I am currently a master’s student in Computer Science and Technology at Hangzhou Dianzi University, advised by Prof. Zhou Yu. I received my bachelor’s degree from Taizhou University, where I was advised by Prof. Shiqing Zhang.
I am currently a research intern with the MNN team at Taobao, Alibaba Group, working on on-device language models, post-training, and multimodal learning.
I have received the National Scholarship, the Outstanding Graduate of Zhejiang Province award, and an Outstanding Graduation Thesis award.
Research Interests
My research focuses on large language model post-training, unified multimodal models, agentic reinforcement learning, and multimodal understanding and generation.
My other research interests include pattern recognition, image quality assessment, and multimodal affect recognition.
Experience
Alibaba Group (Taobao) · Meta Team
Algorithm Research Intern · 2025–2026
Contributed extensively to the MNN-Pocket post-training pipeline, including data, SFT, RL, MOPD, quantization, evaluation, and on-device deployment.
Built an SFT Loop Engine with failure analysis, synthetic data generation, and iterative training, improving average performance across 11 tasks by 7.46 percentage points.
Contributed to multi-teacher distillation and QAT, compressing a 302.8M-parameter model to 166 MB. MNN open-source models accumulated over 6K downloads.
Led Imp-o 1B–3B training and Sana-Turbo on-device acceleration, and advanced device–cloud agentic RL experiments.
On-device Small Models · Post-training & Deployment
As a primary developer, contributed extensively to data construction, SFT, RL, multi-teacher distillation, quantization, and evaluation to bring models to on-device deployment with MNN.
Led the on-device development and optimization of MNN-Sana-Edit-V2, including model adaptation, inference acceleration, and open-source delivery with MNN.
Downloads · Last 30 Days
—
Downloads · All Time
—
Loading download statistics…
Publications
IEEE TPAMI · AcceptedSole First Author
PerDepNet: Personality-guided Cross-domain Multitask Learning Network for Automated Video Depression Detection
Introduces personality-domain information with cross-domain multitask learning and audiovisual fusion to better model depression-related emotional cues.
Trains on-device small models to learn when to consult cloud models and internalize their guidance. Self-evolving RL, consultation cold start, and consultation-aware RL improve long-horizon task success while reducing cloud calls.