Understand, predict, simulate, create, plan, and act in the real world.
DreamX is the unified spatial intelligence model and system portfolio of AMAP-ML, the AI team at Alibaba AMAP. We connect research, engineering, and real-world deployment across maps, mobility, local services, digital content, and interactive worlds.

We advance DreamX through production systems, open-source projects, benchmarks, and publications at ICLR, CVPR, ECCV, ACL, AAAI, SIGGRAPH, ICCV, ICML, KDD, EMNLP, ACM MM, and WWW. We release code and evaluation assets to help the community reproduce, compare, and extend our work.

Join us | Research interns, full-time researchers, and AI engineers in spatial intelligence, LLM agents, reinforcement learning, world models, multimodal learning, embodied AI, recommendation, and generative AI are welcome to get in touch.

30+ Open-source Projects  |  11 ICLR 2026 Papers  |  10 CVPR 2026 Papers  |  5 ECCV 2026 Papers
7 ACL 2026 Papers  |  5 AAAI 2026 Papers  |  4 ICML 2026 Papers  |  1 KDD 2026 Oral Paper
5 ICCV 2025 Papers  |  2 EMNLP 2025 Oral Papers
Focus: Spatial Intelligence · DreamX · LLM Agents · World Models · Multimodal and Generative AI

We define spatial intelligence as the ability of AI to understand the real world and its evolution over space and time; predict future states; generate and simulate digital representations; plan toward human goals; and act through products and embodied systems.

For AMAP, this means connecting maps, mobility, urban environments, local services, digital content, and physical action in one learning and deployment loop. Generative intelligence is an essential part of this mission: it enables AI to represent, create, enrich, and simulate the world rather than standing as a separate product anchor.

Our work is organized around three core problems:

DreamX turns this mission into six model and system families. Each family focuses on a distinct relationship between intelligence and the real world:

This table defines the portfolio at the family level. Public same-name releases and related research artifacts are linked below where available.

The six DreamX families build on a common foundation:

These selected public projects show how the spatial intelligence mission translates into research artifacts, open-source systems, benchmarks, and models. The complete project map below organizes all public artifacts by the three core problems and the shared foundation.

• 2026.06.18 AMAP-ML has 5 papers accepted to ECCV 2026, expanding the team's work across spatial intelligence, generative modeling, and multimodal AI.
• 2026.06.15 DreamX-World releases its 1.0 technical report and open-sources DreamX-World-5B for long-horizon interactive world generation with 1-minute video support.
• 2026.05.18 MobilityBench provides a scalable benchmark for evaluating route-planning agents in real-world mobility scenarios (KDD 2026 Oral).
• 2026.05.12 CoEvolve trains LLM agents through agent-data mutual evolution, using failure signals to synthesize harder tasks as the agent improves (ACL 2026).
• 2026.05.12 Thinking-with-Map strengthens geolocalization with a reinforced parallel map-augmented reasoning agent (ACL 2026 Findings).
• 2026.05.11 DreamX-World releases the 5B-Cam model and inference code for general-purpose interactive world simulation.
• 2026.05.01 UniMRG shows that multi-representation generation strengthens understanding in unified multimodal models, not just generation (ICML 2026).
• 2026.05.01 Train-Free Infinite-Frame Generation extends pretrained video diffusion to arbitrarily long, temporally consistent videos without any retraining (ICML 2026).
• 2026.05.01 D-Evo improves data efficiency in RL with dual difficulty-aware self-evolution that adaptively reshapes both task and sample difficulty (ICML 2026).
• 2026.05.01 EEPO introduces embedding-perturbed exploration for preference optimization in flow models, addressing exploration collapse in continuous generative policies (ICML 2026).
• 2026.04.22 DCW mitigates SNR-t bias and improves diffusion generation quality across model families (CVPR 2026).
• 2026.04.22 EMF extends efficient one-step generation from class-conditioned synthesis to text-conditioned image generation (CVPR 2026).
• 2026.04.10 SkillClaw turns real interaction traces into reusable, evolving skill libraries.
• 2026.04.01 MACE-Dance decouples motion generation and appearance synthesis for high-quality music-driven dance video (SIGGRAPH 2026).
• 2026.03.23 Omni-WorldBench evaluates world models in dynamic 4D interactive settings.
• 2026.03.20 AutoDrive-R2 improves VLA models with reasoning and self-reflection for autonomous driving scenarios (ICLR 2026).
• 2026.03.18 Video-STAR uses tool-augmented reinforcement learning for open-vocabulary action recognition in video (ICLR 2026).
• 2026.03.11 RL3DEdit uses geometry-guided reinforcement learning to make 3D scene edits more multi-view consistent (CVPR 2026).

The project map complements the family-level view with the research and open-source foundations behind DreamX. Each artifact is organized by the primary role it plays in the spatial intelligence stack.

We are looking for people who want to build the next generation of spatial intelligence systems through clean code, reproducible experiments, rigorous evaluation, ambitious problem selection, and real-world product impact.

If you are interested in research internships, full-time roles, or academic collaboration, please email cxxgtxy@gmail.com (homepage) with your CV, representative projects, and research interests.