I am a third-year Computer Science PhD student at USC, working with Volkan Ustun , William Swartout , Gale Lucas , and Burcin Becerik-Gerber . In summer 2026, I was an Applied Scientist Intern at Amazon , working with Peng Tang on scalable mixed user–agent co-learning with the Amazon Quick team in the Agentic AI organization.
Previously, I worked as a Software Engineer at Bubble.io , building LLM agents for no-code web development, and conducted research in robot learning at Columbia University's Creative Machines Lab with Hod Lipson .
My research goal is to create physically and socially grounded AI agents that learn from and collaborate with humans in both physical and simulated environments. I am currently particularly interested in:
- Multi-Agent Interaction modeling and enabling mixed interactions among humans, robots, and AI agents in teaming, collaboration, and cooperative/competitive settings.
- Realistic Human Simulation generating individualized and crowd-level human behaviors with behavioral realism and social awareness for simulation, game NPCs, and training environments.
- Agent Embodiment building agents that interact with the physical world through first-person perception, robotic control, and sequential decision-making.
Selected Work
Full research list →-
FlockWorld: Toward Population-Scale Multi-Agent Video World Models through Artificial Life Simulation
TL;DR A shared video world model jointly generates agents’ egocentric views, using a controllable flocking simulation to study consistency across perspectives.
World ModelsVideo GenerationMulti-AgentArtificial Life -
GameplayQA: A Benchmarking Framework for Decision-Dense POV-Synced Multi-Video Understanding of 3D Virtual Agents
TL;DR A benchmark of 2,365 questions across nine games tests how well multimodal LLMs understand synchronized agent views. The best model scores 71.3%, compared with 80.5% for humans.
Multi-AgentVideo UnderstandingBenchmarkLLMGame AI -
GraphAllocBench: A Flexible Benchmark for Preference-Conditioned Multi-Objective Policy Learning
TL;DR A suite of 19 resource-allocation problems tests how reliably reinforcement-learning agents balance competing objectives and follow user preferences.
BenchmarkReinforcement LearningMulti-Objective OptimizationGraph Neural Networks -
Learning Realistic Lip Motions for Humanoid Face Robots
TL;DR A humanoid face robot learns realistic, multilingual lip movements from human videos. Facial retargeting and a learned self-model connect observed human motion to robot actuation.
RoboticsHuman-Robot InteractionSpeechFacial Expression -
Implicit Behavioral Alignment of Language Agents in High-Stakes Crowd Simulations
TL;DR PersonaEvolve rewrites agent personas to align emergent crowd behavior with expert data, reducing divergence by 84% on average across four metrics versus unsteered agents, without model fine-tuning.
LLMCrowd SimulationMulti-AgentHuman Modeling -
Human-Robot Facial Co-expression
TL;DR Emo anticipates and co-expresses human smiles in real time, using facial-landmark retargeting and a learned inverse self-model to control a 26-actuator robot face.
RoboticsHuman-Robot InteractionFacial Expression -
Reconfigurable Robot Identification from Motion Data
TL;DR A reconfigurable robot infers its own body shape from joint motion alone. Encoder–decoder models trained on 200,000 configurations generalize to previously unseen morphologies.
RoboticsSelf-Modeling
Research community
Technical Program Committee (TPC) Member/Reviewer
- International Conference on Learning Representations (ICLR) 2024, 2025, 2026
- Association for the Advancement of Artificial Intelligence (AAAI) 2024, 2025, 2026
- Conference on Empirical Methods in Natural Language Processing (EMNLP) 2025
- International Conference on Autonomous Agents and Multiagent Systems (AAMAS) 2025, 2026
- International Conference on Machine Learning (ICML) 2026
- Conference on Neural Information Processing Systems (NeurIPS) 2026