Hanzhou Liu
Hanzhou Liu
怀天下抱负 做未来主人
求实创新 励志图强

I am a 6th-year Ph.D. student in Computer Engineering at Texas A&M University.

My research interests and knowledge scope include:

  • Computer vision, including 3D vision and generation
  • Multimodal models, particularly for video generation and visual understanding
  • LLM post-training and tool-augmented reasoning

I am actively seeking full-time industry positions for 2027, and I am also very open to research internships and postdoctoral opportunities. If you are aware of any relevant openings, please feel free to reach out to me at heyhanzhou@gmail.com.

News
Work Experience
Amazon Store logo
Applied Scientist Intern Amazon Store, Seattle May 2026 – Aug 2026
Advised by Rui Song
  • Built the organization’s first tool-augmented reasoning system for seller-enforcement decisions, using a two-stage search process to retrieve supporting evidence for decision verification and rationale generation.
  • Developed an LLM post-training pipeline for seller-enforcement reasoning, combining synthetic reasoning traces, supervised fine-tuning, and GRPO-based reinforcement learning to improve rationale quality.
Urban Resilience Lab logo
Research Assistant Urban Resilience Lab, College Station Aug 2025 – Jan 2026
Advised by Ali Mostafavi
  • Developed FloodSQL-Bench, the first Text-to-SQL benchmark for flood-risk analytics, comprising 443 question–SQL pairs across six difficulty levels and 10 tables with key-based, spatial, and hybrid joins.
  • Designed a metadata-driven, multi-granularity RAG pipeline with table- and column-level retrieval, and evaluated 20+ proprietary and open-source LLMs on multi-table geospatial reasoning.
Selected Works
ICLR 2026 (8-8-6-6)
Stylos Demo
Hanzhou Liu*, Jia Huang, Mi Lu, Srikanth Saripalli, Peng Jiang*  |  ICLR 2026 Top 1.3%
Stylos couples VGGT with Gaussian Splatting for cross-view style transfer, introducing a voxel-based style loss to ensure multi-view consistency.
AAAI-SS 2026
FloodSQL-Bench Demo
Hanzhou Liu, Kai Yin, Zhitong Chen, Chenyue Liu, Ali Mostafavi  |  AAAI Symposium 2026
FloodSQL-Bench is the first Text-to-SQL benchmark for flood-risk analytics, comprising 443 question–SQL pairs across six difficulty levels and 10 tables with spatial/hybrid joins. Evaluates 20+ LLMs with a metadata-driven, multi-granularity RAG pipeline.
arXiv 2025
Lumos3D Demo
Hanzhou Liu, Peng Jiang, Jia Huang, Mi Lu  |  arXiv 2025
Lumos3D introduces a pose-free single-forward framework for 3D low-light scene restoration using a cross-illumination distillation scheme and a specialized Lumos loss, restoring structure and illumination without scene-specific optimization.
NTIRE CVPR 2025
XYScanNet Demo
Hanzhou Liu, Chengkai Liu, Jiacong Xu, Peng Jiang, Mi Lu  |  NTIRE CVPR 2025
XYScanNet maintains competitive distortion metrics and significantly improves perceptual performance. Experimental results show that XYScanNet enhances KID by 17% compared to the nearest competitor.
IRAJ 2025
DiNAT-IR Demo
Hanzhou Liu, Binghan Liu, Chengkai Liu, Mi Lu  |  International Robotics & Automation Journal
DiNAT-IR explores Dilated Neighborhood Attention (DiNA) for image restoration. By introducing a channel-aware module to complement local attention, DiNAT-IR integrates global context without sacrificing pixel-level precision, achieving competitive results across diverse benchmarks.
ICIVC 2022
Crosswalk Stripe Detection Demo
Hanzhou Liu, Mi Lu  |  ICIVC 2022
Proposes a robust crosswalk stripe detection model based on gradient similarity tags to effectively identify pedestrian crossings under challenging outdoor lighting conditions, complex shadows, and perspective distortions.
Educations
Texas A&M University logo
Ph.D. in Computer Engineering Texas A&M University Aug 2021 – 2027
Texas A&M University logo
M.S. in Computer Engineering Texas A&M University Aug 2019 – Jun 2021
Jilin University logo
B.S. in Electrical Engineering Jilin University Aug 2014 – Jun 2018
Hefei No.1 High School logo
High School Hefei No.1 High School 2011 – 2014
Community Services