Ph.D. Research Proposal Exam: Yexiao He

Monday, November 30, 2026
9:00 a.m.
AVW 2460

ANNOUNCEMENT: Ph.D. Research Proposal Exam

Name: Yexiao He

Committee:

Professor Ang Li (Chair)

Professor Sennur Ulukus

Professor Heng Huang

Date/time: Monday, November 30 at 9:00 AM

Location: AVW 2460

Title: Experience-Driven Learning and Adaptation for AI Agents

Abstract: AI agents built on large language and vision-language models can reason, use tools, and complete complex tasks, yet successful execution does not by itself produce lasting improvement. Interaction histories contain useful procedures, factual evidence, and multimodal examples, but also errors and context-specific details. The central challenge is to determine what an agent should retain from experience, when that experience applies to a new task, and whether an update improves future performance without introducing harmful regressions.

This dissertation proposes methods for experience-driven learning and adaptation, with an emphasis on improving agents through external knowledge and skills while keeping the underlying model parameters fixed. Prior work on data selection, structured knowledge, symbolic reasoning, and tool orchestration provides a foundation. Building on preliminary work in Frozen Models, Evolving Expertise (FMEE), the proposed research will investigate three connected directions. First, it will study how to extract, organize, and validate reusable procedures, evidence-supported knowledge, and multimodal examples from task execution and feedback. Second, it will extend experience-driven adaptation to agents operating across heterogeneous tasks, investigating selective experience reuse, conflicting updates, and the balance between learning new capabilities and retaining existing ones. Third, it will investigate these methods in a natural-language interface for accessing and analyzing hospital data, including cardiac reports, with attention to verifiable tool outputs and the use of expert feedback when available.

Healthcare benchmarks and workflows developed with clinical collaborators will provide the primary evaluation settings. Experiments will separate experience collection, update selection, and final evaluation to measure generalization to unseen cases, transfer across tasks, retention of prior capabilities, and computational and feedback costs. Workflow evaluation will emphasize the correctness and traceability of data retrieval and analysis. The expected contribution is a set of methods and evaluation protocols for converting execution experience into reusable agent capabilities, clarifying when adaptation is beneficial and how its failures can be detected.

Audience: Faculty 

remind we with google calendar

 

November 2026

SU MO TU WE TH FR SA
1 2 3 4 5 6 7
8 9 10 11 12 13 14
15 16 17 18 19 20 21
22 23 24 25 26 27 28
29 30 1 2 3 4 5
Submit an Event