----------------------------------------------------------------------------------------------------







       Department of Systems Engineering and Engineering Management



                          The Chinese University of Hong Kong



----------------------------------------------------------------------------------------------------





Date: Friday, June 27, 16:00 pm – 17:00 pm



Venue: ERB 513, The Chinese University of Hong Kong



Speaker: Professor Farzan Farnia, The Chinese University of Hong Kong



Title: Online Learning Strategies for Model Selection in Generative AI



Abstract: Generative AI services are now widely accessible, offering 

users a growing number of models and APIs to choose from. This 

abundance raises a fundamental challenge: how can we systematically 

and efficiently select the most suitable model for a given task, 

especially when model costs vary and user needs are dynamic? In this 

talk, I present the application of online learning and bandit 

algorithms to address this challenge. Specifically, we show how 

multi-armed bandits can be used to evaluate and select among 

unconditional generative models in an online fashion, balancing 

quality and diversity without relying on ground-truth data. We then 

extend this framework to the prompt-aware setting using contextual 

bandits, enabling adaptive model selection based on input prompts 

while accounting for trade-offs such as cost and performance. These 

algorithms provide an efficient foundation for task assignments to 

generative AI services.



Bio: Farzan Farnia is an Assistant Professor of Computer Science and 

Engineering at The Chinese University of Hong Kong. Prior to joining 

CUHK, he was a postdoctoral research associate at the Laboratory for 

Information and Decision Systems, Massachusetts Institute of 

Technology, from 2019 to 2021. He received his master’s and PhD 

degrees in electrical engineering from Stanford University and his 

bachelor’s degrees in electrical engineering and mathematics from 

Sharif University of Technology. His research interests span deep 

generative models, information systems, and convex optimization.