Title: On Modularity for Challenges posed by Constrained Generation from Language Models

Speaker: Prof. Kartik Goyal @ Georgia Institute of Technology

Time : 13:00 - 14:00, Nov 24th, 2025

Location: Online

https://hli.skku.edu/InvitedTalk251124 

Language: English speech & English slides

Abstract:

In this talk, I will focus on our work on constrained generation like machine translation from language models that involve estimation of conditional distribution under autoregressive models which is typically intractable. I will first characterize technical and algorithmic properties regarding conditional generation that suggest unique challenges for encoding (tokenization), decoding (generation and reasoning),  evaluation, and watermarking (deployment) for these tasks. Then, I will propose some solutions for these challenges followed by a discussion around the importance of modularity for these generation problems.

Bio:

Kartik Goyal is an assistant professor at the School of Interactive Computing in the College of Computing at Georgia Institute of Technology since Fall 2023. He is interested in Language Modeling, Natural Language Processing (NLP) and Machine Learning (ML) problems pertaining to modeling data with structured random variables for development of efficient, interpretable, and controllable probabilistic models. He is also excited about the intersection of these problems with various applications in Computational History and Humanities research. Prior to coming to Georgia Tech, he was a research assistant professor at Toyota Technological Institute at Chicago from Fall 2021-Fall 2023. He obtained is doctoral degree in Language Technologies in 2021 from the Language Technologies Institute in the School of Computer Science at Carnegie Mellon University.