Events

CILVR Seminar: Enhancing Multilingual LLMs with Code-Switching

Speaker: Haneul Yoo (NYU)

Location: 60 Fifth Avenue, Room 7th Floor Open Space
Videoconference link: https://nyu.zoom.us/j/98954373715

Date: Wednesday, September 30, 2026

Abstract: Despite recent advances, large language models (LLMs) exhibit uneven performance across languages and often struggle with code-switching—the natural alternation between multiple languages. While code-switching is a common and meaningful mode of human communication, LLMs remain brittle to code-switching inputs, treating them as noise. This talk reframes code-switching as a valuable linguistic signal, exploring how it can be leveraged across the LLM lifecycle, from evaluation, to training, to inference. I will begin with output language alignment (OLA), highlighting systematic failures to respond in the expected language in code-switching contexts. I will also introduce code-switching red-teaming (CSRT), which reveals vulnerabilities in safety and multilingual understanding. I will then present code-switching curriculum learning (CSCL), a data-efficient approach to cross-lingual transfer inspired by human second-language learning. Finally, I will introduce code-switching in-context learning (CSICL), which uses gradual transitions from a target language to English to improve multilingual generalization without additional training.

Bio: Haneul Yoo is an Assistant Professor/Faculty Fellow in the Department of Computer Science, Courant Institute School of Mathematics, Computing, and Data Science at New York University. She is also a Research Fellow with the Social Innovation and Special Projects Team in the Division of Policy, Evaluation and Training, Department of Peace Operations at the United Nations. She received her PhD from the School of Computing at KAIST. With a background in machine learning (ML) and natural language processing (NLP), she has worked on (1) developing language models and methods to improve generalization across languages, (2) designing resources and evaluation frameworks for underrepresented languages and cultures, and (3) applying AI in high-stakes, real-world domains for social good. She has published papers in major NLP conferences, including ACL, EMNLP, and NAACL, and has been named a Rising Star at KAIST exploreCSR supported by Google. She has served as a Social Chair at NeurIPS 2026 and co-organized the Multilingual and Equitable Language Technologies (MELT) workshop at COLM 2025, the Women in Machine Learning (WiML) symposium at ICML 2026, and the Generative AI for the World (GenAI4World) workshop at COLM 2026.