Foundation-Model Decision Agents
Grounding language and multimodal foundation models as agents that retrieve evidence, plan hierarchically, and make decisions in interactive environments.
Explore representative workHi, my name is Zhi Wang, an Associate Professor and PhD Supervisor at Nanjing University. Previously, I received the PhD degree from City University of Hong Kong and the Bachelor degree from Nanjing University. I was a visiting scholar at University of New South Wales, Nanyang Technological University, and Chinese Academy of Sciences.
My research focuses on developing general-purpose decision-making agents across reinforcement learning, post-training and alignment of foundation models, vision-language-action models, and embodied AI. My long-term goal is to bridge foundation models and embodied intelligence, enabling agents to perceive, reason, decide, and act autonomously in complex real-world environments.
Research
Post-Training & Alignment of Foundation Models
Using reinforcement learning to improve reasoning, exploration, and output diversity in language and generative models through principled reward and optimization design.
Explore representative workSelf-Improving Agents
Designing agents that autonomously refine their evidence, reasoning, and tool-use workflows through iterative interaction and feedback.
Explore representative workGeneralizable & Continual Reinforcement Learning
Developing agents that generalize across tasks and adapt over time through in-context, offline, meta, continual, and lifelong reinforcement learning.
Explore representative workNews
- ICML 2026
Two papers on RL for LLM reasoning and agentic RAG were accepted, including one Oral paper.
- ICLR 2026
Three papers on RL for LLM reasoning and in-context RL were accepted.
- NeurIPS 2025
Three papers on in-context RL, language agents, and RL for LLM reasoning were accepted.
- ICML 2025
One paper on hierarchical LLM agents was accepted.
- TPAMI
One paper on interpretable multi-agent reinforcement learning was accepted.
Earlier news
- NeurIPS 2024
One paper on generalist RL agents was accepted.
- ICLR 2024
One paper on efficient multi-agent RL coordination was accepted.
Representative Publications
Foundation-Model Decision Agents
ICML 2025
Divide and Conquer: Grounding LLMs as Efficient Decision-Making Agents via Offline Hierarchical Reinforcement Learning
GLIDER grounds language models as efficient decision-making agents through offline hierarchical reinforcement learning.
Post-Training & Alignment of Foundation Models
ICML 2026 · Oral
Characterizing, Evaluating, and Optimizing Complex Reasoning
A systematic study of how complex reasoning can be characterized, evaluated, and optimized.
ICLR 2026
Diversity-Incentivized Exploration for Versatile Reasoning
Global diversity incentives promote deep exploration for reasoning in a semantically structured space.
ICLR 2026
ExGRPO: Learning to Reason from Experience
Valuable reasoning experiences are prioritized using rollout correctness and entropy to balance exploration and exploitation.
arXiv 2025
The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models
Understanding entropy dynamics motivates targeted controls that preserve exploration during reinforcement learning.
arXiv 2026
Beyond the Dirac Delta: Mitigating Diversity Collapse in Reinforcement Fine-Tuning for Versatile Image Generation
Diversity incentives reconcile strong task alignment with varied outputs throughout on-policy fine-tuning.
Self-Improving Agents
ICML 2026
From Conflict to Consensus: Boosting Medical Reasoning via Multi-Round Agentic RAG
MA-RAG enables self-improving medical reasoning by iteratively refining external evidence and internal reasoning history.
Generalizable & Continual Reinforcement Learning
ICLR 2026
Scalable In-Context Q-Learning
Dynamic programming and world modeling steer in-context RL toward efficient reward maximization and generalization.
NeurIPS 2025
Mixture-of-Experts Meets In-Context Reinforcement Learning
T2MIR combines mixture-of-experts modeling with in-context reinforcement learning to improve scalable task generalization.
NeurIPS 2024
Meta-DT: Offline Meta-RL as Conditional Sequence Modeling with World Model Disentanglement
Transformer sequence modeling and disentangled world models enable efficient generalization in offline meta-RL.
Academic Service
Program Chair
IEEE International Conference on Cybernetics 2026
Area Chair
NeurIPS & AAMAS
Guest Editor
IEEE Transactions on Cybernetics
Associate Editor
IEEE SMC & IEEE ICNSC
Teaching
Course information, announcements, and lecture materials for students at Nanjing University.