News
Recent updates from my research, open-source work, and the DreamX team. This activity log connects papers, released code, product-facing AI systems, and hiring signals to DreamX's broader spatial-intelligence mission. Public releases remain available through the AMAP-ML GitHub organization.
Recent papers Agents, world models, spatial AI, and generative AI
Open releases 40+ reproducible DreamX projects
Product reach 300M+ daily users
Latest DreamX Updates
These items are curated from recent DreamX releases and paper/project updates, with the newest and most product-relevant signals first.
- 2026.08.13 — DreamX-Phi 1.0 introduced a geometry-aware, action-conditioned video world model for robotic manipulation. At release, it ranked first on WorldArena 2.0 Track 1 and tied for second on Track 2; model weights and inference code are planned after the challenge concludes. project
- 2026.08.03 — LongHorizon-Harness introduced a Manage-Execute-Audit loop with durable verified state for reliable long-horizon computer-use agents, improving results across WeaveBench, OSWorld 2.0, and Terminal-Bench 2.1. paper
- 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. code
- 2026.07.20 — OmniDance was selected for an oral presentation at ECCV 2026, advancing multimodal dance-video generation driven by text, image, and music. code
- 2026.07.20 — DreamX added four publications: SCALAR++ in IJCV, Evaluation-Verification Reward at SIGGRAPH Asia 2026, and MAR-GRPO and Peak-End-Net at ACM MM 2026.
- 2026.06.18 — DreamX had five papers accepted to ECCV 2026, adding another strong top-venue signal to the team’s recent research portfolio.
- 2026.05.18 — MobilityBench accepted as an oral paper at KDD 2026, providing a scalable benchmark for route-planning agents in real-world mobility scenarios.
- 2026.05.12 — CoEvolve accepted to ACL 2026, training LLM agents through agent-data mutual evolution.
- 2026.05.12 — Thinking-with-Map accepted to ACL 2026 Findings, strengthening geolocalization with map-augmented reasoning.
- 2026.05.01 — UniMRG accepted to ICML 2026, showing that multi-representation generation strengthens unified multimodal understanding.
- 2026.05.01 — MIGA accepted to ICML 2026, extending pretrained video diffusion to arbitrarily long, temporally consistent videos without retraining.
- 2026.05.01 — D2Evo accepted to ICML 2026, improving data efficiency in reinforcement learning through dual difficulty-aware self-evolution.
- 2026.05.01 — E2PO accepted to ICML 2026, introducing embedding-perturbed exploration for preference optimization in flow models.
- 2026.04.22 — DCW accepted to CVPR 2026, mitigating SNR-t bias in diffusion probabilistic models.
- 2026.04.10 — SkillClaw released an agentic evolver that turns real interaction traces into reusable skill libraries.
- 2026.04.01 — MACE-Dance accepted to SIGGRAPH 2026, decoupling motion generation and appearance synthesis for music-driven dance video.
- 2026.03.23 — Omni-WorldBench released a benchmark for interactive response capabilities of world models.
- 2026.02.06 — GPG accepted to ICLR 2026 and adopted by ByteDance’s VERL as an official reasoning RL algorithm.
- 2026.02.06 — Tree-GRPO accepted to ICLR 2026, replacing independent chain rollouts with tree-search rollouts for LLM agent reinforcement learning.
Selected GitHub Portfolio
The AMAP-ML GitHub organization hosts the DreamX open-source portfolio. The work is organized around three core problems — understand and predict, generate and simulate, and plan and act — supported by a shared foundation of spatial data, multimodal models, reinforcement learning, infrastructure, and evaluation. Selected flagship releases:
Computer-Use Agents
Verified long-horizon computer use through durable task state and a Manage-Execute-Audit loop.
World Models
A general-purpose interactive world model for controllable world simulation.
Spatial AI
Route-planning agent evaluation grounded in real-world mobility scenarios.
Full release index: visit github.com/AMAP-ML for the complete repository list, pinned releases, project pages, and hiring notes.
2025
- USP accepted to ICCV 2025, proposing unified self-supervised pretraining for image generation and understanding. paper · code
- DreamX continued hiring for interns, full-time researchers, and AI engineers in spatial intelligence, LLM agents, reinforcement learning, world models, multimodal learning, embodied AI, recommendation, and generative AI. team
Earlier Highlights
- VisionLLaMA accepted to ECCV 2024, a unified LLaMA-style backbone for vision tasks. paper · code
- MobileVLM released, bringing real-time vision-language models to mobile devices. paper · code
- YOLOv6 open-sourced, an industrial-grade real-time object detection framework. code
- Twins accepted to NeurIPS 2021, revisiting spatial attention design in Vision Transformers. paper · code