Haichuan Wang
I am a third-year CS PhD candidate at Harvard University, where I am fortunate to be advised by Professor Milind Tambe. My research lies at the intersection of Strategic Decision Making and Generative AI. Specifically, my interests are:
When Alignment Is Possible: Align Strategic Agents.
In reward-based alignment, such as RL post-training and controlled decoding, the reward model serves as an incentive mechanism, shaping the agent’s utility through a KL-regularized objective. From this perspective, misalignment arises when these incentives diverge from the behavior humans intend to encourage. This research thread aims to better align agents and agentic systems with human intentions by improving the incentives encoded in reward models and training objectives.
Exemplary Projects
- Optimizing Reward Model As An Incentive Mechanism (ICML 2026; EC26 Incentive-Based AI Alignment Workshop (Oral))
Reward Shaping for (Inference-Time) Alignment: A Stackelberg Game Perspective
When Alignment Is Uncertain: Detect And Correct Misalignment.
To be announced.
When Alignment Is Infeasible: Learn To Best Respond.
Responding effectively to other agents requires understanding their strategies, which can be complex, high-dimensional, and multimodal. I use diffusion and flow-based models to represent these strategies and support better decision making.
When Other Players’ Behavior Data Are Available: Learn Their Strategies And Best Respond.
When data on other players’ behavior are available, we can learn their strategies directly using a machine learning model, reducing reliance on unrealistic behavioral assumptions. Prior approaches often lack uncertainty quantification or impose restrictive parametric assumptions. Diffusion and flow-based models offer a flexible alternative: they capture complex, high-dimensional distributions and adapt to new strategic settings through conditional modeling. In the ranger–poacher game, we use these models to predict where poachers are likely to go in a national park and help rangers design patrol routes. Our results demonstrate that generative models can serve as effective behavioral models and inform decision making for societal benefit.
Exemplary Projects
- Efficiently transferring strategic knowledge across similar players using Optimal Transport (NeurIPS 2025 Spotlight)
Composite Flow Matching for RL with Shifted-Dynamics Data - Playing against learned adversary strategies via robust optimization (UAI 2025)
Robust Optimization with Diffusion Models for Green Security - Learning human behavior under data scarcity and observational bias (IAAI 2025; NeurIPS 2025 Tackling Climate Change with Machine Learning Workshop (Oral))
Generative AI Against Poaching: Latent Composite Flow Matching for Wildlife Conservation
When Other Players’ Behavior Data Are Not Available: Play The Nash Equilibrium Strategy.
When data on other players’ behavior are unavailable, a Nash equilibrium strategy is the most robust in two-player zero-sum games. In continuous-action settings, existing deep learning approaches often use simple neural network architectures that can suffer from mode collapse, limiting their ability to capture complex, multimodal strategies. We use diffusion models to parameterize players’ policies within iterative game-solving algorithms, achieving substantially stronger empirical convergence in high-dimensional action spaces.
Exemplary Project
- Computing Nash equilibria in high-dimensional continuous-action games
DiffNash: Finding Nash Equilibrium in High-Dimensional Continuous Action Spaces with Diffusion Policies
I am a MATS 10 Fellow mentored by Professor Peter Henderson, working on post-training. Previously, I graduated summa cum laude from the University of Chicago, where I won the David S. Hu Thesis Award in Economics. I had a triple major in Computer Science (Honors), Economics (Honors), and Mathematics. During my undergraduate years, I worked with Professor Haifeng Xu on game theory and with Professor Anthony Zhang and Dr. Gordon Liao on Decentralized Finance.
