Murray Kang

Murray Kang

CS PhD Candidate

Hello! I am a Ph.D. candidate advised by Lianhui Qin at the University of California San Diego (UCSD). I am currently a Student Researcher at Google DeepMind, working on embodied foundation models in collaboration with the Gemini Robotics team.

Previously, I was a Student Researcher on Google DeepMind’s GenAI Team during summer 2025. Before joining UCSD, I completed my undergraduate degree in computer science at the University of Washington (UW), where I worked with Luke Zettlemoyer and Terra Blevins.

Research Topics

My research goal is to build agentic AI systems that reason, act, and learn through interaction, using experience to develop better intuition and continually improve.

  • Model Architecture and Latent Reasoning: LaDiR (ICLR 2026), LaDi-RL (arXiv 2026), NF-CoT (arXiv 2026), Scaffolding Minds (arXiv 2026), and mmLaDiR (release very soon).
  • RL Algorithm and Agentic Environment: Flow of Reasoning (ICML 2025), GFlowVLM (CVPR 2025), LaDi-RL (arXiv 2026), and DeliveryGym (release very soon).
  • Embodied Foundation Models, Agents, and Simulation Environments: PAN World Model (Technical Report, 2025), AgentSpec (arXiv 2026), SimWorld Studio (arXiv 2026), DeliveryGym (release very soon), and Google DeepMind 2026 project with Gemini Robotics (internal).
  • Recursive Self-Improvement (RSI): SimWorld Studio (arXiv 2026) and DeliveryGym (release very soon).

News

  • [2026.08] We released Scaffolding Minds, my first-author paper on multimodal latent reasoning with Google DeepMind.
  • [2026.07] I gave an invited talk at Apple on my LaDiR paper.
  • [2026.06] Our VideoScience-Bench paper was accepted to ECCV 2026.
↓ SHOW MORE ↓ ↑ HIDE ↑
  • [2026.06] I started my second Student Researcher role at Google DeepMind, working on embodied foundation models with the Gemini Robotics team.
  • [2026.06] We released AgentSpec on understanding embodied agent scaffolds through controlled composition.
  • [2026.06] We released NF-CoT on latent reasoning with normalizing flows.
  • [2026.05] We released SimWorld Studio, my first-author paper on automatic environment generation with an evolving coding agent for embodied agent learning.
  • [2026.04] Our TSRBench paper was accepted to ICML 2026.
  • [2026.02] We released LaDi-RL, my first-author paper on diversity-preserving reinforcement learning with latent diffusion.
  • [2026.01] My first-author paper LaDiR was accepted to ICLR 2026.
  • [2025.11] We released PAN, a world model for general, interactable, and long-horizon simulation.
  • [2025.06] I joined Google DeepMind's GenAI Team as a Student Researcher, working on latent reasoning for vision-language models.
  • [2025.05] My first-author paper Flow of Reasoning was accepted to ICML 2025.
  • [2025.02] My first-author paper GFlowVLM was accepted to CVPR 2025.
  • [2024.09] Our FinBen paper was accepted to the NeurIPS 2024 Datasets and Benchmarks Track.
  • [2024.05] I joined Honda Research Institute as a Research Intern, working on GFlowNets for vision-language model reasoning.
  • [2024.01] My first-author paper Translate to Disambiguate was accepted to EACL 2024 as an Oral presentation.

Experience

Google DeepMind Google DeepMind | Jun 2026 – Nov 2026
Position: Student Researcher
Topic: Embodied Foundation Models
Collaboration: Gemini Robotics Team
Hosts: Abhijit Ogale and Priyank Jaini

Google DeepMind Google DeepMind | Jun 2025 – Sep 2025
Position: Student Researcher
Topic: Latent Reasoning for Vision Language Models
Host: Yinpeng Chen

Honda Research Institute Honda Research Institute, USA | May 2024 – Sep 2024
Position: Research Intern
Topic: GFlowNets Post-Training for Reasoning in Vision-Language Models
Hosts: Kwonjoon Lee and Enna Sachdeva

Feel free to reach out to me for collaborations, discussions, or even a casual chat about research!

Visitors

Visitor Count