Abstract
[Google Meet]
連結點我
Multimodal LLMs have achieved remarkable performance across many benchmarks, but its reliability and novel problem solving---capabilities that are central to real-world applications---remains unclear. In this talk, I will present our recent work on probing this question, with a focus on error correction and consistency of MLLMs, and coverage shrinkage in RLVR, a widely used and efficient algorithm for training reasoning LLMs. I will also highlight key challenges in deploying LLMs—and mainstream Machine Learning more broadly—in low-resource settings.
連結點我
Multimodal LLMs have achieved remarkable performance across many benchmarks, but its reliability and novel problem solving---capabilities that are central to real-world applications---remains unclear. In this talk, I will present our recent work on probing this question, with a focus on error correction and consistency of MLLMs, and coverage shrinkage in RLVR, a widely used and efficient algorithm for training reasoning LLMs. I will also highlight key challenges in deploying LLMs—and mainstream Machine Learning more broadly—in low-resource settings.
Bio
Khoa D Doan is currently an Assistant Professor in the College of Engineering and Computer Science (CECS) at VinUniversity, Vietnam and also the Associated Director of VinUni-Illinois Smart Health Center (VISHC), a joint initiative between VinUniversity and the University of Illinois Urbana-Champaign (UIUC). His research focuses on developing computational frameworks that enable the safe/secure and practical deployment of ML models in constrained and especially low-resource applications. Prior to his academic path, he held multiple industry positions, spanning from software developer to research scientist. He received his Ph.D. in Computer Science (machine learning) from Virginia Tech and his M.S. in Computer Science (high-performance distributed computing) from the University of Maryland, College Park.