59.3293° N 18.0686° E Stockholm

Hand-drawn portrait of Zirong Chen

Zirong (Rex) Chen

é™ˆć­èŁ

Trustworthy AI, agentic memory, and neurosymbolic reasoningTillförlitlig AI, agentminne och neurosymboliskt resonemang

Stockholm, SE

RoleRoll
Postdoctoral FellowPostdoktor
DeptAvd
EECS-CLS
CentreCentrum
TECoSA
InstLÀrosÀte
KTH Royal Institute of Technology
BasedBaserad i
Stockholm, SE

Stockholm local timeLokal tid i Stockholm

:--

Fika with me?Fika med mig?

AboutOm mig

Hello! My name is Zirong (Rex) Chen (é™ˆć­èŁ).

Hej! Jag heter Zirong (Rex) Chen (é™ˆć­èŁ).

I am a Postdoctoral Fellow at the Department of Computing and Learning Systems (EECS-CLS) and TECoSA (Center for Trustworthy Edge Computing Systems and Applications), KTH Royal Institute of Technology, where I work with Dr. David Broman and Dr. Rafia Inam. My current research focuses on trustworthy assurance, agentic memory, neurosymbolic reasoning, and faithfulness alignment.

Jag Àr postdoktor vid Department of Computing and Learning Systems (EECS-CLS) och TECoSA (Center for Trustworthy Edge Computing Systems and Applications), KTH, dÀr jag arbetar med Dr. David Broman och Dr. Rafia Inam. Min aktuella forskning fokuserar pÄ trustworthy assurance, agentic memory, neurosymboliskt resonemang och faithfulness alignment.

I received my Ph.D. in Computer Science from Vanderbilt University in June 2026, advised by Dr. Meiyi Ma. My doctoral research focused on trustworthy NLP-powered systems for public safety, including reliable LLMs and neurosymbolic approaches to emergency response training.

Jag tog min doktorsexamen i datavetenskap vid Vanderbilt University i juni 2026 med Dr. Meiyi Ma som handledare. Min doktorsforskning fokuserade pÄ tillförlitliga system för samhÀllsskydd med sprÄkteknologi, bland annat tillförlitliga sprÄkmodeller och neurosymboliska metoder för utbildning inom rÀddningstjÀnsten.

I earned my bachelor's degree (B.E. in Computer Engineering) in China and my master's degree (M.S. in Computer Science) at Georgetown University. I was a basketball athlete in high school and college.

Jag tog min kandidatexamen (B.E. i datateknik) i Kina och min masterexamen (M.S. i datavetenskap) vid Georgetown University. Jag spelade basket bÄde pÄ gymnasiet och universitetet.

Appointments & educationAnstÀllningar och utbildning

  • Aug. 2026–presentAug. 2026–nu Postdoctoral FellowPostdoktor
    EECS-CLS & TECoSA, KTH Royal Institute of Technology, Stockholm
  • 2021–2026 Ph.D. in Computer ScienceDoktorsexamen i datavetenskap
    Vanderbilt University, Nashville, TN
    Dissertation:Avhandling: Towards Trustworthy NLP-Powered Intelligent Systems for Public Safety
  • 2019–2021 M.S. in Computer ScienceMasterexamen i datavetenskap
    Georgetown University, Washington, DC
  • 2015–2019 B.E. in Computer EngineeringKandidatexamen i datateknik
    Jinling Institute of Technology, Nanjing, China

NewsNyheter

Last updated October 2026.Uppdaterad oktober 2026.

2026
2025
  • Nov.Invited talk "Neurosymbolic AI for Safety-Critical Systems" at DTAI Lab, KU Leuven by Dr. Luc De Raedt and Dr. Jessa Bekker. See you (virtually) in Belgium!Inbjuden förelĂ€sning ”Neurosymbolic AI for Safety-Critical Systems” vid DTAI Lab, KU Leuven, pĂ„ inbjudan av Dr. Luc De Raedt och Dr. Jessa Bekker. Vi ses (digitalt) i Belgien!
  • Nov.Our LwP (Learning with Preserving) is accepted by AAAI-2026 as Oral Spotlight! See you in Singapore! Shoutout to first authors David Wang and Siwoo Bae.VĂ„r LwP (Learning with Preserving) har antagits till AAAI-2026 som Oral Spotlight! Vi ses i Singapore! Stort grattis till förstaförfattarna David Wang och Siwoo Bae.
  • Oct.Okt.Invited guest talk on our Angie work at Dr. Ehsan Hoque's ROC-HCI Lab at University of Rochester.Inbjuden gĂ€stförelĂ€sning om vĂ„rt Angie-arbete vid Dr. Ehsan Hoques ROC-HCI Lab pĂ„ University of Rochester.
  • Aug.We presented LogiDebrief (oral and poster at IJCAI) in Montreal, returned to Vandy, Anchor Down!Vi presenterade LogiDebrief (bĂ„de muntligt och som poster pĂ„ IJCAI) i Montreal och Ă€r tillbaka pĂ„ Vandy. Anchor Down!
  • Jul.I won the IJCAI volunteer scholarship!Jag fick IJCAI:s volontĂ€rstipendium!
  • Jun.Our project won the Vanderbilt Catalyst Funds!VĂ„rt projekt tilldelades Vanderbilt Catalyst Funds!
  • MayMajI joined Epiq AI Labs as an AI Research Fellow (intern), researching legal AI topics, see you in NYC!Jag började pĂ„ Epiq AI Labs som AI Research Fellow (praktikant) och forskar om juridisk AI. Vi ses i New York!
  • Apr.Our work LogiDebrief is accepted by IJCAI-2025! See you in Montreal, CA!VĂ„rt arbete LogiDebrief har antagits till IJCAI-2025! Vi ses i Montreal, Kanada!
  • Apr.Our project is awarded by Google again (Gemma3 Academic Program)!VĂ„rt projekt har fĂ„tt anslag frĂ„n Google igen (Gemma3 Academic Program)!
  • Mar.Angie now has two solutions: desktop and web!Angie finns nu i tvĂ„ versioner: skrivbord och webb!
  • Feb.Sim911 and our newly developed debriefing function are now also available online!Sim911 och vĂ„r nyutvecklade Ă„terkopplingsfunktion finns nu Ă€ven pĂ„ webben!
  • Jan.Our work Sim911 is selected as Oral Spotlight (4.9%) at AAAI-25!VĂ„rt arbete Sim911 valdes ut som Oral Spotlight (4,9 %) pĂ„ AAAI-25!
Show earlier news · 2024, 2023Visa Ă€ldre nyheter · 2024, 2023
2024
  • Dec.Our paper Sim911 is accepted by AAAI-25! See you in Philadelphia.VĂ„r artikel Sim911 har antagits till AAAI-25! Vi ses i Philadelphia.
  • Oct.Okt.Our project is awarded by the Google Academic Research Award!VĂ„rt projekt tilldelades Google Academic Research Award!
  • Sep.We were invited (permanently) to present our work at the LIVE session.Vi blev inbjudna (permanent) att presentera vĂ„rt arbete pĂ„ LIVE-sessionen.
  • Sep.We were invited to the Innovation Panel organized by Davidson County and the Sheriff's Office.Vi blev inbjudna till innovationspanelen som anordnades av Davidson County och Sheriff’s Office.
  • Aug.Our project SimER/Angie got funded by NSF (pre-announcement).VĂ„rt projekt SimER/Angie har beviljats anslag frĂ„n NSF (förhandsbesked).
  • Aug.We were awarded by the OpenAI Research Access Program.Vi tilldelades stöd genom OpenAI Research Access Program.
  • Jul.We launched and deployed SimER/Angie V2 at DEC!Vi lanserade och driftsatte SimER/Angie V2 pĂ„ DEC!
  • Jun.We hosted a Workshop on Future Technologies for Emergency Response and Public Safety!Vi arrangerade en workshop om framtidens teknik för rĂ€ddningstjĂ€nst och samhĂ€llsskydd!
  • MayMajWe were invited to the TN State Government of Education's K-12 Event at Peabody!Vi blev inbjudna till delstaten Tennessees utbildningsmyndighets K-12-evenemang pĂ„ Peabody!
  • Apr.SimER/Angie prototype first participates in real-world training!Prototypen av SimER/Angie anvĂ€ndes för första gĂ„ngen i skarp utbildning!
  • Mar.We won the Vanderbilt Generative AI Seed Grant!Vi tilldelades Vanderbilt Generative AI Seed Grant!
  • Feb.We were at AAAI 2024 in Vancouver, CA.Vi var pĂ„ AAAI 2024 i Vancouver, Kanada.
  • Feb.Our work Auto311 is selected as Spotlight Oral Presentation (6%) at AAAI 2024!VĂ„rt arbete Auto311 valdes ut till Spotlight Oral Presentation (6 %) pĂ„ AAAI 2024!
  • Jan.Serve as TA for Automated Verification.Undervisningsassistent i kursen Automated Verification.
  • Jan.I won the AAAI Scholarship and Volunteer Award!Jag fick AAAI:s stipendium och volontĂ€rutmĂ€rkelse!
2023
  • Dec.I won the Vanderbilt Graduate Student Travel Grant!Jag fick Vanderbilt Graduate Student Travel Grant!
  • Dec.Our paper Auto311 is accepted by AAAI-24! See you in Vancouver!VĂ„r artikel Auto311 har antagits till AAAI-24! Vi ses i Vancouver!

Research projectsForskningsprojekt

SimDispatch: Multi-Agent Dispatching SimulationSimDispatch: Fleragentsimulering för utlarmning

SimDispatch is a multi-agent simulation system for 9-1-1 dispatcher training. It separates control transitions from agent actions, constrains event states through a topology, and uses LLM agents with distinct personas to role-play dispatching scenarios in real time. The system is deployed at Metro Nashville Department of Emergency Communications and is under active refinement.

SimDispatch Àr ett fleragentsystem för simulering av 9-1-1-utlarmning. Det skiljer styrning av tillstÄndsövergÄngar frÄn agenternas handlingar, begrÀnsar hÀndelsetillstÄnd genom en topologi och anvÀnder sprÄkmodellbaserade agenter med olika roller för att spela upp utlarmningsscenarier i realtid. Systemet Àr driftsatt vid Metro Nashville Department of Emergency Communications och vidareutvecklas löpande.

Multi-Agent Simulation, Reliable LLMs, Safety-Critical Training

Personalized Adaptive Curriculum for 9-1-1 TrainingPersonanpassad adaptiv utbildning för 9-1-1

PACE is a personalized adaptive curriculum engine for 9-1-1 call-taker training. It frames curriculum adaptation as a recommendation problem using a contextual bandit, learning from each trainee's training history to account for different learning patterns and training needs.

PACE Àr en motor för personanpassade, adaptiva utbildningsplaner för 9-1-1-operatörer. Den formulerar anpassningen som ett rekommendationsproblem med en kontextuell banditmodell och lÀr sig av varje deltagares trÀningshistorik för att ta hÀnsyn till olika inlÀrningsmönster och utbildningsbehov.

Adaptive Learning, Trustworthy AI, Safety-Critical Training

Learn moreLĂ€s mer
PACE system architecture showing personalized adaptive curriculum engine for 9-1-1 call-taker training
Trainee proficiency modelling drives adaptive scenario difficulty across the curriculum.Modellering av deltagarens fÀrdighetsnivÄ styr scenariernas svÄrighetsgrad genom hela utbildningen.

Logic-informed Call-taking Performance DebriefingLogikstödd Äterkoppling pÄ samtalshantering

LogiDebrief combines Signal Temporal Logic with modular LLM calls to analyze and debrief 9-1-1 and 3-1-1 calls. In real-world testing, it achieves over 95% accuracy without performance degradation on long contexts. Deployed for daily operations and call-taker training, it has analyzed more than 1,701 calls and saved over 311 hours of operational time.

LogiDebrief kombinerar signaltemporal logik med modulÀra anrop till sprÄkmodeller för att analysera och ge Äterkoppling pÄ 9-1-1- och 3-1-1-samtal. I tester i verklig verksamhet uppnÄr systemet över 95 % trÀffsÀkerhet utan försÀmring vid lÄnga kontexter. Det anvÀnds i den dagliga verksamheten och i operatörsutbildning och har analyserat fler Àn 1 701 samtal och sparat över 311 arbetstimmar.

Trustworthy AI, AI Safety, Reliable LLMs

Learn moreLĂ€s mer
LogiDebrief system architecture showing LLM integration with Signal Temporal Logic for automated debriefing
LLM runtime decomposed into modular function calls, each checked against Signal Temporal Logic.SprÄkmodellens exekvering uppdelad i modulÀra funktionsanrop, vart och ett kontrollerat mot signaltemporal logik.

Customized Emergency Response TrainingSkrÀddarsydd utbildning för rÀddningstjÀnst

We develop Angie with Metro Nashville Department of Emergency Communications, bringing together PACE, LogiDebrief, Sim911, and Auto311 across more than 400 workstations with over 200 active users. Sim911, the LLM-enabled call-taking simulator, has generated over 600 simulated calls and saved more than 80 hours of instructor time. Over 90% of training participants rated its simulations as comparable or superior to human-led sessions.

Vi utvecklar Angie tillsammans med Metro Nashville Department of Emergency Communications. Plattformen förenar PACE, LogiDebrief, Sim911 och Auto311 pĂ„ fler Ă€n 400 arbetsstationer med över 200 aktiva anvĂ€ndare. Sim911, den sprĂ„kmodellbaserade samtalssimulatorn, har genererat över 600 simulerade samtal och sparat mer Ă€n 80 timmar av instruktörernas tid. Över 90 % av utbildningsdeltagarna bedömde simuleringarna som likvĂ€rdiga med eller bĂ€ttre Ă€n mĂ€nskligt ledda övningar.

Safety-Critical AI, Trustworthy LLMs, Reliable Systems

Learn moreLĂ€s mer
SimER system overview showing AI-driven simulation environment for call-taker training with Metro Nashville DEC
Angie / SimER, the simulation environment deployed with Metro Nashville Department of Emergency Communications.Angie / SimER, simuleringsmiljön som driftsatts tillsammans med Metro Nashville Department of Emergency Communications.

Trustworthy Emergency Response SystemsTillförlitliga system för rÀddningstjÀnst

Auto311 is a confidence-guided system for automated non-emergency call handling, developed with Metro Nashville's emergency communications, IT services, and government teams. It uses confidence measurements from deep neural networks to guide call handling and achieved 94.49% accuracy in emulated evaluation, outperforming direct LLM calls.

Auto311 Àr ett konfidensstyrt system för automatiserad hantering av icke-akuta samtal, utvecklat tillsammans med Metro Nashvilles larmcentral, IT-tjÀnster och förvaltning. Det anvÀnder konfidensmÄtt frÄn djupa neurala nÀtverk för att styra samtalshanteringen och uppnÄdde 94,49 % trÀffsÀkerhet i en emulerad utvÀrdering, bÀttre Àn direkta anrop till sprÄkmodeller.

AI Safety, Trustworthy ML, Reliable Decision Systems

Learn moreLĂ€s mer
Auto311 system diagram showing confidence-guided automated processing for non-emergency calls
Confidence-guided routing and handling of non-emergency 311 calls.Konfidensstyrd dirigering och hantering av icke-akuta 311-samtal.

Adaptable City Knowledge

The Adaptable City Knowledge project aims to create a dynamic, data-driven platform that helps cities respond to evolving challenges. From traffic management to disaster preparedness, this initiative fosters collaboration between city officials, businesses, and residents to promote resilience.

Projektet Adaptable City Knowledge bygger en dynamisk, datadriven plattform som hjÀlper stÀder att möta förÀndrade utmaningar. FrÄn trafikstyrning till krisberedskap frÀmjar initiativet samarbete mellan kommun, nÀringsliv och invÄnare för ökad motstÄndskraft.

AI/ML, NLP, Life-long Learning

Learn moreLĂ€s mer
CitySpec system architecture for intelligent requirement specification in smart cities with formal verification
City requirements captured in natural language and formalised into verifiable specifications.Kommunala krav fÄngade i naturligt sprÄk och formaliserade till verifierbara specifikationer.

Selected publications & manuscriptsUtvalda publikationer och manuskript

ContactKontakt

EmailE-post
zirongc [at] kth.se
OfficeKontor
LindstedtsvÀgen 30, TECoSA, KTH Royal Institute of Technology
AddressAdress
Södermalm, Stockholm, SE
LinksLĂ€nkar
Google Scholar · ORCID · LinkedIn

Visitor mapBesökskarta