Friday, November 13, 2026
2:00 PM – 3:00 PM CST
InfraAI Lecture Series: Vishal Misra on Bayesian Geometry and Transformers
About this event
A UT Austin InfraAI Center lecture by Columbia's Vishal Misra, From Prediction to Computation: Bayesian Geometry, Transformers, and Mathematical Discovery, on what computations transformers learn and how training shapes attention.
From next-token prediction to computation
The InfraAI Center at UT Austin hosts Vishal Misra, a Columbia Engineering professor of computer science and electrical engineering and vice dean for computing and AI, for a special lecture titled "From Prediction to Computation: Bayesian Geometry, Transformers, and Mathematical Discovery."
Language models are trained only to guess the next token, yet they show reasoning, in-context learning and even mathematical discovery. Misra asks what computations transformers actually learn. He presents controlled experiments where the ideal Bayesian predictor is known exactly, showing how transformers learn inference and how gradients shape attention geometry, then checks whether the same signatures appear in large models. He also covers limits, including a case where a transformer can recognize an underlying program but cannot reliably reuse it, and connects the work to AI-assisted mathematics.
Practical details
- When: Friday, November 13, 2026, 2:00 pm to 3:00 pm.
- Where: GDC 2.216, Gates Dell Complex, UT Austin campus.
- Access: The campus calendar page does not say whether the lecture is open to the public and lists no registration. Confirm with the InfraAI Center contact on the page before attending.
Organized by
The long-running AI colloquium of the Department of Computer Science at The University of Texas at Austin. FAI brings visiting and local researchers to campus through the academic year for hour-long technical talks spanning machine learning, robotics, language, and human-AI interaction. Talks are free, open to the public, held in the Gates Computer Science Complex, and simulcast on Zoom.