Linguistics/Statistics Seminar: Fei Wang

Title: From Risk to Resilience: Addressing Misalignment in (Multimodal) Large Language Models

 

Abstract: As (multimodal) large language models (LLMs) become central to intelligent systems, their use is expanding from everyday applications to specialized domains, such as assisting linguistic and statistical analysis. Alignment plays a crucial role in the successful development of LLMs, ensuring that model behavior matches our expectations and remains consistent with various objectives. However, misalignment persists as a significant challenge that undermines the trustworthiness and reliability of these models. This talk will explore methods to tackle the misalignment problem by addressing three key research questions: (1) How to enhance contextual faithfulness of misaligned LLMs through causal inference? (2) How to ensure a reliable alignment process by mitigating semantic shortcuts in the multimodal scenario? (3) How to integrate missing alignment objectives to achieve precise control over model behavior in diverse contexts, such as customizing computational linguistic tools? Particularly, this talk will systematically address these challenges with causal intervention, conditional alignment, and constraint integration. In addition, this talk will shed light on the responsible development of LLMs across linguistic, statistical, and interdisciplinary contexts.

 

Bio: Fei Wang is a Ph.D. candidate in computer science at the University of Southern California. His research focuses on identifying and mitigating fundamental risks in the development and deployment of (multimodal) large language models to enhance their robustness, controllability, and safety, ensuring trustworthy and responsible outcomes. His work has been recognized with an Amazon ML PhD Fellowship, an Annenberg PhD Fellowship, a USC Graduate School Research Award, and a USC CS Best Research Award. Additionally, he has instructed tutorials and served as area chairs at top tier NLP conferences, including EMNLP, NAACL, and ACL. More information can be found at https://feiwang96.github.io/