Xiaoyang Cao 曹骁扬

Hi! I’m a master’s student at MIT. I am affiliated with the Institute for Data, Systems, and Society (IDSS), the Center for Computational Science and Engineering (CCSE), and the Laboratory for Information and Decision Systems (LIDS). I am advised by Michiel Bakker. Before MIT, I received my B.S. in Mathematics and Physics from Tsinghua University in 2025.

I study how intelligence arises and how to make AI safe. I’m currently exploring this through multi-agent systems and scalable oversight.

Feel free to reach out at xycao [at] mit [dot] edu.

Portrait of Xiaoyang Cao

Photo credits: Ziqi Wei

M.S. student at MIT IDSS, CCSE, and LIDS

Cambridge, MA

Publications

  1. Two robots in a corridor: under uniform responsibility shares both wait and are delayed, while under learned shares the empty robot yields so the loaded one proceeds

    Who Bears the Burden? Learning Responsibility for Shared Constraints in Multi-Agent Reinforcement Learning

    Xiaoyang Cao, Jingqi Li, Zhe Fu, Alexandre M. Bayen

    Preprint 2026 arXiv Website

  2. Two pipelines compared: under outcome-only RL the reader takes a role-violating shortcut and answers from memory, while Role Anchor preserves the intended flow through query and passages

    Do Modules Stay in Their Lane? Role Drift in Compound LLM Systems

    Xiaoyang Cao, Siddarth Srinivasan, Michiel A. Bakker

    Preprint 2026 arXiv

  3. Diagram of RE-PO's E-step and M-step for robust preference optimization

    RE-PO: Robust Enhanced Policy Optimization as a General Framework for LLM Alignment

    Xiaoyang Cao, Zelai Xu, Mo Guang, Kaiwen Long, Michiel Bakker, Yu Wang, Chao Yu

    ICLR 2026 arXiv Website

  4. PCBF objective-space decomposition around the Pareto front with upper and lower bounds

    Pareto Control Barrier Function for Inner Safe Set Maximization Under Input Constraints

    Xiaoyang Cao, Zhe Fu, Alexandre M. Bayen

    ACC 2025 arXiv Slides

  5. Virtual node accelerating message passing across a traffic graph

    Virtual Nodes Improve Long-term Traffic Prediction

    Xiaoyang Cao, Dingyi Zhuang, Shenhao Wang, Jinhua Zhao

    TRB Annual Meeting 2025 arXiv Slides

Research Interests

Multi-Agent Emergence.

I want to build methods that organize many agents so that strong abilities emerge from simple interaction rules.

My starting point is Anderson’s “more is different”: at scale, new phenomena emerge that you can’t predict from the parts. Wolfram’s A New Kind of Science makes the same case computationally. Simple rules, iterated, produce genuinely complex behavior. I want to study this in multi-agent AI systems, where a group of agents achieves what no single agent could. Right now I’m interested in agent swarms.

Scalable Oversight.

I want to build oversight methods that scale together with the model itself.

Goodhart’s law says that once a measure becomes a target, it stops being a good measure. Almost every way we train and audit AI runs into this (RLHF, LLM judges, debate, weak-to-strong supervision, etc.). This is a safety problem, because we lose the ability to tell when the system has gone wrong. It is also a problem for self-improvement, because a system can only improve as far as its supervision signal can see. I want to know where a supervision signal can come from when the system is stronger than its evaluator, and how to keep it trustworthy as both keep changing.

News

New preprint on learning responsibility for shared constraints in multi-agent RL is on arXiv.
Our paper on role drift in compound LLM systems was covered by VentureBeat.
New preprint on role drift in compound LLM systems is on arXiv.
RE-PO, our framework for robust LLM alignment, was accepted at ICLR 2026.
Started my M.S. at MIT.
Admitted to MIT for Fall 2025.
Our paper on Pareto Control Barrier Functions was accepted at ACC 2025.
Our paper on virtual nodes for traffic prediction was accepted at TRB 2025.

Education