1. Foundational Breakthroughs in LLMs. We focus on expanding the frontiers of intelligence in LLMs. We are pretraining LLMs based on novel architectures, with the goal of developing LLMs that support unbounded context windows, continuously improve themselves, and dynamically adapt their architectures after deployment.

2. Agents and Agent Economies. We study how intelligence emerges and evolves in individual agents and multi-agent systems through learning and interaction. We also explore and shape societal structures in which humans and intelligent agents coexist, together with the economic foundations that support this coexistence.

1. Tonghan Wang’s doctoral dissertation received the Victor Lesser Distinguished Dissertation Award, which is presented annually to just one PhD graduate worldwide in recognition of outstanding doctoral research in agents and multi-agent systems. (Read the award announcement.)

2. He received the ACM EC 2024 Best Paper Award (AI Track) and has earned five Spotlight distinctions at NeurIPS and ICLR, as well as three Oral distinctions at AAAI, AAMAS, and EC. To date, he has published more than 30 papers at leading conferences, including NeurIPS, ICLR, ICML, EC, AAAI, and AAMAS.

1. Foundational Theory. Tonghan Wang has applied deep learning to tackle a major open problem in game theory and computational economics: optimal auction mechanism design. His work provides an empirical counterpart to Professor Andrew Yao’s 2017 theoretical research on incentive-alignment guarantees regarding dominant-strategy incentive compatibility (DSIC). DSIC is a cornerstone of microeconomics and is widely regarded as one of the “holy grails” of computational economics and game theory. In 1981, Roger Myerson derived the optimal mechanism for single-item auctions—a foundational contribution that led to his receipt of the 2007 Nobel Memorial Prize in Economic Sciences. Wang’s work introduces the world’s first deep-learning-based mechanism design framework for multi-item, multi-bidder settings that provides rigorous DSIC guarantees.

2. LLMs and Agentic AI. His research examines novel LLM architectures and their training methods, as well as training, inference, memory, collaboration, and incentive mechanisms in agentic AI systems. In his work on LLMs, he has identified the functional roles of individual neurons and leveraged these insights to develop efficient new architectures and training methods. In his work on agents, he investigates the processes through which multiple autonomous agents explore, divide labor, communicate, organize, and evolve intelligence within shared environments. When agents’ interests are not fully aligned, he employs mechanism design and contracts to achieve socially optimal outcomes.