Speaker: Zora Wang Carnegie Mellon University
Time: 2024-12-18 22:00-2024-12-18 23:30
Venue: 线上 (https://us06web.zoom.us/j/85954598443)

Abstract:

Despite the potential of language model-based agents to solve real-world tasks such as web navigation, current methods still struggle with long-horizon tasks with complex action trajectories. In contrast, humans can flexibly solve complex tasks by learning reusable task workflows from past experiences and using them to guide future actions. I will first present Agent Workflow Memory (AWM), to enable agents to similarly induce and use commonly reused workflows in memory, under both supervised and unsupervised scenarios, achieving high success rate and versatile generalization across tasks and domains. To further boost agent generation efficiency, we expand workflows from memory to its action space in TroVE, by dynamically constructing reusable tools during the task-solving process. TroVE not only produces tailored tools for specific data distributions, but also accelerates and improves the accuracy of human verification compared to baseline approaches that generate primitive programs. I will conclude with a discussion of future directions, including advanced representations of workflows and the development of adaptive digital agents.

Short Bio:

Zora Wang (王芷若) is a PhD student at Carnegie Mellon University, Language Technologies Institute, advised by Professors Graham Neubig and Daniel Fried. Her research interest is using programmatic approaches to solve real-world digital problems, such as software engineering on code repositories, and building intelligent agents to navigate through personal devices. Her recent work focuses on building adaptive agents that learn increasingly complex task workflows in an online, unsupervised manner. Prior to CMU, she has worked on tabular intelligence at Microsoft Research (Asia) for two years. She has presented her work at and served as reviewers for top-tier NLP/ML conferences including NeurIPS, ICLR, ACL, EMNLP, and KDD. She has co-organized the first Agent Workshop at CMU, gave tutorials about "Language Models for Tabular Data" at SIGIR, and has been recognized with CMU Presidential Fellowship.