Jiayin Liu

Jiayin Liu

Undergraduate, Tsinghua University

I am a senior at Tsinghua University, majoring in Theoretical and Applied Mechanics with a minor in Mathematics. This fall I am a visiting research assistant at Duke University, working with Prof. Xiang Cheng. In Spring 2025 I was an exchange student at Princeton University, where I worked with Prof. Christine Allen-Blanchette. I have also worked with Prof. Tao Du at Tsinghua IIIS.

Research

I work on theory and methods for flow-based generative models, including guidance, reward-based fine-tuning, and sampling. I am also interested in stochastic processes on manifolds and in generative modeling with physical and geometric structure. My projects so far have studied the consistency between guided sampling and reward fine-tuning of diffusion models, Hamiltonian latent dynamics for trajectory and video generation, and generative design for origami.

Papers and projects

Toy example on a tree-shaped data distribution with five panels: pretrained samples without and with CFG, samples after NFT, samples after NFT with CFG added, and samples after updating the guided fieldFive panels on the same toy tree: samples with a matched reference cover the rewarded branches, the intended update field, samples with a mismatched reference concentrate on a small part, the reference error field, and samples moved by the error term alone leave the rewarded branches

GuidanceNFT: NFT-Based Reward Fine-Tuning of Guided Diffusion Models

Jiayin Liu, Zeyu Michael Li, Bingshuo Qian, Xiang Cheng

Under reviewDuke University, with Prof. Xiang Cheng · Jul 2026 – Oct 2026

Reward fine-tuning with NFT and classifier-free guidance can undo each other. We trace the failure to a mismatch in the reference velocity that NFT's two regression terms share, and correct it with a reference estimated online by flow matching on guided samples. On a toy tree-shaped distribution (above), updating the guided field moves samples toward the reward target without pushing them off the data (top row), while a mismatched reference distorts the update and concentrates samples on fewer branches (bottom row).

Project page arXiv PDF Code Models

SPS-GAN architecture: a motion sample and system information pass through a phase-space map to an initial state, a port-Hamiltonian neural network rolls it out, and the rollout is decoded by a trajectory generator or an image generatorDecoded double pendulum positions while sweeping the first active latent direction: one arm angle changesDecoded double pendulum positions while sweeping the second active latent direction: the other arm angle changesAnimation: a generated video of a double pendulumDecoded three-body positions while sweeping the first active latent direction: the bodies move along the orbitDecoded three-body positions while sweeping the second active latent direction: the bodies move radiallyAnimation: a generated video of three bodies orbiting

Discovering Phase Space Structure in Learned Hamiltonian Systems

Yulong Yang*, Jiayin Liu*, Vineet Bansal, Christine Allen-Blanchette (* equal contribution)

NeurIPS 2026Princeton University, with Prof. Christine Allen-Blanchette · Jan 2025 – Oct 2025

Structured Phase Space GAN: a conditional generative model that samples an initial state in a learned latent phase space and evolves it with a port-Hamiltonian network. Cyclic-coordinate regularization encourages a compact latent representation; analysis of its active directions reveals interpretable mechanical coordinates from trajectories or video.

Project page arXiv PDF

Method: a vertex unit (one token, with its folding condition) and the target curve feed an autoregressive diffusion model that adds one vertex unit per token; the units make the crease pattern, which is folded in simulation so its end vertex follows the targetAnimation: a generated crease pattern folding in simulation, its end vertex tracing the dashed 3D twist targetAnimation: the same design generated one vertex unit at a time; after each unit the path of the next vertex is drawn, until the end vertex follows the target

Inverse origami design for trajectory following

Tsinghua IIIS, with Prof. Tao Du · Oct 2025 – Jun 2026

Given a target 3D trajectory, generate a rigid-foldable crease pattern whose folding motion follows it, with an autoregressive diffusion model that writes the pattern one vertex unit at a time. Above: the method (top); a generated design folding along a 3D twist, and the same design built unit by unit (bottom).

Project page Code Models

Education

Honors and awards