Researchers from Moonshot AI and Tsinghua University just introduced Prefill-as-a-Service (PrfaaS). The system breaks the requirement for expensive, high-speed local connections by offloading heavy initial memory setup to remote, specialized clusters. It uses smart scheduling
@jiqizhixin
-
Pop-up Attacks Hijack AI Assistants: LaSM Defense Method
By
–
What if a simple pop-up could completely hijack your AI assistant? Researchers from Shanghai Jiao Tong University have a new fix. They found that pop-up attacks misdirect an AI's "attention" in specific layers of its neural network. Their method, LaSM, defends by
-
AgentWard: Full-Stack Security OS for Autonomous AI Agents
By
–
Worried about the security of autonomous AI agents? Tsinghua researchers present AgentWard (玄甲). This full-stack security OS builds five coordinated defense-in-depth layers directly into the agent workflow—securing everything from startup to decision-making. It verifies
-
HumDex: Portable Teleoperation System for Humanoid Robot Training
By
–
What if training dexterous humanoids for complex, real-world tasks just got dramatically easier? USC Physical Superintelligence Lab and WorldEngine AI present HumDex! They built a portable teleoperation system that effortlessly captures human full-body and precise hand
-
LifeSim: Evaluating AI Assistants Through Realistic User Simulation
By
–
How do we truly evaluate AI assistants for long-term, real-world personalization? Researchers from Fudan University present LifeSim. It's a groundbreaking user simulator that models human cognition (Belief-Desire-Intention) in virtual environments to generate realistic,
-
AgentConductor: Dynamic Multi-Agent LLM Communication System
By
–
What if multi-agent LLMs could dynamically reconfigure their communication for any task? Researchers from Shanghai Jiao Tong University and Meituan present AgentConductor. This new system uses an LLM orchestrator and reinforcement learning to build smart, adaptive
-
Seedance 2.0 Advances Video Generation for World Complexity
By
–
Seedance 2.0: Advancing Generation for World Complexity Paper:
-
Seedance 2.0 Multimodal AI Generates Dynamic Scenes
By
–
Seedance 2.0 tech report is here! Seedance 2.0 is a groundbreaking multi-modal AI model. It unifies text, image, audio, and video inputs to generate dynamic scenes, offering deep control over motion, style, characters, and even immersive binaural audio. Seedance 2.0 delivers
-
Mind DeepResearch Technical Report on Advanced AI Methods
By
–
Mind DeepResearch Technical Report Paper: https://
arxiv.org/abs/2604.14518 -
MindDR: Multi-Agent Framework Enhances Small AI Model Performance
By
–
How can a smaller AI model punch above its weight in complex research tasks? The MindDR Team at Li Auto Inc. presents Mind DeepResearch (MindDR), a powerful multi-agent framework. It uses a collaborative trio of specialized agents (Planning, DeepSearch, Report) and a multi-stage