Xiaoyu Liu

Xiaoyu Liu

Research Engineer | Foundation Models, Multimodal Alignment, and Causal Reasoning

Hi, this is Xiaoyu.

I am a research engineer at Google, where I have worked on post-training and alignment for large multimodal models since 2025. Prior to joining Google, I received my Ph.D. at the University of Maryland, College Park, advised by Professor Furong Huang on causal inference, trustworthy AI and their applications in LLM. I received my bachelor's degree in mathematics and computer science from Colorado State University, where I worked on optimization and computational methods with Professor Michael Kirby.

Email CV Google Scholar LinkedIn GitHub

News

  • Released PersonaLedger, a privacy-preserving synthetic transaction dataset and benchmark for financial AI, generated through persona-conditioned LLMs with rule-grounded feedback.
  • One paper was accepted to NeurIPS: ViCrit, a verifiable RL proxy task that trains VLMs to localize subtle visual caption hallucinations.
  • Two papers on causal debiasing and model understanding were accepted: Prompting Fairness for LLMs at ICLR and CSRec for large recommendation models at SIGIR.
  • One paper was accepted to EMNLP: DISCO, a domain- and difficulty-aware RL method for imbalanced alignment data.
  • Three works on VLM hallucination benchmarking were released: HallusionBench at CVPR, Mementos at ACL, and AutoHallusion at EMNLP.
  • One paper on causal debiasing for LLM-generated medical reports was accepted to Communications Medicine, and one paper on eliminating spurious correlations in LLMs was accepted to ACL.

Selected Publications

Full list