Zhengyan Huan

Zhengyan Huan

Department of Electrical and Computer Engineering, Tufts University

Email CV GitHub Google Scholar

About Me

I am a PhD candidate in the Department of Electrical and Computer Engineering at Tufts University, advised by Professor Shuchin Aeron. Before joining Tufts, I received an M.S. in Electrical Engineering from the University of Pennsylvania and a B.Eng. in Electronic Information Engineering from Hefei University of Technology.

Research Interests

My research interests include generative models, reinforcement learning, and inverse problems. I focus on developing reliable generative and inference methods for problems involving physical, structural, or task-specific constraints. Methodologically, I am interested in flow matching, diffusion models, and learning-based optimization, with a particular emphasis on constrained generation and inference.

Selected Projects

Comparison of generated dynamics and derivative statistics from standard flow matching and Primal-Dual Flow Matching against ground truth

Primal-Dual Flow Matching for Sample-Wise Constrained Generation

Accepted by NeurIPS 2026

Scientific generative models must match data distributions while satisfying sample-wise requirements, such as physical laws or molecular validity, at a desired rate. We develop Primal-Dual Flow Matching (PDFM), a training-time method that directly controls constraint-satisfaction probability while preserving distributional fidelity. PDFM provides theoretical guarantees on constraint satisfaction and the optimality gap and requires only a membership oracle (a yes-or-no feasibility test), without convexity, differentiable distances, or projections.

Observation ensembles mapped back to their corresponding ensemble of hidden ground-truth signals

Inverse Problems Conditioned on Observation Ensembles: Applications and Methods

2026 · TMLR · Paper

Inverse problems recover hidden signals from indirect observations, but standard methods often assume a known/fixed prior or repeatedly invoke an expensive forward model. We introduce the Ensemble-conditioned Inverse Problem (EIP), where an observation ensemble provides information about its unknown prior. Our ensemble inverse generative models condition posterior sampling on both an individual observation and a learned ensemble summary, enabling non-iterative inference without the forward model at test time and generalization to unseen priors.

Comparison of feasible training data, standard flow matching with many infeasible samples, and the proposed objective with fewer infeasible samples

Constraint-Aware Flow Matching via Randomized Exploration

2026 · TMLR · Paper · Poster

Generative models can match a target data distribution yet still produce samples that violate important rules, such as physical feasibility or required image attributes. We develop Flow Matching via Randomized Exploration (FM-RE), a training-time method that improves constraint satisfaction without changing sampling. When constraints are available only through a membership oracle (a yes-or-no feasibility test), FM-RE uses randomized exploration to learn a constraint-aware mean flow. It requires neither convex constraints, differentiable distance functions, nor projection operators.

Illustration of the quickest change detection reward function before and after the change point

Policy Gradient and Actor-Critic Approaches to Quickest Change Detection

2026 · IEEE TSP · Paper

Quickest change detection seeks to detect a change point in a sequential data stream with minimal delay while controlling false alarms, with applications to cyberattack, solar-flare, and credit-card-fraud detection. Classical methods often require unavailable pre- and post-change distributions. We develop one policy-gradient and two actor-critic methods that learn a sequential alarm policy from labeled time-series data. By defining the stopping time as the first alarm time and enforcing a “keep alarming after the first alarm” structure, our approach aligns learning with the QCD objective by focusing on the first alarm rather than independently classifying each time step.

Non-resilient and resilient robot-monitoring policies with reward and utility curves; dashed lines are non-resilient and solid lines are resilient

Resilient Constrained Reinforcement Learning

2024 · AISTATS · Paper

Constrained reinforcement learning seeks policies that maximize reward while satisfying requirements such as safety or resource limits. In practice, these constraint specifications (e.g., required thresholds) are often difficult to set and may even make the problem infeasible. We develop resilient constrained reinforcement learning, which jointly learns a policy and adjusts nominal constraint specifications by balancing reward improvement against the cost of relaxing constraints. This approach supports applications with uncertain requirements, including resource allocation and robot monitoring.

Publications

2026

2025

2024

2022