How can AI agents confidently navigate complex software without making costly mistakes? Nankai University, Nanjing University, The University of New South Wales, and Microsoft present CUWM. This Computer-Using World Model helps AI agents "see the future" of their actions on
@jiqizhixin
-
Foundation Models Learn Space Through Active Exploration Framework
By
–
How do Foundation Models truly understand a new space when they can only see parts of it? New research from Northwestern, Stanford, UWashington, & Cornell unveils crucial insights! They propose Theory of Space, a framework for AI agents to actively explore and build robust
-
REDSearcher: Novel Framework for Advanced LLM Search Agents
By
–
Tired of LLMs getting lost on deep, complex search tasks, and costing a fortune doing it? The REDSearcher Team presents a novel framework designed to make LLMs elite long-horizon search agents. It cleverly generates high-quality, complex search tasks, actively trains models to
-
Tencent Shuts Down AI Lab, Consolidates into Hunyuan Unit
By
–
Breaking: Tencent has shut down its Tencent AI Lab, folding parts of the team into its Hunyuan unit. Once a flagship AI research hub founded in 2016 with the vision “Make AI Everywhere,” the lab powered everything from game AI like “Juewu” (surpassing pro players in Honor of
-
Diffusion Models Recover Human Faces From Bedroom Training Data
By
–
Can a diffusion model trained on bedrooms actually help recover human faces? A team from Rutgers, Duke, and the University of Michigan just dropped new insights showing that even "weak" or mismatched diffusion models can still provide powerful guidance for inverse problems.
-
Toyota and Tsinghua advance robot learning with co-training strategies
By
–
How can robots master new, complex tasks and generalize effortlessly without endless, costly data collection? Toyota Research Institute and Tsinghua University just published a massive study on this! They explored co-training strategies for Large Behavior Models, teaching
-
Parallel-Probe: Faster AI Reasoning Without Performance Loss
By
–
Can we make AI reasoning much faster and more efficient without losing its smarts? Researchers from the University of Maryland, Washington University in St. Louis, and UNC Chapel Hill introduce Parallel-Probe. This innovative, training-free controller uses "2D probing" to
-
VideoWorld 2: Learning Transferable Knowledge from Videos
By
–
World 2: Learning Transferable Knowledge from Real-world Videos Paper: https://
arxiv.org/abs/2602.10102
Project: https://
maverickren.github.io/VideoWorld2.gi
thub.io/
… Our report: https://
mp.weixin.qq.com/s/FX8XjEKRrbN9
PRDNDk7C5g
… #PapersAccepted by Jiqizhixin -
VideoWorld 2: AI learns complex tasks from real-world videos
By
–
Wow, AI could learn complex, long-term tasks and transfer that knowledge, just by watching real-world videos!
— 机器之心 JIQIZHIXIN (@jiqizhixin) 19 mars 2026
Researchers from ByteDance Seed and Beijing Jiaotong University just unveiled VideoWorld 2.
Their new dynamics-enhanced Latent Dynamics Model wisely separates visual… https://t.co/Orsi7Et6td pic.twitter.com/LonQcVwm60Wow, AI could learn complex, long-term tasks and transfer that knowledge, just by watching real-world videos! Researchers from ByteDance Seed and Beijing Jiaotong University just unveiled VideoWorld 2. Their new dynamics-enhanced Latent Dynamics Model wisely separates visual
-
SkillBench: Measuring LLM Agent Skills Performance Across 86 Tasks
By
–
Do "Agent Skills" actually make your LLM agents perform better? Researchers from BenchFlow and a diverse team from multiple institutions present SkillBench, a rigorous benchmark of 86 tasks across 11 domains. It precisely measures how well 'Agent Skills'—structured procedural