Xinyao Liao
Hi, I'm Xinyao Liao.
I am a PhD student in Computer Science at the National University of Singapore, working with Prof. Angela Yao in the Computer Vision & Machine Learning (CVML) Group. My research explores visual generative models, reinforcement learning, and learning systems that can improve at inference time and through experience.
Now
Recent updates
A compact timeline of research and academic milestones.
Featured research
Learning to generate — and improve.
I am especially interested in objectives and inference procedures that make generative models more reliable, adaptive, and efficient.
Visual Prefix Guidance
Training-free inference-time guidance that contrasts generated and corrupted prefixes to reduce autoregressive prefix drift.
Variational Policy Alignment
Pixel-aware reinforcement learning for autoregressive image generation, aligning token policies with image-space quality.
Step-Level Reward for T2I RL
Turning trajectory-level supervision into step-level learning signals for more effective reinforcement learning of diffusion models.
Publications
Research archive
VPG: Visual Prefix Guidance for Autoregressive Image and Video Generation
arXiv 2026Visual Prefix Guidance is a training-free inference-time method for visual autoregressive generation that contrasts predictions from generated and corrupted prefixes to reduce prefix drift and exposure bias.
Beyond research
Curious about systems, ideas, and the world around them.
Before my PhD, I studied Computer Science / Information Security and Philosophy at HUST. Outside research, I enjoy reading, films, music, climbing, and building small things for fun.