Andrew Lan, UMass Amherst

Today (Thursday May 1st) 4-5 pm | 1400 18th Ave S A1013

Artificial intelligence (AI) and in particular large language model (LLM)-powered tools have demonstrated immense potential to improve teaching and learning in educational applications. For example, many digital learning platforms have adopted LLM-based chatbots for automated tutoring and feedback in conversational settings. 

A challenge remains in how to rapidly evaluate these tools during their development process, since rigorous randomized experiments (e.g., A/B tests) tend to have prolonged experimental cycles. Building an authentic, LLM-based simulated student agent can be a key enabler of rapid evaluation of learning tools: it can be used by teachers, content designers, and human tutors to practice their tutoring strategies and find effective feedback mechanisms, better understand progress and struggles of students, develop more targeted content, and adjust curricula. As a result, students may benefit from better tutoring strategies and learning content through a rapid evaluation cycle.

 

In this talk, Dr. Lan will detail two baby steps towards this goal, based on his work at the UMass machine learning for education (ML4Ed) group. First, towards instructing LLMs to make student-like errors, he will introduce DiVERT, a framework that learns an interpretable representation of errors, as text, behind distractors in math multiplechoice questions. Second, towards developing student models for open-ended learning activities, he will introduce dialogue knowledge tracing, a first attempt to track student knowledge evolution over the course of a 1-on1 dialogue between a tutor and a student. Dr. Lan will wrap up his talk by discussing numerous future research directions that need to be explored before an authentic, LLM-based simulated student agent can become a reality. 

Min Chi, North Carolina State University

Monday May 5th 4-5 pm | 1400 18th Ave S A1013

How do we design AI models that collaborate with humans to support, rather than replace, human decision-making? Models that are not only accurate, but also interpretable, adaptive, and aligned with human values? Human decisions rarely occur in isolation; they evolve, influenced by goals, feedback, and changing environments. In high-stakes domains such as education and healthcare, understanding and supporting these patterns is not just valuable—it is essential. 

In this talk, Dr. Chi will present a line of work that incorporates the fundamental characteristics of human-centered tasks into the design of deep learning and reinforcement learning frameworks. These models are built to reflect the sequential, goal-driven, and context-sensitive nature of human behavior. Drawing on tasks from adaptive learning platforms and early intervention in septic shock, she will demonstrate how integrating behavior modeling with decision-making frameworks leads to more effective prediction and intervention strategies. These methods could help bridge the gap between what human intend to do and the real-world outcomes of their actions.

You can find the latest information about the LIVE Learning Innovation Incubator and the LIVE Learning Innovation Series on our website and sign-up for our mailing list to receive periodic updates and invitations to our events.

LIVE: The Learning Innovation Incubator at Vanderbilt

Visit us online  |  Contact us

 

Manage your preferences | Opt Out using TrueRemove™
Got this as a forward? Sign up to receive our future emails.
View this email online.

2201 West End Ave None | Nashville, TN 37203 US

This email was sent to meagan.m.postema@vanderbilt.edu.
To continue receiving our emails, add us to your address book.