Research
My research centers on large foundation models and their industrial deployment, aiming to advance AI capabilities from the perspectives of data, algorithms, and systems. The following four directions are closely connected and jointly drive next-generation multimodal intelligence in intelligent agents, recommendation, and content creation.
LLM & VLM Data Governance & Synthesis
High-quality data is the cornerstone of LLMs and vision-language models (VLMs). We build large-scale multimodal corpus engineering pipelines that collect, clean, deduplicate, and evaluate heterogeneous data (text, image, audio, and video) for pre-training and fine-tuning.
We also develop data-centric methods such as informative token selection, synthetic data generation, and automated quality inspection, which have supported the training of large foundation models at Alibaba with consistently high data-quality standards.
Representative Papers
-
Squeeze Out Tokens from Sample for Finer-Grained Data GovernancearXiv / Under review, 2025
-
Turbo: Informativity-Driven Acceleration Plug-In for Vision-Language ModelsEuropean Conference on Computer Vision (ECCV), 2025, Oral
Projects & Applications
-
Billion-Scale Multimodal Corpus Engineering
End-to-end data pipeline for Alibaba's foundation models, covering acquisition, cleaning, multilingual alignment, instruction-following data generation, and automated quality control. Enabled efficient convergence of large-scale models.
-
Data Governance Framework for Foundation Models
Patented framework for multimodal data governance, supporting deduplication, toxicity filtering, privacy desensitization, and informative-token selection to reduce training cost while improving model quality.
LLM & Multimodal Pre-training
We design pre-training paradigms and efficient training systems for large language models (LLMs) and multimodal large language models (MLLMs), with a focus on contrastive language-image learning, cross-modal alignment, and scalable distributed training.
Our work has been deployed in Alibaba's AI search and intelligent assistant, serving a large user base while substantially reducing compute cost, and has powered the AIGC platform with a large volume of generated assets.
Representative Papers
-
Advancing Myopia To Holism: Fully Contrastive Language-Image Pre-trainingIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
-
FOLDER: Accelerating Multi-Modal Large Language Models with Enhanced PerformanceInternational Conference on Computer Vision (ICCV), 2025
-
CateKV: On Sequential Consistency for Long-Context LLM Inference AccelerationInternational Conference on Machine Learning (ICML), 2025
-
Wave-Particle (Continuous–Discrete) Dualistic Visual Tokenization for Unified Understanding and GenerationarXiv / Under review, 2025
Projects & Applications
-
Unified Multimodal Pre-training Framework
Integrated vision, language, and speech modalities with large-scale global optimization and generative joint training, achieving multi-scale cross-modal semantic alignment and significant generalization gains.
-
Training and Inference Acceleration
Designed training and inference acceleration algorithms that substantially reduce computational cost while maintaining comparable model performance, including distributed parallelism, mixed-precision strategies, and efficient token selection.
Video Generation
We develop diffusion-based generative models for high-quality and controllable video synthesis, including virtual try-on, human image animation, and camera-controllable background generation.
These techniques have been integrated into Alibaba's AIGC platform, generating a large volume of assets and creating substantial business value in e-commerce content creation.
Representative Papers
-
Tunnel Try-on: Excavating Spatial-temporal Tunnels for High-quality Virtual Try-on in VideosThe 32nd ACM International Conference on Multimedia (ACM MM), 2024
-
Wear-Any-Way: Manipulable Virtual Try-on via Sparse Correspondence AlignmentThe 18th European Conference on Computer Vision (ECCV), 2024
-
Beyond Static Scenes: Camera-Controllable Background Generation for Human MotionarXiv / Under review, 2026
Projects & Applications
-
Virtual Try-On System
Diffusion-based image and video virtual try-on deployed in e-commerce scenarios, supporting manipulable pose and garment correspondence.
-
AIGC Content Generation Platform
Core algorithms of Alibaba's AIGC platform, generating a large volume of marketing assets with controllable motion and background.
Agent Systems
We build intelligent agent systems for e-commerce, including product content understanding, buyer/seller agents, and intelligent customer service.
These agents combine multimodal perception, sequential decision making, and reinforcement learning to support tasks such as automated product understanding, personalized recommendation, seller assistance, and customer service dialogue, all deployed at scale within Alibaba's ecosystem.
Representative Papers
-
Model-based Constrained MDP for Budget Allocation in Sequential Incentive MarketingThe 28th ACM International Conference on Information and Knowledge Management (CIKM), 2019
-
Learning Temporal Point Processes via Reinforcement LearningThe Thirty-second Annual Conference on Neural Information Processing Systems (NeurIPS), 2018, Spotlight
-
Tile Networks: Learning Optimal Geometric Layout for Whole-page RecommendationInternational Conference on Artificial Intelligence and Statistics (AISTATS), 2022
Projects & Applications
-
Product Content Understanding Agent
Multimodal agent systems that automatically understand and generate product content from images, videos, and text, powering listing quality and search relevance in Alibaba's e-commerce platform.
-
Buyer / Seller Agent
Intelligent agents that assist buyers in product discovery and decision making, and help sellers optimize content, pricing, and operations through automated recommendations and insights.
-
Intelligent Customer Service
LLM-based customer service agents capable of multi-turn dialogue, after-sales support, and personalized responses, deployed at scale across Alibaba's consumer services.
-
Sequential Incentive Marketing Agent
Model-based constrained MDP for budget allocation deployed in financial marketing, optimizing long-term user engagement under strict budget constraints.