Xin Du · 杜鑫
Menu
Xin Du

Xin Du

Associate Professor · PhD Supervisor
Tongji University · Shanghai Research Institute for Intelligent Autonomous Systems

First author of more than ten papers at leading AI conferences, including ICML, NeurIPS, AAAI, and ACL, and in major journals. ICML 2024 Oral.

Appointments

  • 2026 - present　Associate Professor (tenured), Tongji University
  • 2023 - 2026　Assistant Professor, Waseda University
  • 2021 - 2023　JSPS Research Fellow

Education

  • 2020 - 2023　PhD in Engineering, The University of Tokyo (Advanced Interdisciplinary Studies)
  • 2018 - 2020　MS in Information Science, The University of Tokyo (Mathematical Informatics)
  • 2012 - 2017　BE, Tongji University

Research

I study state evolution, complexity, task-relevant compression, and efficient sequence representations in foundation models and agents. I also use semantic representations to model events and tail risk in complex financial markets.

01 · DYNAMICS

Agent Dynamics

Instability, order parameters, and early-warning signals in long-horizon generation and agent loops, with online regulation of internal states.

ICML 2026 02 · COMPLEXITY

Foundation Models and Complexity

Correlation dimension on statistical manifolds as a probe of pretraining phases, long-range structure, hallucination, and degeneration.

NeurIPS 2025 · PRR 2024 03 · COMPRESSION

Semantic Compression and Clustering

Information bottlenecks and conditional generative distributions for discrete indexing, document clustering, and generative retrieval.

ICML 2024 Oral · AAAI 2025 04 · REPRESENTATION

Nonlinear Semantics and Linear Language Models

Function-valued semantic fields for polysemy, and selection, writing, and finite-state memory in Mamba, DeltaNet, and related models.

NeurIPS 2022 05 · FINANCE

Semantic Models of Financial Markets

Joint representations of news, events, and prices for transferable stock geometry, portfolio construction, and tail-risk control.

ACL 2020 · KBS 2022

Prospective Students

Applications are welcome for Tongji University’s 2027 intake: recommended-admission master’s, direct-entry PhD, and regular PhD tracks.

Research topics include:

  • macroscopic states and critical transitions in long-running foundation models and agents
  • online monitoring, regulation, and training methods for generative processes
  • information compression, knowledge representation, clustering, and retrieval
  • foundational problems at the intersection of AI and complex systems

Applicants should have a solid background in mathematics and programming and sustained interest in foundational problems of large models.

Please send a CV, transcript, and brief research-interest statement to duxin@tongji.edu.cn.

Join us → Tongji University @lab