Seminars

View all Seminars  |  Download ICal for this event

From Constrained Learning to Human-Like Large Language Models

Series: Department Seminar

Speaker: Dr. Jivnesh Sandhan, Kyoto University, Japan

Date/Time: May 20 10:00:00

Location: CSA Auditorium, (Room No. 104, Ground Floor)

Abstract:
In the first half of this talk, I will discuss how to apply contextual rules in sequence labeling problems. I will show how constraints can be activated at specific locations in a sequence and how they can be prioritized using a novel soft-masked attention module. I will illustrate how this constrained learning framework can be used to solve a very challenging tokenization problem in Sanskrit. I will also address an important question: do we still need classical engineered smaller models, or are large language models (LLMs) the answer to every problem? In the second half, I will discuss how human-like LLMs are. LLMs can generate impressive text, but we often get an intuitive sense of whether something like a social media post or an email is written by a human or an LLM. If LLMs are to be integrated into human society, their human-likeness becomes crucial. To understand this, I will invite you to wear a psychology hat and examine questions such as whether LLMs can act as human proxies in psychological experiments. I will extend this discussion to real-world applications, including education, mental health and customer service. Finally, we will see a new zero-day vulnerability in LLM safety that goes beyond standard jailbreaking and refusal bypass, demonstrating how latent persona drift can lead to unexpected behavioral shifts.

Speaker Bio:
Dr. Jivnesh Sandhan is an assistant professor at the department of Intelligence Science and Technology, Kyoto University. Prior to this, he served as a visiting assistant professor in the dept. of Computer Science at IIT Dharwad. He earned his Ph.D. at IIT Kanpur in 2023, following a dual degree (B.Tech-M.Tech) from the same institute in 2018. He has published 18 papers, including 10 publications at top-tier A*/A (7 first-authored) venues such as ACL, EMNLP, EACL and AACL. His current research primarily focuses on computational psychometrics, LLM jailbreaking and Sanskrit computational linguistics. He has contributed extensive academic service as an area chair for ARR at ACL 2026, EMNLP 2025, AACL 2025 and EACL 2026, and as a program committee member for BMCC 2026, ISCLS 2026, WSC 2025, and BHASHA-AACL 2025. His research contributions have been recognized with the Outstanding Area Chair Award and a Best Resource Paper Award finalist distinction at EMNLP 2025, the Best Poster Award at IndoML 2022.

Host Faculty: R Govindarajan