Sheng Liu

Sheng Liu, PhD

shengl@stanford.edu
GitHub
Twitter
Google Scholar

I am a Postdoctoral Researcher at Stanford University, working with Prof. James Zou and Prof. Lei Xing. I earned my PhD in Data Science from New York University.

My research focuses on developing machine learning methods with theoretical foundations that enable AI to move from rigid tools to reliable, collaborative partners, with an emphasis on robustness to handle messy data, inference-time steering to incorporate feedback and evolving context, and agency to act as proactive assistants for complex tasks. I am particularly interested in their applications in disease and healthcare. Outside of academia, I love playing tennis and am also a certified scuba diver and surfer.

I am on the academic job market in the 2025 Fall cycle. Please feel free to reach out if you think I could be a good fit!

We are actively seeking students and collaborators. A background in machine learning, large foundation models, AI agents, and AI for medicine and healthcare is preferred. If you're interested in joining us, kindly send me an email.

Research Overview

Research Interests:

Machine learningAI for medicineReliable AIAI for science

Latest News

2025

  • 2025 10 Orgnizing a workshop on AI and healthcare at ICCV 2025 in Hawaii!
  • 2025 09 Co-authored work on trainable multi-agent system is now online! We received more than 900 stars on GitHub already!
  • 2025 04 Happy to serve as an Area Chair for ICLR 2026.
  • 2025 09 Co-authored work on revealing neurocognitive and behavioral patterns by unsupervised manifold learning from dynamic brain data is accepted at Nature Computational Science!
  • 2025 09 Our paper on investigating reasononing and hallucination of MLLMs is accepted at NeurIPS 2025.
  • 2025 09 Invited talks at Symposium on AI Medicine at Stanford University.
  • 2025 08 Invited talks at National University of Singapore, department of mathematics.
  • 2025 08 Invited talks at University of Macau.
  • 2025 08 Invited talks at Hong Kong University and Hong Kong University of Science and Technology.
  • 2025 06 Invited talks at Hippocratic AI
  • 2025 04 Invited talks at CILVR Lab and Center for Advanced Imaging at NYU as well as Columbia University Missing NYC so much!
  • 2025 04 Happy to serve as an Area Chair for NeurIPS 2025.
  • 2025 03 Serving as an Area Chair for ACML 2025.
  • 2025 03 Serving as the Local Chair for Conference on Parsimony and Learning (CPAL 2025). See you at Stanford!
  • 2025 03OctoTools is accepted by KnowledgeNLP 2025, Foundation Models in the Wild, Reasoning and Planning for LLMs workshops at ICLR 2025.
  • 2025 02Our agentic framework for tool usage OctoTools is online now. Demo and Github ⭐️
  • 2025 02TextGrad is now accepted by Nature .
  • 2025 02Our paper on Reducing hallucinations in VLM via latent space steering is accepted at ICLR 2025 as a Spotlight paper. (Code)
  • 2025 02Our paper on theoretically evaluating the data reconstruction problem is accepted at AISTATS 2025. We derived algorithmic upper bounds and the matching lower bounds on various defense methods.
  • 2025 02Large scale multimodal dataset for medicine MedTrinity-25M is accepted at ICLR 2025.(Code, Data)
  • 2025 01Guest lecture at UC Santa Cruz by Prof. Yuyin Zhou.

2024

Active Research Areas

1. Robust and Reliable AI

The real world is complex and noisy: data can be imperfect, labels may be inaccurate, and user prompts are often ambiguous. Directly deploying AI models in such environments can be dangerous. We develop methods with theoretical justification to enhance the robustness and reliability of AI models by robust pre-training and inference-time steering for real-world applications.

Related Publications:

Reducing hallucinations in vision-language models via latent space steering

Sheng Liu, Haotian Ye, Lei Xing, James Zou

ICLR (2025) Spotlight

In-context vector: making in-context learning more effective and controllable

Sheng Liu, Haotian Ye, Lei Xing, James Zou

ICML (2024)

Adaptive early-learning correction for segmentation from noisy annotations

Sheng Liu*, Kangning Liu*, Weicheng Zhu, Yiqiu Shen, Carlos Fernandez-Granda

CVPR (2022) Oral

Early-learning regularization prevents memorization of noisy labels

Sheng Liu, Jonathan Niles-Weed, Narges Razavian, Carlos Fernandez-Granda

NeurIPS (2020)

2. AI Systems and AI agents

We develop AI software platforms to assist human experts in clinical practice and promote human–AI collaboration. We also optimize AI systemts with large language models (LLMs).

Related Publication:

Optimizing generative AI by backpropagating language model feedback

Mert Yuksekgonul*, Federico Bianchi*, Joseph Boen*, Sheng Liu*, Pan Lu*, Zhi Huang*, Carlos Guestrin, James Zou (*equal contribution)

Nature (2025)

OctoTools: An Agentic Framework with Extensible Tools for Complex Reasoning

Pan Lu* , Bowen Chen*, Sheng Liu*, Rahul Thapa, Joseph Boen, James Zou (*equal contribution)

NAACL KnowledgeNLP (2025) Best paper award

3. AI for human disease and health

Medicine presents high-stakes, complex challenges where accuracy and reliability are paramount. We develop AI models to support clinical decision-making in areas such as radiation oncology and Alzheimer’s disease, aiming to improve treatment planning, diagnosis, and patient outcomes. Our work integrates domain knowledge with advanced machine learning techniques to ensure safe, effective, and trustworthy AI solutions for real-world medical applications.

Related Publication:

Automated radiotherapy treatment planning guided by GPT-4Vision

Sheng Liu*, Oscar Pastor-Serrano*, Yizheng Chen, Matthew Gopaulchan, Weixing Liang, Mark Buyyounouski, Erqi Pollom, Quynh-Thu Le, Michael Gensheimer, Peng Dong, Yong Yang, James Zou, Lei Xing (*equal contribution)

AAPM (2024) Best in Medical Physics

Generalizable deep learning model for early Alzheimer’s disease detection from structural MRIs

Sheng Liu, Arjun V Masurkar, Henry Rusinek, Jingyun Chen, Ben Zhang, Weicheng Zhu, Carlos Fernandez-Granda, Narges Razavian

Nature Scientific Reports (2023)