Yu Gu (谷雨)

About

My research focuses on developing practical language agents to assist humans in real-world tasks and enhance productivity. Two key challenges are grounding and planning. Existing work has primarily focused on action grounding. However, situational grounding at a higher level, which aligns the agent’s internal state representation with the actual environment, remains underexplored. In planning, search is essential for optimizing performance but presents significant challenges in agent tasks due to the complexity of environmental interactions. Robust world modeling plays a crucial role in improving both situational grounding and search. My current efforts are dedicated to advancing methods in both situational grounding and planning.

As an AI researcher in general, I’m a fan of compression is intelligence, but NOT a fan of scaling is all you need. To achieve better compression, the following questions have to be answered.

• What’s the alternative to scaling law for achieving better compression? Should we combine weights-based compression with symbol-based compression (e.g., language)? And how?
• Closely related to the first question, what tasks are interpolative in vector space, and what are not? Our understanding of this question is constantly evolving (e.g., how many people considered learning natural language grammar an interpolative task a decade ago?). Fortunately, I believe most agent tasks today are still viewed as non-interpolative, which motivates us to explore solutions beyond scaling law using agent tasks as testbeds.

Selected Publications