Department of Computational and Data Sciences
Department Seminar
Speaker : Dr. Sweta Agrawal, Senior Research Scientist at Google DeepMind
Title : Towards Reliable, Context-Aware, and Collaborative Multilingual Intelligence
Date & Time : July 17th, 2026 (Friday), 11:00 AM
Venue : # 102, CDS Seminar Hall
ABSTRACT:
Large Language Models have fundamentally transformed Natural Language Generation, yet they remain limited by two major bottlenecks: a static “train-once-and-deploy” learning paradigm and a generic “one-size-fits-all” approach to text generation. Effective communication is neither static nor generic. It must adapt to specific domains, evolving knowledge, and precise audience needs across languages. In this talk, I will present my research on bridging the gap between general-purpose AI capabilities and the nuanced user-centric needs of real-world applications.
A central theme of my work is enabling multilingual systems to move beyond generic generation by grounding outputs in context. I will highlight my core contributions in controlling audience-specific text complexity, leveraging context for machine translation in challenging domains, and embedding user preferences directly into the generation process, all under realistic resource constraints. Because standard evaluation metrics often fail to capture practical utility in these settings, I will also present robust, user-centric evaluation frameworks that measure whether a model’s output actually meets the desired constraints and satisfies user needs.
Looking toward the next frontier, I will outline my goal to transition LLMs from passive text processors into active, collaborative and intentional agents. This requires moving beyond handling explicitly provided contexts to navigating implicit, open-ended, and diverse multicultural contexts prevalent in real-world scenarios. I will discuss how we can build this next generation of reliable language technologies by equipping models with the capacity to deliberately integrate contextual information across multiple levels — from prompt to internal representation to model parameters — and the modularity to adapt to this evolving knowledge. Transitioning to this paradigm requires fundamentally rethinking how we measure success. I will detail how we can construct dynamic evaluation frameworks that assess long-horizon trajectories, knowledge retention, and real-world task success derived from implicit user feedback.
BIOGRAPHY:
Dr. Sweta Agrawal is a Senior Research Scientist at Google DeepMind, where she focuses on scaling and diversifying pretraining data to ensure Gemini delivers high-quality performance across global languages. Before that, she was a Postdoctoral Researcher at the Instituto de Telecomunicações. She earned her Ph.D. in Computer Science from the University of Maryland, College Park. Her works have received an EMNLP Outstanding Paper Award and top rankings at WMT evaluation campaigns. She has also held research roles at Meta, Unbabel, and Adobe, and actively serves as an Area Chair for leading AI/ML/NLP conferences.
Host Faculty: Dr. Danish Pruthi
ALL ARE WELCOME
