Events

TAUR Lab Seminar: Progressive Adaptation of AI Agents

Speaker: Zora Zhiruo Wang

Location: 60 Fifth Avenue, Room 204
Videoconference link: https://nyu.zoom.us/j/92408704679

Date: Wednesday, April 15, 2026

AI agents have shown remarkable promise, yet a critical ingredient remains overlooked: the ability to continuously adapt. This talk presents a principled framework for building agents that learn from experience, and rise to meet the complexity and diversity of real-world tasks.
In the first part, I introduce an online adaptation paradigm that moves agents from static executors to continual learners. Rather than treating each task in isolation, agents accumulate reusable workflow memory (AWM) and programmatic skills (ASI), yielding measurable gains in success rate, efficiency, and human verifiability.
In the second part, I take adaptation beyond simplified benchmarks and into the wild. By studying the distributional gap between how agents are developed and how humans actually work, and through a head-to-head comparison of AI and human workflows across occupational skills, a striking contrast emerges: agents tend to program their way through tasks, while humans prefer interfaces. I discuss how human-agent teaming can bridge this gap in the near term, before turning to a longer-horizon question: how do we build agents that reach expert-level competence? Our approach lies in distilling human expertise through structured agent-human interaction, and presents a systematic comparison of context- versus weight-based approaches to doing so.
Ultimately, the gap between agents and expert humans is not a fixed target but a moving frontier, and an approach built on continuous adaptation is the only one that can keep pace.

Speaker’s Bio: Zora Zhiruo Wang is a PhD student at Carnegie Mellon University, Language Technologies Institute, advised by Professors Graham Neubig and Daniel Fried. Her research focuses on using programmatic approaches to solve real-world problems, spanning adaptive agents and AI for human work. Zora has been recognized with Google PhD Fellowship and CMU Presidential Fellowship. She has organized a series workshops about Deep Learning for Code (at ICLR, NeurIPS, and ICML), and delivered tutorials on LLMs for tabular data at SIGIR 2024, and has presented her work at and served as reviewers for top-tier NLP/ML conferences.

In addition, Zora will be available for 1-1 meetings on April 15th. Feel free to book available slots to meet with her via the following spreadsheet: https://docs.google.com/spreadsheets/d/1AWGusKJ0GIL3ULRZb5-XgkQjok_DmTBHmzUSRZakdOQ/edit?usp=sharing