Autonomous Agents · Multi-Agent Systems · LLM Interpretability

Hi, I'm Dewen Liu.

I am an undergraduate researcher at Fudan University, majoring in Intelligent Science and Technology. My research focuses on autonomous agents, multi-agent systems, large language models, and AI interpretability.

Currently, I am working on the scientific evaluation of LLM-based multi-agent systems, agent workflow reconstruction, and mechanistic interpretability for foundation models. My long-term goal is to build AI systems that are not only capable, but also scientifically understandable, controllable, and trustworthy.

I am always happy to meet new friends and collaborators. Feel free to reach out if you are interested in agents, LLMs, multi-agent collaboration, or interpretability.

News

  • 2026 AgentXRay was accepted to ICML 2026.
  • 2026 Submitted a NeurIPS 2026 paper on scientific evaluation of LLM-based multi-agent systems.
  • 2025 Won MCM Outstanding Winner, SIAM Award, and COMAP Scholarship Award.
  • 2025 Received a Chinese invention patent as the first inventor: CN 120163631 B.
  • 2024–2025 Awarded the National Scholarship for two consecutive years.

Research Interests

Autonomous Agents

Agent design, tool use, self-improvement, evaluation, and deployment in complex reasoning environments.

Multi-Agent Systems

Scientific evaluation of collaboration, factor-controlled MAS analysis, communication, topology, and coordination.

LLM Interpretability

Feature explanations, causal effects, SAE-based analysis, and agent-assisted diagnosis of model behavior.

Research Experience

Towards a Science of Collective AI

Co-first author · Submitted to NeurIPS 2026 · arXiv:2602.05289

Developed a scientific evaluation framework for LLM-based multi-agent systems, including a collaboration-gain metric that separates genuine cooperation from resource scaling and a factor library for controlled MAS analysis.

Self-Evolving SAE Feature Explanation System

First author · Ongoing · Planned submission to AAAI 2027

Building a two-sided feature explanation framework that connects input-side SAE activations with output-side causal effects, assisted by self-evolving agents for diagnosis and adaptive optimization.

AgentXRay: White-Boxing Agentic Systems via Workflow Reconstruction

Contributing author · ICML 2026

Contributed to discussions and experiments for reconstructing black-box agent workflows into editable white-box systems using combinatorial search over agent roles, models, reasoning modes, and tools.

Selected Publications

Towards a Science of Collective AI: LLM-based Multi-Agent Systems Need a Transition from Blind Trial-and-Error to Rigorous Science

Jingru Fan*, Dewen Liu*, Yufan Dang, Huatao Li, Yuheng Wang, Wei Liu, Feiyu Duan, Xuanwen Ding, Shu Yao, Lin Wu, Ruijie Shi, Wai-Shing Leung, Yuan Cheng, Zhongyu Wei, Cheng Yang, Chen Qian, Zhiyuan Liu, Maosong Sun.

*Equal contribution.

arXiv:2602.05289 · Submitted to NeurIPS 2026

AgentXRay: White-Boxing Agentic Systems via Workflow Reconstruction

Ruijie Shi, Houbin Zhang, Yuecheng Han, Yuheng Wang, Jingru Fan, Runde Yang, Yufan Dang, Huatao Li, Dewen Liu, Yuan Cheng, Chen Qian.

ICML 2026

Honors & Awards

  • MCM Outstanding Winner, SIAM Award, and COMAP Scholarship Award, 2025.
  • National Scholarship, two consecutive years, 2024 and 2025.
  • CUMCM National Second Prize, 2025.
  • Chinese Invention Patent, first inventor, CN 120163631 B, 2025.
  • Siyuan Program, Fudan Excellence Student Training Program.
  • Outstanding Science Popularization Volunteer, Shanghai municipal level, 2024.

Education

Fudan University

B.Sc. in Intelligent Science and Technology · 2023–Present

GPA: 3.88/4.0; Rank: 1/27. Core coursework includes programming, data structures and algorithms, computer architecture, and statistical learning.

Beyond Research

Outside research, I enjoy badminton, detective fiction, board games, and meeting people from different backgrounds. I am always open to friendly conversations, research discussions, and new collaborations.