the proactive capacity of LLMs under such constraints, enabling them not only to respond to user queries but also to actively acquire knowledge, adapt to domain-specific needs, and collaborate effectively with humans. To this end, it presents a unified framework spanning three dimensions: cost-efficient data acquisition, proactive domain assistance, and proactive team-AI collaboration. For data efficiency, we develop PUtree, PU-ADKA, and ROSE to improve model performance under limited supervision and constrained expert budgets.
For domain assistance, we design systems such as D3LM, TASA, and CamPilot that enable LLMs to actively elicit missing information, personalize interactions, and generate structured outputs across legal, educational, and multimodal settings. For collaborative intelligence, we introduce CoLabScience and DTR, which enable LLMs to identify when and how to intervene in ongoing scientific discussions and achieve low-latency, high-quality interaction through offline rehearsal. Together, these contributions transform LLMs from passive responders into proactive collaborators and provide a principled approach for deploying LLM systems in real-world, high-stakes, and cost-sensitive environments.