DeepSeek-V4 tech report is here! The team replaced standard attention with a hybrid compressed system and used a new Muon optimizer to make training faster and more stable. They also introduced improved layer connections to help the model handle complex reasoning more
@jiqizhixin
-
StreamingVLA: Parallel Vision-Language-Action for AI Robots
By
–
What if your AI robot could think and act simultaneously, instead of waiting between steps? Researchers from Tsinghua University & Lenovo present StreamingVLA. The new method lets a vision-language-action model run its "see," "think," and "act" stages in parallel, like a
-
NES: AI Framework Predicts Your Next Code Edit
By
–
What if your IDE could predict your next code edit before you even type it? Researchers at Ant Group present NES, a new AI framework that learns from past editing patterns. It uses two models: one to guess where you'll edit next, and another to suggest what to change—all
-
World Engine: Synthetic Edge Cases for Autonomous Driving Training
By
–
What if you could train a self-driving car on its hardest moments, not just its longest drives? OpenDriveLab, Huawei, NVIDIA & others present World Engine. Instead of just adding more miles of normal data, it generates massive volumes of synthetic edge cases—like cut-ins and
-
Can LLMs Truly Emulate Individual Human Online Personas?
By
–
Can LLMs truly think and act like a specific person online? Researchers from Northeastern, USC, Columbia & others present OPeRA, a new dataset that captures real people’s shopping habits—their persona, screen view, action, and internal reasoning. It’s the first public
-
Memory Intelligence Agent: AI Learning from Experience Like Humans
By
–
What if an AI could learn from its own memory like a human, getting smarter with every task? Researchers from East China Normal University, Shanghai AI Lab, and others present MIA: the Memory Intelligence Agent. It uses a "Manager-Planner-Executor" team. The Manager stores
-
Cross-Architecture Distillation Recipe for Mamba Models
By
–
Attention to Mamba: A Recipe for Cross-Architecture Distillation Paper: https://
arxiv.org/abs/2604.14191 -
Mamba Models Match Transformer Performance Without Attention
By
–
Can you get a Mamba model to perform like a Transformer without adding Attention? Researchers from Apple, MILA, and Flat Iron Institute (including Abhinav Moudgil and Ningyuan Huang) have a breakthrough answer. They introduce a two-step distillation recipe: first, they convert
-
AI Model Self-Corrects Visual Grounding Errors With Confidence Scoring
By
–
What if a model could catch and correct its own mistakes while learning? Researchers from Peking University present a new AI method for visual grounding. Instead of just matching words to image regions, their system uses a "confidence score" to flag its unreliable guesses. It
-
KMLP: Hybrid AI Model for Web-Scale Tabular Data
By
–
What if a single AI model could automatically learn from billions of messy data rows, eliminating the need for manual feature engineering? Researchers from Zhejiang University and Ant Group present KMLP, a new hybrid architecture for web-scale tabular data. It uses a clever