BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:AI Center Seminar - AI Fundamentals series - Prof. Liu Zhiyuan
DTSTART:20260911T103000
DTEND:20260911T113000
DTSTAMP:20260920T204437Z
UID:a6569a3775836fbd222f60577b4a4064768cf50d5be1367bea713cdc
CATEGORIES:Conferences - Seminars
DESCRIPTION:Prof. Liu Zhiyuan\nThe talk is jointly organized by the EPFL
  AI Center and the EPFL Urban Transport Systems Laboratory (LUTS) as par
 t of the AI Fundamentals seminar series.\n\nHost: Prof. Nikolaos Gerolimi
 nis \n\nTitle \nTRIP: A Progressive Traffic Reasoning Foundation Model B
 ased on Dual Modeling of Physical and Semantic Spaces\n\nAbstract\nWhile l
 arge language models (LLMs) are transforming society\, their reliance on n
 ext-token prediction limits their adherence to the physical constraints of
  transportation systems. Addressing this from an interdisciplinary perspec
 tive\, this presentation identifies three core challenges in applying LLMs
  to transportation: information density\, physical constraint satisfaction
 \, and the disconnection between semantic reasoning and traffic physics. W
 e propose TRIP (Transport Reasoning with Intelligence Progression)\, a uni
 fied framework for complex traffic tasks. First\, we introduce an informat
 ion-theoretic "information density" metric to develop a divide-and-conquer
  architecture that integrates LLMs with domain-specific traffic models. Se
 cond\, to align semantic reasoning with physical laws\, we propose a "prog
 ressive reasoning" paradigm where a physics-embedded simulator verifies th
 ought nodes. To manage the exponential growth of the thought tree under li
 mited computational budgets\, we formulate the generation and verification
  process as a Partially Observable Markov Decision Process (POMDP) and dev
 elop an efficient dynamic-programming-based tree search algorithm. Experim
 ental results demonstrate that this method outperforms conventional reason
 ing paradigms in tasks such as intersection channelization and signal timi
 ng. Finally\, we present the deployment and engineering optimization of TR
 IP in real-world scenarios.\n\nBio\nProf. Liu Zhiyuan is currently the Cha
 ir Professor of Southeast University (SEU) and has been awarded the former
  National Science Fund for Distinguished Young Scholars. He graduated from
  Southeast University with a bachelor's degree in traffic engineering\, an
 d a Ph.D. from the Department of Civil Engineering of NUS. He worked as an
  assistant professor in the Department of Civil Engineering at Monash Univ
 ersity\, Australia. In 2015\, he returned full-time to the School of Trans
 portation of Southeast University as a professor. From May 2019 to July 20
 23\, he served as Vice Dean\, School of Transportation\, and from July 202
 3 to March 2025\, he served as Acting Dean\, Graduate College of Elite Eng
 ineers\, and is currently Secretary\, School of Statistics and Data Scienc
 e. As the first or corresponding author\, he has published more than 200 p
 apers in SCI/SSCI journals such as Nature Sustainability\, Transportation 
 Science\, TR Part B/C/E\, IEEE TKDE\, etc. Prof. Liu currently serves as t
 he Managing Editor of Multimodal Transportation\, an international journal
  published by Elsevier\, Area Editors of TR Part C and Associate Editors o
 f TR Part E.
LOCATION:ELE 117 https://plan.epfl.ch/?room==ELE%20117 https://epfl.zoom.u
 s/j/62295160272?pwd=XDxkLbTvMkORbz6zCOkRnaBTIRGJlG.1
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR