I am Fan LIU, a graduate student at HKUST(GZ). My research focuses on LLM agents for data science, particularly code generation for data science. Broadly, I study how to build autonomous data science agents that can perceive and reason over data, make effective modeling decisions, and operate reliably in real-world environments. My work spans three main directions:

• Data-centric: methods that improve how agents perceive, interpret, reason over, and adapt to data, providing the perceptual foundation for autonomous data science agents. Representative examples include multimodal data perception for scientific discovery ([ICLR 2026] Towards Multimodal Data-Driven Scientific Discovery Powered by LLM Agents) and data governance mechanisms such as unlearning in graph-structured environments ([WWW 2025] Subgraph Federated Unlearning, OpenReview).

• Model-centric: methods that strengthen the modeling, reasoning, and decision-making capabilities of autonomous data science agents. Representative examples include agentic mathematical modeling ([NeurIPS 2025] MM-Agent: LLM as Agents for Real-world Mathematical Modeling Problem, arXiv, Code) and improving inference-time computation for LLM reasoning ([NeurIPS 2025] Bag of Tricks for Inference-time Computation of LLM Reasoning, arXiv, Code).

• Systems-centric: infrastructure that enables autonomous data science agents to operate reliably in real-world environments, including execution infrastructure for data and environment interaction, control infrastructure for long-horizon task execution, and platform infrastructure for scalable training, deployment, and governance. A representative system is DSLighting, an end-to-end data science agent platform (PyPI).

I have published over 10 papers at top-tier venues, including ICLR, NeurIPS, KDD, WWW, ECML PKDD, ICMLW, and TFS. For more details, please refer to [Google Scholar]. If you are interested in my research, feel free to reach out for discussions, collaborations, internship opportunities, or related inquiries.

I am on the job market for postdoctoral and industry positions.

Email: liufanuestc AT DOT com

Selected Works

NeurIPS 2025

MM-Agent: LLM as Agents for Real-world Mathematical Modeling Problem

MM-Agent is an LLM agent framework for real-world mathematical modeling. It decomposes open-ended modeling into problem analysis, model formulation, computational solving, and report generation, enabling end-to-end solutions for real-world mathematical modeling tasks.

Paper Code Demo 501 stars

Data Science Agent Harness

DSLighting

DSLighting is an LLM-driven autonomous data science execution engine that turns task descriptions and datasets into iterative code generation, execution, evaluation, and refinement workflows.

Code Docs 46 stars

Recent Works

Education and Experience

Awards, Acknowledgements, and Services