Senior Director & Head of DreamX, Alibaba
I build foundation AI systems that move from original research to reproducible open source and large-scale map, mobility, and interactive AI products. I lead DreamX at Alibaba AMAP, a 100+ member product-facing AI team building spatial intelligence models and systems that understand and predict, generate and simulate, plan and act in the real world. Our work is grounded in AMAP products serving 300M+ users every day. My research record includes 100+ papers published or accepted at leading AI conferences and journals.
16,000+ Total Citations
6,000+ As First Author
100+ Top-Conference & Journal Papers
300M+ Daily Users
Citations: Google Scholar · Publication count includes confirmed acceptances as of September 27, 2026; see Publications for the counting scope.
Recent Updates
2026.09.26 DreamX had 14 papers accepted to NeurIPS 2026: 11 in the Main Track and 3 in the Evaluations & Datasets Track, including one Spotlight. Team announcement. 2026.09.03 DreamX-Creator 1.0 released model weights and inference code for native joint audio-video generation with a 7B model and an autoregressive one-step 2K refiner. 2026.08.13 DreamX-Phi 1.0 introduced a geometry-aware, action-conditioned video world model for robotic manipulation and reported leading results on WorldArena 2.0. 2026.08.03 LongHorizon-Harness introduced a Manage-Execute-Audit loop with durable verified state for reliable long-horizon computer-use agents. 2026.07.23 DreamX-World 1.0 went live as an interactive world model, following the release of its technical report and open-source 5B model supporting one-minute generation. View all updates
Featured Projects
Current Focus
LLM ReasoningA simple and strong reinforcement learning baseline for model reasoning — no critic, no reference model, no KL penalty. Adopted by ByteDance's
VERL framework as an official algorithm.
World ModelA general-purpose interactive world model that creates diverse, high-fidelity virtual environments with camera-controlled navigation and prompt-driven world events.
Spatial AIA scalable benchmark for evaluating route-planning agents in real-world mobility scenarios, connecting agent research with AMAP's spatial intelligence anchor.
AgentAgentic skill evolution from real interaction traces, turning reusable skills into collective libraries across sessions, devices, and agents.
Earlier Impact
DetectionIndustrial-grade real-time object detection framework with a full training-to-deployment toolchain, broad open-source adoption, and follow-up deployment work on RepVGG-style quantization.
Vision-LanguageA compact vision-language assistant designed for real-time on-device deployment, with 1B/3B models evaluated on mobile hardware such as Snapdragon 888.
ArchitectureSystematically interleaves local and global attention in a general-purpose vision backbone, outperforming Swin with a simpler, deployment-friendly design. The same compute-context pattern is now common in efficient LLM architectures.
FoundationA unified LLaMA-style vision backbone introducing AS2DRoPE, an early resolution-aware extension of 2D rotary position embedding validated across generation and understanding. This multidimensional RoPE design later became mainstream in frontier multimodal and generative Transformers.
Research Journey
2024 – Present · Alibaba AMAP
DreamX: Spatial Intelligence Models & Systems
Leading DreamX, a 100+ member product-facing AI team organized around one mission: spatial intelligence. The portfolio connects three core problems — understanding and predicting the world, generating and simulating the world, and planning and acting in the world — with shared foundations in spatial data, multimodal models, reinforcement learning, infrastructure, and evaluation. Representative systems include DreamX-Phi, DreamX-World, MobilityBench, LongHorizon-Harness, SkillClaw, and GPG. DreamX Agent powers Navigation Live, while AIGC and Creator support visual content enhancement for AMAP's Saojiebang and physically grounded video generation. The work contributes to products serving 300M+ users every day. Published 60+ papers at top venues and open-sourced 40+ projects.
2020 – 2024 · Meituan
Vision Transformers, Multimodal Models & Industrial AI
Built the Visual Intelligence team from scratch. Created Twins (NeurIPS 2021), CPVT (ICLR 2023), VisionLLaMA (ECCV 2024); reproduced LLaMA 7B and built MobileVLM for on-device deployment; open-sourced YOLOv6; shipped autonomous delivery and drone perception systems.
2017 – 2020 · Xiaomi
Neural Architecture Search & AutoML
Founded Xiaomi's AutoML team. Produced a series of influential NAS works — FairNAS (ICCV 2021), FairDARTS (ECCV 2020), DARTS- (ICLR 2021), FALSR — establishing new standards for fair and robust architecture search. Featured by Lei Jun.
2013 – 2017 · KingStar
Power Grid AI & Reinforcement Learning
Core contributor to the "Complex Power Grid Autonomous-Collaborative Automatic Voltage Control" project. Contributed 20 invention patents. Awarded the National Science and Technology Progress First Prize (2018).
2012 – 2013 · IBM Research China
Large-Scale Data Analytics
Research scientist working on large-scale data analytics and machine learning solutions.
Recognition
- Top 100 AI Scholars, AMiner 2023 — selected from hundreds of thousands of AI researchers worldwide
- 3 first-authored papers on PaperDigest's Most Influential lists: FairNAS, Twins, CPVT
- Area Chair: ICLR, NeurIPS | Senior Program Committee: AAAI, IJCAI
- 50+ invention patent families with Chinese filings, including 9 with international filings
DreamX Technical System
Understand & Predict the World — Route-planning agents (MobilityBench), map-augmented geolocalization (Thinking-with-Map), urban scene understanding, mobility forecasting, recommendation, and industrial map systems
Generate & Simulate the World — Native joint audio-video generation with 2K refinement (DreamX-Creator), interactive world simulation (DreamX-World), GUI world models (Code2World), scene-text editing (FluxText), and 3D editing (RL3DEdit)
Plan & Act in the World — Action-conditioned robotic world modeling (DreamX-Phi), verified long-horizon computer use (LongHorizon-Harness), agent skill evolution (SkillClaw), reasoning and self-reflection for physical action (AutoDrive-R2), and LLM reasoning (GPG)
Shared Foundation — Spatial data and knowledge, multimodal foundation models, reinforcement learning, generative modeling, infrastructure, and evaluation across projects including SpatialGenEval, Omni-WorldBench, FASA, USP, and RealQA
Team & Opportunities
I lead DreamX at Alibaba, a 100+ member product-facing AI team building AMAP’s spatial intelligence models and systems. DreamX releases are hosted publicly through the AMAP-ML GitHub organization.
Our philosophy: We build systems where research quality, engineering discipline, open-source reproducibility, and product deployment reinforce each other. Many core projects ship with reproducible code, and our work contributes to AMAP products serving 300M+ users every day.
Open Source
We maintain 40+ DreamX projects on GitHub, organized around understanding and prediction, generation and simulation, planning and action, and a shared technical foundation.
Hiring
We are always looking for talented interns, full-time researchers, and AI engineers in LLM agents, reinforcement learning, world models, multimodal learning, spatial intelligence, and generative AI. Drop me an email if interested.
Education
- M.S. in Electrical Engineering, Tsinghua University, 2012
- B.S. in Electrical Engineering, Southeast University, 2010
