Help AI systems improve without drifting
Reliable systems should revise weak behavior, preserve what already works, and adapt across changing tasks and contexts.
Zhenhao Chen
Machine Learning PhD Student at MBZUAI · Abu Dhabi, UAE
Building AI systems that revise, adapt, and generalize.
I am a Machine Learning PhD student at MBZUAI, advised by Prof. Kun Zhang and Dr. Mingming Gong. My research spans causal reasoning, self-correction, robust representation, world models, and open-ended discovery.
Research
Research agenda
Learn structure beyond one modality
World models can capture latent structure from text, perception, interaction, and other signals without being tied to a single medium.
Stress-test behavior under shift
Robust evaluation makes capability claims concrete across style shifts, long-tail cases, and changing observation conditions.
Organize open-ended search
Scientific AI systems need memory, structure, and targeted guidance to turn exploration into reusable knowledge.
Selected work
Featured publications
View all publicationsICML 2026 Oral · FMs for Science, ICLR 2026 Workshop
CausalGame: Benchmarking Causal Thinking of LLM Agents in Games
Zhenhao Chen*, Yongqiang Chen*, Chenxi Liu*, Junchi Yu, Xiangchen Song, Zijian Li, Jialin Li, Philip Torr, Bo Han, Kun Zhang
An interactive benchmark for evaluating how LLM agents design experiments, reason from biased evidence, and recover hidden mechanisms.
LLM agentsAI scientistbenchmarkexperimentation
ICML 2025
Reflection-Window Decoding: Text Generation with Selective Refinement
Zeyu Tang*, Zhenhao Chen*, Xiangchen Song, Loka Li, Yunlong Deng, Yifan Shen, Guangyi Chen, Peter Spirtes, Kun Zhang
A decoding strategy that lets language models selectively revisit and refine past tokens during generation.
reasoningdecodingLLMs
arXiv Preprint
Confidence Matters: Revisiting Intrinsic Self-Correction Capabilities of LLMs
Loka Li*, Zhenhao Chen*, Guangyi Chen*, Yixuan Zhang, Yusheng Su, Eric Xing, Kun Zhang
A study of when models can detect and correct their own errors without external oracles, with confidence as a key factor.
self-correctionalignmentLLMs
ICML 2024
CaRiNG: Learning Temporal Causal Representation under Non-Invertible Generation
Guangyi Chen*, Yifan Shen*, Zhenhao Chen*, Xiangchen Song, Yuewen Sun, Weiran Yao, Xiao Liu, Kun Zhang
A causal representation learning approach for temporal data and video understanding under non-invertible generation.
representation learningtemporal datacausal learning
Experience
Building research systems across labs and industry
May 2026 - Present
Research Intern
TEG, Tencent
Building and training a copilot agent for the Honor of Kings series, with a focus on LLM policy learning and enhanced reasoning for strategy game copilots.
Researching LLM agents, inference-time reasoning, self-correction, alignment, and causal representation learning under the supervision of Prof. Kun Zhang and Dr. Mingming Gong.
Sep 2020 - Apr 2021
Research Assistant
Tianjin University
Contributed to a MindSpore implementation of GPT-2, including text generation, deployment tools, summarization fine-tuning, and model compression for edge devices.
Education
Academic path
Aug 2023 - Present
Doctor of Philosophy in Machine Learning
Supervisors: Prof. Kun Zhang and Dr. Mingming Gong.
Aug 2021 - Jun 2023
Master of Science in Machine Learning
GPA: 3.95/4.0.
Sep 2016 - Jun 2020
Bachelor of Engineering in Computer Science and Technology
GPA: 88/100.
Service
Academic service
Journal reviewer: IEEE Transactions on Pattern Analysis and Machine Intelligence, Pattern Recognition, ACM Computing Surveys.
Conference reviewer: NeurIPS, ICLR, ICML, AISTATS.
Student leader of the Center for Integrative Artificial Intelligence at MBZUAI, organizing weekly seminars and maintaining the official webpage.
Skills
Research and engineering toolkit
Research
LLM agentsInference-time reasoningSelf-correctionAlignmentAI scientist systemsCausal representation learningMultimodal learning
Engineering
PythonC/C++RustShellSQLLinuxDockerGit
ML Systems
PyTorchTransformersDeepSpeedvLLMAccelerateWandBPandasNumPyScikit-learn
Projects
Systems workbench is being rebuilt.
Selected AI systems projects will live here as concise case studies with problem framing, system design, evaluation notes, and links.
Open projectsWriting
Writing is being curated.
Future notes will focus on reliable AI, causal reasoning, evaluation, research taste, and the engineering of systems that adapt under shift.
Open writing