PhD · MBZUAI NLP/Abu Dhabi, UAE

Zirui Song

/ˈziːˌruː.i/ /sɔːŋ/

Building the last mile between AI and the physical world, starting with chemistry.

  • Autonomous chemistry labs
  • Reaction simulation
  • Embodied agents
  • Trustworthy models

I’m a PhD student at Mohamed bin Zayed University of Artificial Intelligence, advised by Prof. Xiuying Chen and Prof. Xiaojun Chang. My research spans multimodal AI, embodied agents, and trustworthy models. Previously, I earned my bachelor’s degree in Software Engineering with First Class Honours from the University of Technology Sydney.

My focus is shifting toward entrepreneurship. I’m exploring how AI can understand and act in the physical world, starting with chemistry. My two bets are autonomous chemistry labs that turn model decisions into experiments, and simulation models that predict how chemical reactions unfold.

Now·maintaining ChemLab-Assets·PhD at MBZUAI·open to co-founders and collaborators

Open source · actively maintained

ChemLab-Assets

203 simulation-ready chemistry laboratory instruments, for robots and LLM agents. The open-source project I build and maintain with collaborators at MBZUAI.

mbzuai-nlp / ChemLab-Assets

Agents that work in a chemistry lab need more than meshes. Every instrument here ships its geometry together with working joints, front-panel menus, an executable behaviour model, and a machine-readable contract.

A robot can operate an instrument by hand in a 3D simulator, or an LLM agent can run it from a shell, and the instrument behaves the same either way.

instruments
203
working joints
1,330
front panels
106
agent tasks with checkers
276
$ chemlab run ika_plate_rctdigital set-temp 80C
ok: set point 80.0 °C (was 25.0 °C); now 22.0 °C
$ chemlab run ika_plate_rctdigital heat on
ok: ika_plate_rctdigital heating on (set point 80.0 °C, now 22.0 °C)
# lab time moves only when the agent waits
$ chemlab wait 600
t = 150.0 s -> 750.0 s (waited 600 s)
$ chemlab run ika_plate_rctdigital temp
ika_plate_rctdigital: 80.2 °C (set point 80.0 °C); heating on
# a mistake: the error says what is wrong and how to fix it
$ chemlab run ika_plate_rctdigital set-temp 400C
error[usage/OUT_OF_RANGE]: set point 400.0 °C is outside the range (0.0 °C to 310.0 °C).
fix: give a set point from 0.0 °C to 310.0 °C.
instrument changed: no
An IKA hotplate driven through chemlab, abridged from the tour in the project README.

01 · Embodied mode

Robots operate the instruments

Doors, lids, sashes, keys, and knobs are real joints with grasp points and colliders. All 203 instruments were converted to USD and run in NVIDIA Isaac Sim, where the screen a robot sees is drawn by the instrument’s own menu engine.

02 · Agentic mode

LLM agents run the lab from a shell

One command, chemlab, puts the whole lab behind plain verbs with units and range checks. Errors say what is wrong and how to fix it, lab time moves only when the agent waits, and 276 tasks are scored by reading the instrument’s own state.

Release loglast push

  1. v1.0first release, 198 instruments
  2. v1.1reference data for 17 more instruments
  3. v1.2103 built-in reactions
  4. v1.3203 instruments, finer 3D models
  5. v1.4front-panel menus for 106 instruments
  6. v1.5chemlab CLI for agents, labware, reagents

Zirui Song · Jiadong Zhang · Jiahao Zhang · Xiaosong Ma · Xiaojun Chang · Xiuying Chen — MBZUAI. Code under MIT, assets under CC BY-NC 4.0.

What I’m building

Building the last mile between AI and the physical world.

Chemistry is where I’m placing that bet: connecting what a model predicts with what happens in a real experiment. I’m pursuing two complementary routes toward that goal.

01Act in the physical world

Autonomous chemistry labs

I want to build labs where AI can plan experiments, coordinate automated equipment, and learn from the results.

The Agentic Model would be the lab’s decision-making brain, turning scientific goals into experimental steps and adapting to observations. A catalysis foundation model would provide specialized knowledge for reasoning about catalysts and reaction conditions. Lab automation would connect those decisions to physical experiments.

02Predict the physical world

Chemical reaction simulation

My second bet is a Simulation Model that models chemical reactions and aims to predict their behavior under real experimental conditions.

The goal is to explore how changes in catalysts and reaction conditions affect outcomes before running an experiment. Predictions would guide what to test in the lab, while experimental observations would help evaluate and improve the model.

  1. Simulate
  2. Experiment
  3. Learn

Together, these routes aim to connect prediction, experimentation, and learning: use simulation to guide experiments, then use physical evidence to improve the next prediction.

Research

Selected publications

Selected work on embodied reasoning, multimodal AI, trustworthy models, and human-aligned evaluation.

* Equal contribution. See Google Scholar for the full publication list.

News

Recent updates

  • Released ChemLab-Assets v1.5.0: chemlab, a command line that lets LLM agents run all 203 instruments.
  • First release of ChemLab-Assets, simulation-ready chemistry laboratory instruments for robots and LLM agents.
  • 3 papers were accepted by EMNLP 2026.
  • 2 papers were accepted by ICML 2026.
  • 6 papers (1 oral) were accepted by ACL 2026.
Earlier updates
  • One paper was accepted by AAAI 2026.
  • One paper was accepted by NeurIPS 2026.
  • 3 papers (1 oral) were accepted by EMNLP 2025.
  • I was admitted to the degree of Bachelor of Engineering (Honours) in Software Engineering with First Class Honours.
  • I won the Dean's List 2025 prize (Top 2% of students) from UTS.
  • One paper was accepted by ECAI 2026.
  • One paper was accepted by ACL 2025.
  • One paper was accepted by Nature Computational Science.
  • Admitted to MBZUAI, where I commenced my PhD studies in August 2025.
  • One paper was accepted by NAACL 2025.
  • One paper was accepted by Communications Chemistry.
  • First day as a visiting student at MBZUAI under the supervision of Prof. Xiuying Chen.
  • One paper was accepted by EMNLP 2024.
  • One paper was accepted by ECCV 2024.
  • Prof. Ling Chen accepted me as an undergraduate research assistant at the Australian Artificial Intelligence Institute (AAII).
  • I was selected as an international exchange student majoring in Software Engineering at UTS.

Blog

Writing

Notes from research and building, posted as I go.

  • The newest posts are listed on the blog page.

All posts

Education · Experience

Background

Education

PhD student in NLP at MBZUAI, supported by the UAE Government Scholarship. Previously, Software Engineering at UTS with First Class Honours and the Dean’s List 2025 award.

Experience

Research experience at Alibaba, MBZUAI, UTS, and NTU, spanning multimodal AI, embodied agents, and trustworthy models.

Education and experience in detail

Education

2025.08 – 2029.05

MBZUAI logo PhD, Mohamed bin Zayed University of Artificial Intelligence

Expected · NLP Department · UAE Government Scholarship

2021.06 – 2025.05

UTS logo B.E. (Honours), University of Technology Sydney

Software Engineering · First Class Honours · GPA 3.90/4.00 · Dean's List 2025 (Top 2%)

Experience

2026.03 – now

Alibaba logo Alibaba Group

Algorithm Engineer (Research Intern) · Supervised by Xiang Wang

2024.09 – now

MBZUAI logo MBZUAI

Visiting student → PhD · Supervisor: Prof. Xiuying Chen · Trustworthy MLLMs

2023.10 – 2025.02

UTS logo University of Technology Sydney

Research Intern · Prof. Ling Chen & Prof. Meng Fang · Multimodal Agents

2023.03 – 2024.01

NTU logo Nanyang Technological University

Research Intern · Supervisor: Prof. Alex Kot · Multimodal

Honors and academic service

Honors & awards

  • 🎖️
    Dean's List 2025 (Top 2% of students), University of Technology Sydney · certificate
  • 🥈
    Silver Medal, Kaggle — LLM Science Exam [51 / 2664], 2024
  • 🥇
    School Second Class Scholarship, 2022

Academic service

Conference Reviewer

NeurIPS 2027ICML 2026ACL 2026ECCV 2026 EACL 2026ICLR 2026AAAI 2026EMNLP 2025 COLM 2025ACM MM 2025NeurIPS 2025ACL 2025 NAACL 2025ICME 2025IJCAI 2025EMNLP 2024

Journal Reviewer

IEEE TPAMI IEEE TAI

Let’s connect

Working on a similar question?

I’d love to hear from potential co-founders, researchers, and people working in chemistry labs who see opportunities for AI and automation. Tell me what you’re working on and where our interests might meet.

ziruisong2003@gmail.com

Based in Abu Dhabi, United Arab Emirates.