Undergraduate · NYU Shanghai
I work on language models for code
and how we evaluate them.
At NYU Shanghai, I work on code generation, LLM-generated tests, and behavior-preserving repair of research software. I also work on computer vision, with experience in medical imaging and multimodal systems.
An executable benchmark of 1,196 verified tasks across 12 rare and legacy languages. Separates solution synthesis from test generation and uses a Docker-based verification pipeline with model cascading.
Advanced Capstone Project · NYU Shanghai
Building tooling for behavior-preserving repair of research code. Reconstructed historical and modern environments to compare model outputs, gradients, and one-step training updates during a PyTorch 1.4-to-2.8 migration.
An API-only migration ran all 14 configurations but preserved behavior in 10; a reference patch passed all 14 behavior checks.
Serendipity Research Group · NYU Shanghai
Studying how same-model and cross-model tests expose faults in fixed code candidates, with and without access to candidate code. Built isolated execution and a proxy oracle requiring agreement among three qualified reference programs.
A 36-problem expansion yielded 52 qualified faulty programs across 21 problems for strict test-strategy comparisons; fault detection and structural coverage are analyzed separately.
Algorithm Engineer Intern (Research) · Medical and Healthcare AI Lab
Developed nnU-Net v2 pipelines for cerebral microbleed and lacune detection in multimodal MRI. Worked on spatial alignment, data validation, controlled ablations, and lesion-level evaluation.
Serendipity Research Fellowship · NYU Shanghai
Built benchmark generation and verification tooling for rare and legacy programming languages, including Docker-based pass@k evaluation, feedback-driven resampling, and API cost controls.
Earlier research includes long-video understanding and chest X-ray interpretation. See the full CV
Course project Sep — Dec 2025
Exploring how depth representation affects image generation. Co-designed an 18-channel conditioning signal for Stable Diffusion 1.5 ControlNet, combining raw depth, 16 ordinal layers, and occlusion boundaries.
Compared with a single-channel baseline on 50k COCO images under a matched 20k-step training budget.
Research project May — Aug 2025
A prototype that turns academic figures into editable PowerPoint slides. Combines OCR, SAM segmentation, region-level vision-language classification, and recursive segmentation of small objects.
Research Assistant · CSDSE Department, NYU Shanghai
Get in touch
For research conversations and collaboration, email is the best way to reach me.
zy3077@nyu.edu