Yucen Wang

I am Yucen Wang (汪钰岑), a Ph.D. student in Computer Science and Technology at Nanjing University, affiliated with the LAMDA Group in the School of Artificial Intelligence, advised by Prof. De-Chuan Zhan. I received my M.Sc. from Nanjing University in 2023, also with LAMDA. I was a research intern at Microsoft Research Asia, working with Kaixin Wang and Li Zhao on world models, reinforcement learning and embodied AI.

njuwyc@gmail.com  /  wangyc@lamda.nju.edu.cn  /  WeChat ID: Plan2Explore

Google Scholar  /  GitHub  /  LinkedIn

Research Interests

My long-term research direction focuses on building general-purpose embodied agents through foundation models, world models, and reinforcement learning.

• Learning interactive world models from visual and robotic data to support prediction, imagination, planning, policy evaluation and improvement.
• Building Vision-Language-Action models and World Action Models for scalable robot learning, RL post-training, and real-world deployment.

My broader interests include model-based RL, representation learning in visual RL, latent action modeling and video generation models.

Seeking Opportunities

I expect to graduate in 2027. I am open to research collaborations, visiting or research internships, postdoctoral positions, and future full-time opportunities in academia or industry, around world models, reinforcement learning, and robot foundation models. Happy to connect and discuss potential opportunities.

Publications

Introduces Raw-Dream, a world-model-based VLA RL post-training paradigm that pairs a task-agnostic video world model trained on diverse untargeted behavior data with zero-shot VLM rewards, enabling unseen-task adaptation while mitigating world model hallucinations through initial-noise-based uncertainty filtering.

Introduces CoLA-World, which pioneers the joint training of latent action models and world models by directly repurposing a pretrained video generation model as the forward dynamics model, resolving representational collapse and enabling a synergistic co-evolution framework.

Grounds VLM semantic representations into a world-model latent space, using predicted temporal distances as intrinsic rewards for reward-free policy learning and multimodal goal conditioning through world model imagination.

Constructs an Implicit Action Generator to infer latent behaviors of visual distractors and train separated world models, allowing policy optimization in a clean latent space for visual control with complex task-irrelevant dynamics.

Develops a Vision-Language-Latent-Action framework that injects physically grounded latent actions into VLA training.

Systematically reviews reward modeling techniques in deep reinforcement learning, organizing methods by reward source, modeling mechanism, and learning paradigm.

Learns separated environment dynamics by action-determined task relevance, reducing the effect of visual distractors on complex visual imitation tasks.

Experience

Microsoft Research Asia, Research Intern, Machine Learning Group, Apr. 2025 - May 2026. Worked on building action-conditioned world models from pretrained video foundation models, learning latent action models and latent-action-based world models, and RL post-training of VLA models in video world models.

Nanjing University, M.Sc. and Ph.D. student at LAMDA Group, 2020 - present. Advised by Prof. De-Chuan Zhan, focusing on reinforcement learning and world models.

Website template adapted from Jon Barron.