Academic
About Me
I’m interested in building decision-making capabilities in large models for complex tasks, with a focus on using reinforcement learning to achieve behavioral alignment — ensuring models don’t just produce correct answers, but do so safely, reliably, and in line with human values. My current research interests include:
- Design and training mechanisms for reward models;
- Planning, memory, and tool use in agent frameworks;
- Joint reasoning and policy generation under multimodal inputs.
Research Interests
Decision Models
Multimodal large models (e.g., VLMs) excel at tasks combining vision and language, making them a foundational component for complex decision systems. Meanwhile, emerging architectures like JEPA—which model in continuous vector spaces—may provide the underlying foundation for more general-purpose decision-making in the future.
Reward Models
A reward model acts as the 'judge' in reinforcement learning, determining which outputs deserve positive feedback. By carefully designing the reward function, we can guide models toward safer, more reliable, and human-aligned responses—making it a critical step in alignment.
Agent Frameworks
An agent framework defines how a model interacts with its environment: from receiving a task and planning, to calling tools, executing actions, and reflecting on outcomes. A well-designed framework can significantly enhance an agent’s autonomy, robustness, and problem-solving capability.
Selected Publications
Paper in Progress...
Grinding hard on my research!
A new paper is in the works—stay tuned!