Guowei Xu

I'm a senior undergraduate student (from 2023 Fall) in Institute for Interdisciplinary Information Sciences (IIIS, Yao Class) at Tsinghua University, majoring in Artificial Intelligence. Prior to this, I was a member of the Chinese national team for the International Physics Olympiad (IPhO).

I have been incredibly fortunate to be advised by Prof. Huazhe Xu (Tsinghua Embodied AI Lab), Prof. Li Yuan (PKU-YuanGroup), Prof. James Zou (Stanford AI for Science Lab), and Prof. Yilun Du (Harvard Embodied Minds Lab).

I am actively seeking PhD opportunities for Fall 2027. I'm always eager to connect with fellow researchers and explore potential collaborations or exchange ideas.

News

• [Oct. 2026] 📝 New preprint MPT is out, a framework for jointly learning action policies and self-orchestration in decentralized agent swarms.
• [Sept. 2026] 🎉 Our papers Reward Subsystem and EoM are accepted to NeurIPS 2026!
• [Jun. 2026] 🎤 Honored to serve as a Session Chair at ACL 2026!
• [Jun. 2026] 🎉 Our paper BES is accepted to ICML 2026 Workshop AI4Math!
• [Jun. 2026] 📝 New preprint Economy of Minds is out, exploring emerging multi-agent intelligence with economic interactions.
• [May 2026] 📝 New preprint BES is out, on self-improving language models with bidirectional evolutionary search.

Research Interests

My long-term goal is to understand and build intelligence that can solve real-world tasks.

Intelligence need to be understood. I study the signals inside models and use them to guide learning, from dormant neurons that reveal when RL agents stop exploring [DrM] to a sparse subsystem of value and dopamine neurons that encodes reward inside LLMs [Reward Subsystem].

Intelligence can learn from human experience. People reason in stages, refine how they learn, and break hard problems apart before recombining the pieces. I bring these strategies to models through structured visual reasoning [LLaVA-CoT], optimizers that are themselves optimized [metaTextGrad], and bidirectional evolutionary search for self-improvement [BES].

Intelligence emerges from interaction. Inspired by Minsky's Society of Mind, I study how collective capabilities arise among many simpler agents, from economic exchanges [EoM] to decentralized swarms that learn to organize themselves [MPT].

I also believe that intelligence should solve real-world tasks. Looking ahead, I want to keep reducing the human in the loop, toward self-evolving systems and agent societies that explore their environments and keep improving on their own.

Education

Publications & Preprints

Papers sorted by recency. Representative papers are highlighted.

Selected Open-source Projects

Selected Talk and Reports

Selected Awards

Teaching Assistant

Community Services