Public Signals & Selected Coverage
Selected public research and open-source signals, plus media reports around my work. These items follow the same research trajectory highlighted across this site: efficient AI systems, vision and multimodal foundation models, open-source industrial frameworks, and DreamX's current work on spatial intelligence models and systems.
Public DreamX Signals
- DreamX ECCV 2026 acceptance wave
Five papers accepted to ECCV 2026, including OmniDance as an oral presentation, strengthening the team’s recent portfolio across spatial, generative, and multimodal AI
- DreamX on the AMAP-ML GitHub Organization
AMAP’s unified spatial-intelligence team and portfolio, spanning understanding and prediction, generation and simulation, planning and action, and shared foundations
- GPG in ByteDance VERL
GPG adopted as an official reasoning reinforcement-learning algorithm in the VERL framework
- MobilityBench
Real-world route-planning agent benchmark accepted as an oral paper at KDD 2026
- DreamX-World
DreamX-World 1.0 live demo, technical report, and an open-source 5B model for one-minute interactive world generation
Agent Systems
- 多模态Agent
Multimodal Agent System
Foundation Models & Vision
- 全面超越ViT，美团、浙大等提出视觉任务统一架构VisionLLAMA
VisionLLaMA: A unified architecture for vision tasks that surpasses ViT
- 端侧实时运行、3B媲美7B！美团、浙大等提出MobileVLM V2：更快、更强的端侧视觉语言模型
MobileVLM V2: Faster and stronger mobile vision language model
- 骁龙888实时运行，美团、浙大等打造全流程移动端多模态大模型MobileVLM
MobileVLM: Real-time multimodal model running on Snapdragon 888
- 美团提出基于隐式条件位置编码的Transformer，性能优于 ViT 和 DeiT
Implicit conditional position encoding Transformer outperforming ViT and DeiT
- Twins：重新思考高效的视觉注意力模型设计
Twins: Rethinking efficient visual attention model design
Object Detection
- 更准更快的YOLOv6来了，美团出品并开源
YOLOv6: Faster and more accurate, open-sourced by Meituan
AutoML & Neural Architecture Search
- 小米AI实验室成果速递
Xiaomi AI Lab research highlights
- 雷军强推:小米造最强超分辨率算法，现已开源
Lei Jun recommends: Xiaomi’s strongest super-resolution algorithm, now open-sourced
- 超越MnasNet、Proxyless：小米提出全新神经架构搜索算法FairNAS
FairNAS: A new NAS algorithm surpassing MnasNet and Proxyless
- 两个月三项成果，对标谷歌！独家对话小米AutoML团队，如何让模型搜索更公平
Exclusive interview with Xiaomi AutoML team: Making model search fairer