Welcome to Yuheng Zha’s page!

I’m looking for internships and full-time positions for training vision-language models, world models, and agentic systems. Specifically, I expect to get enough computing resources for pre-training the WM-VLM, to enable the physical intelligence.

Bio

I’m a PhD candidate at UC San Diego, advised by Professor Zhiting Hu. I got my Bachelor’s degree (with Honors) from Zhejiang University.

Research Interests

My research interest lies in VLMs, World Models and Agentic systems. Currently, I’m actively doing research in

• Internal world models that enable VLM doing interleaved visual-textual reasoning
  • WM-VLM: Probing Internal World Models for Interleaved Visual-Textual Reasoning
• Reasoning with Vision-Language Models: Enhancing the reasoning capabilities of vision-language models through reinforcement learning with various reward signals.
  • Vision-G1: Towards General Reasoning Vision-Language Models via Reinforcement Learning
• Agentic Systems: Building autonomous agents that can perceive, reason, and act in complex environments (e.g., real websites) using multimodal inputs.
  • CocoaBench: Evaluating Unified Digital Agents in the Wild

My previous research also includes

• Evaluating the factual consistency of generative language models and building a unified model for many NLP tasks
  • AlignScore: Evaluating Factual Consistency with A Unified Alignment Function
  • Text Alignment Is An Efficient Unified Model for Massive NLP Tasks
• Training video world models with natural language actions and video states
  • Pandora: Towards General World Model with Natural Language Actions and Video States
• Visual Concept Learning
  • CReLeRI: Explainable, Concept-centric, Representation, Learning, Reasoning, and Interaction Video Analysis System