Fengxian Ji

About

Building agents that help advance scientific discovery.

Hello, I am Fengxian Ji. My research focuses primarily on AI scientist agents, with an additional interest in applying artificial intelligence to finance.

I repeatedly exiled myself into nothingness and finally came to understand: human agency alone could not transcend all things. The vines of nihilism entwine us not because they are a void, but because they are an inescapable part of reality. May there be more tolerance among people, and more resilience within the self amid suffering. One must save oneself rather than wait to be saved, for we are all equally entangled in the vines.

News

Scroll for more

  • Style Wins, Substance Loses examines stylistic bias in LLM-based scientific idea evaluation through SciStyleBench.

  • ServImage connects image generation and editing evaluation with real-world commercial value.

  • LabGuard grounds natural-language laboratory rules into executable runtime guards for embodied agents.

  • The Library of AI Scientist provides a curated collection of papers and emerging research directions.

Research

Research Interests

01

AI Scientist Agents

Agentic systems that support and automate stages of the scientific research process.

02

AI for Finance

Applications of language models and intelligent agents in financial settings.

Research Output

Selected Publications

Overview of the SciStyleBench framework
SciStyleBench · Framework

Manuscript

Style Wins, Substance Loses: A Diagnosis of LLM-as-Judge in Idea Generation

Fengxian Ji*, Yuke Li*, Jingpu Yang*, Juanfan Wu, Fan Zhang, Zhexuan Cui, Yu Xie, Min Peng, Qianqian Xie, Xiuying Chen, and Zhuohan Xie

Introduces SciStyleBench to diagnose stylistic bias in LLM-based scientific idea evaluation across 600 ideas, 15 style variants, and three evaluation settings, together with metrics and a style-aware judging module.

AI ScientistLLM-as-JudgeEvaluation
Overview of the ServImage benchmark and evaluation framework
ServImage · Benchmark

Manuscript

ServImage: An Image Generation and Editing Benchmark from Real-world Commercial Imaging Services

Fengxian Ji*, Jingpu Yang*, Zirui Song*, Lang Gao, Junhong Liang, Zhenhao Chen, Jinghui Zhang, and Xiuying Chen

Connects image-generation evaluation with economic value using 1,070 paid design tasks, 2,050 designer deliverables worth over $295K, 33K candidate images, and human payment annotations.

Image GenerationImage EditingBenchmark
Code & Project
Overview of the LabGuard framework
LabGuard · Framework

Manuscript

LabGuard: Grounding Natural-Language Laboratory Rules into Runtime Guards for Embodied Laboratory Agents

Jingpu Yang*, Fengxian Ji*, Zhengzhao Lai*, Zhexuan Cui, Guangxian Ouyang, Qian Jiang, Fan Zhang, Min Peng, Qianqian Xie, Preslav Nakov, and Zhuohan Xie

Transforms natural-language laboratory rules into typed executable specifications and runtime monitors, with 812 annotations derived from 203 seed rules and evaluation in embodied laboratory tasks.

Embodied AgentsLaboratory SafetyRuntime Guards

* Equal contribution.

Open Source

Selected Project

Background

Experience

Remote Research Assistant

Wuhan University

Maintaining a long-term research collaboration with Dr. Zhuohan Xie.

Visiting Student

Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)

Advised by Prof. Xiuying Chen.

Recognition

Honors & Awards

  • Outstanding Undergraduate Graduation Thesis
  • National Scholarship for Undergraduates
  • School Second-Class Scholarship

Contact

Let’s connect.

I am open to conversations and collaborations related to AI scientist agents and AI for finance.

jifengxian1224@gmail.com