What if LLMs could reason without writing out every single step? Researchers from Harbin Institute of Technology propose a new method called Latent Reasoning Tuning (LRT). Instead of generating long, costly text chains, LRT teaches the model to use compact internal
@jiqizhixin
-
M4Human: Millimeter-Wave Radar Detects 3D Body Motion
By
–
What if you could track a person's full 3D body shape through walls, in the dark, and with total privacy? Researchers from NTU, Edinburgh, MIT, and others present M4Human! They built the world's largest radar dataset for human motion. It uses millimeter-wave radar signals to
-
Single AI Teaches Robot Hands Grasp Any Object
By
–
What if a single AI could teach any robot hand to grasp any object? Researchers from the National University of Singapore and Zhejiang University present T(R,O) Grasp. Their new method uses a "graph diffusion" model to learn the spatial relationship between a hand and an
-
WildClawBench: Testing AI Agents in Real Computer Environments
By
–
Can your AI assistant actually do real work, or just answer questions? The InternLM team at Shanghai AI Lab presents WildClawBench, a new benchmark that ditches simple Q&A. It tests AI agents in a real computer environment (with a browser, terminal, and files) where they must
-
PiLoT: AI System Enables Drones to Locate Themselves Using Camera
By
–
What if a drone could find itself and any target on a map using just a camera?
— 机器之心 JIQIZHIXIN (@jiqizhixin) 21 avril 2026
Researchers from the National University of Defense Tech, Zhejiang & Westlake Universities present PiLoT.
It's a new AI system that matches live drone video directly to a 3D map, like visual GPS,… pic.twitter.com/NKocOKB2JMWhat if a drone could find itself and any target on a map using just a camera? Researchers from the National University of Defense Tech, Zhejiang & Westlake Universities present PiLoT. It's a new AI system that matches live drone video directly to a 3D map, like visual GPS,
-
AdaGen: AI-Powered Optimization for Faster Image Generation
By
–
Can AI learn to optimize its own creative process for faster and higher-quality image generation? AdaGen is here to move beyond the rigid, manually designed rules currently used in models like Diffusion and Transformers. AdaGen uses a lightweight reinforcement learning policy
-
OneStory: Maintaining Plot Consistency in AI Video Generation
By
–
How do we stop AI video generators from forgetting the plot between camera cuts? Meta AI and the University of Copenhagen introduce OneStory to solve the challenge of multi-shot consistency. The model treats video creation as a next-shot task, using a smart memory system to
-
TDM-R1: Fast AI Image Generator Learning from Human Feedback
By
–
What if fast AI image generators could learn from simple human likes and object counts as easily as they learn from complex math? Researchers from HKUST, CUHK Shenzhen, and Xiaohongshu present TDM-R1 to do just that. Most lightning-fast AI models struggle to use real-world
-
Elementary Functions From Single Operator Mathematical Discovery
By
–
All elementary functions from a single operator Paper: https://
arxiv.org/pdf/2603.21852
v2
… Our report: https://
mp.weixin.qq.com/s/vl0fyvKVaLJd
LNqA6t-NqA
…
