I am intrigued by the underlying "physics" of Large Language Models (LLMs), focusing on how they represent knowledge, absorb knowledge, process information, and make predictions. A major part of my research is improving the efficiency and accuracy of updating pretrained LLMs, ensuring that their knowledge remains robust, consistent, and up-to-date while minimizing costs and time.

My work in this area centers on designing new paradigms for multi-modal interactions and deriving empirical scaling laws for multi-modal foundational models. I focus particularly on video understanding and generation, aiming to seamlessly integrate language, visual, and temporal modalities.

A key aspect of my research is exploring how multimodal interactions are learned, including the mechanisms by which information flows and aligns across modalities to create cohesive representations. I also investigate protocols for efficient and complete multimodal interactions, ensuring that each modality contributes optimally to the task at hand while minimizing redundancy and maximizing interpretability.

My goal is to develop systems capable of effectively analyzing and generating content across diverse modalities, with applications in tasks such as video captioning, video-based reasoning, and video synthesis, thereby advancing the capabilities of multi-modal AI.

I use NLP to tackle challenges in social and scientific domains. For example, I analyze social media data to study the spread of misinformation and its societal impacts, helping to develop tools that counteract disinformation.

Additionally, I explore how NLP can contribute to scientific advancements, such as facilitating drug design and interpreting neural signals. These interdisciplinary applications demonstrate the transformative potential of NLP in both improving lives and advancing science.