Xinrun Wang

Hiring! I am looking for self-motivated PhD/Master students to join my group at SMU to work on i) fundamental decision making (i.e., single- and multi-agent reinforcement learning), ii) decision making for applications (e.g., RL for FinTech, MARL for urban security, AI for science), and iii) foundation agents (e.g., computer control agents and auto research agents). The requirement for SMU PhD admission can be found here. Visiting CSC students/researchers and remote interns are also welcome — please email me if you are interested.

About Me

I am an Assistant Professor and Lee Kong Chian Fellow at the School of Computing and Information Systems, Singapore Management University, where I lead the DIGA research group. My work builds highly-capable agents — empowered by (multi-agent) reinforcement learning and (multimodal) large language models — for complex decision-making tasks, and applies them to finance, urban security, and science.

Research Interests

(Multi-Agent) (Reinforcement) Learning, (Distributed) Foundation Models/Agents, Computer Science of Language Models, and AI for Science.

Pages we actively maintain:

News

AI for Materials Science

Speculations

  1. All decision-making scenarios — single-agent, cooperative multi-agent, competitive multi-agent (i.e., game theory), and mixed cooperative-competitive — will be mastered through one algorithm, one model, and even one set of parameters of the model.
  2. All the methods improving over a base foundation model — i.e., prompt engineering such as CoT, self-reflection, and alignment — should be addressed by a universal method, maybe reinforcement learning (RL).
  3. For any given problem A (maybe NP-hard), we can find another NP problem B such that if the foundation model can solve B, it can solve A. Therefore, AGI/ASI can be achieved by focusing on NP problems only.
  4. The foundation agent should emerge from a foundation model capable of solving difficult tasks. Structurally, it comprises three components: (i) a highly capable multimodal language model, (ii) minimal embodiment for tool creation and use, and (iii) optimization via reinforcement learning to drive the emergence of reasoning and tool-related capabilities. Mechanistically, it follows a trinity architecture: (i) a multimodal LLM, (ii) a memory module, and (iii) a world model, where memory and world model together form a universal harness that grounds and contextualizes the agent's behavior.
  5. Artificial General Intelligence (AGI) will ultimately be i) self-improving, i.e., searching with regular retraining, on decisions and alignment through interacting with (virtually) embodied environments such as video games, and ii) distributed intelligently across massive open-sourced domain-specific models.
  6. The current AI revolution represented by LLMs is nearing its end in the service industry, or the tertiary sector. Only by entering the secondary sector, represented by industry and manufacturing, can AGI ultimately be achieved. This will be piloted by AI for science, with AI for manufacturing as the main force.