🔬 Research Experience
School of Electronic and Computer Engineering, Peking University, Research Intern
2025.06.17 - Present
Supervisor: Prof. Daquan Zhou
Projects:
- Text-to-Sounding Movie Generation: Developing a high-fidelity text-to-sounding movie generation model that generates synchronized visual and auditory content from textual descriptions.
School of Statistics, Renmin University of China, Research Intern
2025.02.11 – 2025.05.16
Supervisor: Prof. Feng Zhou
Projects:
- TPP-SD: Accelerating Transformer Point Process Sampling with Speculative Decoding: Developed a novel framework that accelerates Transformer temporal point process sampling by adapting speculative decoding techniques from language models. Experiments on both synthetic and real datasets demonstrate that TPP-SD produces samples from identical distributions as standard autoregressive methods, but with 2-6× speedup.
GSAI, Renmin University of China, Research Intern
2024.03.08 – 2025.01.21
Supervisor: Prof. Hongteng Xu
Projects:
- USPTO-LLM (WWW 2025): Constructed the first chemical reaction dataset (USPTO-LLM, 247K entries) containing abundant reaction condition information by using LLM APIs to extract data from the USPTO patent database. Validated the dataset quality on graph-based and sequence-based retrosynthesis models. The dataset is open-sourced at https://zenodo.org/records/14396156.
💼 Industry Experience
AISphere, Algorithm Engineer Intern
2025.07.28 - Present
- Working on text-to-sounding movie generation models, focusing on enhancing the synchronization and quality of generated audiovisual content.
Intuitive Fosun, R&D Intern
2025.01.15 – 2025.02.15
Leveraged OCR to extract high-frequency operating parameters from the transducer screen of the daVinci surgical robot. Processed video frames using OpenCV to minimize noise from parameter fluctuations, achieving an OCR accuracy of 99%.
Trained a single-layer Transformer Encoder to extract and classify key information from 2,100 after-sales feedbacks of daVinci surgical robots into 11 categories, which greatly helped after-sales engineers identify issues and provide solutions.