In this talk, I will present a summary of my research over the past five years on developing AI foundation models across language, code, and multimodal domains. I will begin by introducing our early work on multilingual pre-trained models, including the Unicoder series for natural language and the CodeBERT/CodeGPT series for programming languages, along with their applications in AI agent systems. Next, I will highlight our recent work on improving the training efficiency of large language models (LLMs), aimed at reducing reliance on expensive GPU resources—an effort recognized as the runner-up for Best Paper at NeurIPS 2024. I will then discuss our broader efforts to enhance the reasoning abilities of LLMs through techniques such as self-criticism, data synthesis, tool-augmented verification, and multimodal chain-of-thought reasoning. Following that, I will introduce our progress in building foundation models for multimodal understanding and generation—particularly in text-to-video and image-to-video generation—through the Unicoder-VL, NUWA, and Step-Video series. I will conclude by sharing my perspectives on promising directions toward artificial general intelligence (AGI) and look forward to an engaging discussion.

Speaker: Dr. Nan DUAN
Date: 16 May 2025 (Friday)
Time: 10:30am – 11:30am
Venue: LAU 6-209
Poster: Click here