Date：2025/9/12 14:20-15:30
Location：R103, CSIE
Speakers：Prof. Kai-Wei Chang (UCLA)
Host：莊永裕教授
Abstract:
Recently developed vision-language systems have made impressive progress. However, recent work shows that their successes sometimes come from exploiting superficial cues rather than genuinely understanding images and context. In this talk, I will argue that multimodal intelligence requires attention to details—models must both perceive fine-grained visual information and integrate it with language understanding and reasoning. I will cover vision-language reasoning in diverse scenarios, including mathematical reasoning in visual contexts, social commonsense reasoning, physical commonsense reasoning, and reasoning over safety (customizing multimodal guardrails). I will further highlight our OpenVLThinker framework, one of the first open-source vision-language models to exhibit chain-of-thought reasoning through iterative self-improvement. I will conclude by discussing current limitations and future directions for building reliable multimodal models.
Biography:
Kai-Wei Chang is an Associate Professor in the Department of Computer Science at the University of California Los Angeles and an Amazon Scholar at Amazon AGI. His research interests include designing trustworthy natural language processing systems and developing multimodal models for vision-language applications. Kai-Wei has published broadly in NLP, AI, and ML. His awards include the Sloan Fellow (2021), AAAI Senior Member (2023), EMNLP Best Long Paper Award (2017), and KDD Best Paper Award (2010). He is elected as an officer of SIGDAT, the organizer running EMNLP. Kai-Wei obtained his Ph.D. from the University of Illinois at Urbana-Champaign in 2015 and was a post-doctoral researcher at Microsoft Research in 2016. Additional information is available at http://kwchang.net/.