Peilin Cai’s Personal Website

More broadly, my research is driven by a simple goal: to enable models that both perceive and generate the real world in a scalable, physically grounded way. From structured 3D reconstruction during my undergraduate years at Wuhan University to my current work on large language models, world models, and embodied agents at USC, a recurring theme has been bridging raw sensory data with structured representations of scenes, actions, and goals.

I am especially interested in unified models of perception and generation for lifelong scene understanding and robust, context-aware agents.

I carried out two research projects of great personal significance at Prof. Yue Zhao's FORTIS Lab: SecDOOD (ICCV 2025 Poster) and PERSONABENCH (NeurIPS 2025 MTI-LLM Spotlight). The former proposed a secure on-device OOD detection framework that requires no gradient backpropagation, while the latter introduced the first benchmark for evaluating the personalization capabilities of LLMs in multi-turn conversational settings.

At GVL, I developed The Earth Simulator, a street-view world model that turns a handful of raw, pose-free images into long-horizon, camera-controllable exploration videos grounded in 3D geometry. By combining a persistent 3D Gaussian spatial memory with a generative video model, we aim to preserve real-world structure while achieving photorealistic, temporally stable rollouts from sparse, in-the-wild driving footage.

I have strong coding skills and a solid background in computer vision and natural language processing, along with extensive experience training, deploying, and running inference with LLMs and VLMs. I am also continuing to strengthen my research skills in robotics at the GVL Lab.

If you are interested in collaborating, please feel free to reach out. My preferred email is peilinca@usc.edu.