Jialuo Li

I am a first-year graduate student in Computer Science at the Georgia Institute of Technology, where I am advised by Prof. Humphrey Shi. I received my undergraduate degree from the Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University (also known as the Yao Class, directed by Turing Award laureate Andrew Chi-Chih Yao). Previously, I interned at New York University, where I was advised by Prof. Saining Xie.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

Humans exhibit an extraordinary capacity to integrate information from multiple sensory modalities, such as vision, auditory, and tactile inputs, to navigate and interpret their environments with remarkable efficiency. This multimodal integration leverages the complementary strengths of each sensory channel, facilitating a coherent and comprehensive understanding of complex surroundings. Inspired by this cognitive prowess, my long-term objective is to develop AI systems that emulate human-like multimodal synthesis, thereby enhancing robustness and adaptability in both generative and understanding tasks.

Specifically, I focus on multimodal generative models, ranging from text-to-image synthesis to Multimodal Large Language Models (MLLMs) and unified models.

DIG optimizes long-form video understanding by distinguishing between "global" and "localized" queries, applying different strategies for each with a training-free approach.

SAGE introduces a human-inspired agentic framework that replaces resource-heavy frame processing with iterative reasoning and a diverse toolkit for efficient long video understanding by leveraging a novel synthetic data pipeline and a specialized RL strategy.

PAI-Bench is a new benchmark designed to evaluate Physical AI capabilities across video generation and understanding. The study finds that while current models produce high-quality visuals, they lack the physical common sense and reasoning required to truly understand real-world dynamics.

We introduce SciScore, a reward model that improves generative models' scientific accuracy. Trained in two stages, it achieves human-level performance in evaluating scientific image realism on Science-T2I dataset.

CHOPS uses small and large LLMs to create an efficient, safe LLM agent that accesses user data, interacts with systems, and delivers accurate responses.

This homepage is designed based on Jon Barron's homepage and deployed on GitHub Pages. Last updated: Jan 21, 2026.
© 2026 Jialuo Li