Xinyuan Liu

I'm a second-year PhD student in Computer Science at Arizona State University, where I'm advised by Ransalu Senanayake in the LENS Lab. Previously, I was a research assistant at Vanderbilt University with Jie Ying Wu, working on multimodal models for surgical robotics. I received my M.S. in Computer Science from Rice University and my B.S. in Software Engineering from Huazhong University of Science and Technology.

Email  /  CV  /  Scholar  /  Github  /  LinkedIn

Research

I'm interested in data-efficient learning and reliable embodied AI: how intelligent systems learn useful behaviors and make better decisions from limited data and feedback. My work spans adaptive model selection, imitation and reinforcement learning, and multimodal evaluation of physical and cognitive capabilities. * indicates equal contribution.

copy@article{sadikoglu2026root,
  title={{ROOT}: Discovering Rewards for User-Specified Embodied Behaviors},
  author={Sadikoglu, Eren and Taparia, Aditya and Liu, Xinyuan and Senanayake, Ransalu},
  journal={arXiv preprint arXiv:2610.04250},
  year={2026}
}

Reward design cast as an observation-guided search over a tree of reward programs, trained policies, and rollouts, so learned behaviors match what the user asked for, such as natural gait and posture. Evaluated on four embodiments, including a real Unitree Go2.

copy@article{milkey2026physicslens,
  title={{PhysicsLENS}: Diagnosing Physical Property Blindness in Video Generation Models},
  author={Milkey, Isaiah and Sagar, Som and Taparia, Aditya and Liu, Xinyuan and Wen, Jiqing and Senanayake, Ransalu},
  journal={arXiv preprint arXiv:2610.01162},
  year={2026}
}

A robotics-grounded benchmark that asks whether video generation models follow a physical property stated in text but hidden from the frame, such as mass or viscosity. Plausible-looking videos often ignore it.

copy@article{liu2026physical,
  title={Physical Agentic {AI}: An Architecture for Orchestrating a Robot Crew with {LLMs}},
  author={Liu, Xinyuan and Sadikoglu, Eren and Chatterjee, Riana and Senanayake, Ransalu},
  journal={arXiv preprint arXiv:2608.22657},
  year={2026}
}

An LLM plans for a crew of robots but never actuates them; a procedural orchestrator checks every step against declarative workflow contracts before dispatch. Demonstrated on a Unitree G1 humanoid and Go2 quadruped.

copy@inproceedings{nguyen2026cupid,
  title={{CUPID} in the Model Zoo: Online Matchmaking for Selecting Your Dream {LLM}},
  author={Son Nguyen and Xinyuan Liu and Ransalu Senanayake},
  booktitle={Forty-third International Conference on Machine Learning},
  year={2026}
}

A dueling-bandit framework that learns a user's latent preferences from pairwise feedback and matches them to the right LLM under cost and time budgets.

copy@inproceedings{nguyen2026llm,
  title={{LLM} Routing as Reasoning: A {MaxSAT} View},
  author={Nguyen, Son and Liu, Xinyuan and Senanayake, Ransalu},
  booktitle={ICLR Workshop on Logical Reasoning of Large Language Models},
  year={2026}
}

Language-conditioned LLM routing cast as weighted MaxSAT, where natural-language feedback induces hard and soft constraints over model attributes.

copy@article{liu2026humanoid,
  title={Humanoid Factors: Design Principles for {AI} Humanoids in Human Worlds},
  author={Liu, Xinyuan and Sadikoglu, Eren and Senanayake, Ransalu and Huang, Lixiao},
  journal={arXiv preprint arXiv:2602.10069},
  year={2026}
}

A framework of physical, cognitive, social, and ethical humanoid factors for designing and evaluating humanoids that share human spaces.

copy@inproceedings{shahbazi2026neuralaugmented,
  title={Neural-Augmented {Kelvinlet} for Real-Time Soft Tissue Deformation Modeling},
  author={Shahbazi, Ashkan and Pereira, Kyvia and Heiselman, Jon S. and Akbari, Elaheh and Benson, Annie and Seifi, Sepehr and Liu, Xinyuan and Johnston, Garrison L. H. and Wu, Jie Ying and Simaan, Nabil and Miga, Michael and Kolouri, Soheil},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}

A physics-informed neural simulator that combines Kelvinlet analytical priors with FEM data for real-time soft-tissue deformation under surgical grasping.

Miscellanea

Website template from Jon Barron.