PIXAR: Auto-Regressive Language Modeling in Pixel Space
We present PIXAR, the first pixel-based autoregressive LLM that understands and generates text-in-images, reaching GPT-2–level performance without relying on symbolic tokenization.
CV
Education
- Ph.D in Designing Responsible NLP, University of Edinburgh, 2029 (expected)
- MSc in Computer Science, University of Edinburgh, 2023
- BEng in Electronics and Computer Science, University of Edinburgh, 2021
Work experience
- 2022-2023: Machine Learning Engineer
- ByteDance, Beijing
- Duties includes: intigrate multimodal features into the recommendation system
- 2021-2022: Natural Language Processing Engineer
- ByteDance, Shenzhen
- Duties included: develop a pipeline of training and deployment of BERT-based models on a large scale.
- Summer 2020: Game Programmer Intern
- NetEase Games, Guangzhou
- Duties included: build tools to facilitate game numerical design.
Publications
Learning to Simulate Natural Language Feedback for Interactive Semantic Parsing
We propose simulating natural language feedback for interactive semantic parsing, enabling scalable training without costly human annotations and improving text-to-SQL error correction.
Brain tumor segmentation with attention-based U-Net
We present an attention-enhanced U-Net model integrating Squeeze-and-Excitation and CBAM modules to improve MRI brain tumor segmentation, delivered with an app featuring text-to-speech and chatbot support.