Researchers introduce a novel technique, Mean-Variance Split Residuals, enabling the training of Diffusion Transformers with an unprecedented 1000 layers, pushing the boundaries of generative model architecture.
A new framework, MACE-Dance, utilizes cascaded expert models to generate high-quality, music-driven dance videos by separately handling motion and appearance, showcasing advanced multimodal AI capabilities.
This paper presents Flow-OPD, a method for on-policy distillation that significantly improves the efficiency and performance of flow matching models, crucial for advanced generative AI applications.
Robinhood is reportedly preparing for a second retail venture IPO, capitalizing on the current AI market rally and investor interest in technology-driven financial services.
Google Finance is rolling out its new AI-powered features to Europe, offering enhanced financial insights and tools to users across the continent.
OpenAI provides insights into the significant expansion of ChatGPT's adoption across various sectors and user demographics during the first quarter of 2026.
HyperEyes introduces an efficiency-aware reinforcement learning approach for multimodal search agents, optimizing performance across various granularities for complex tasks.
A new research direction explores how Large Language Models can autonomously improve other LLMs through agentic discovery, enabling more robust and scalable AI systems during testing.
HumanNet presents a massive dataset and methodology for scaling human-centric video learning to one million hours, promising significant advancements in understanding human behavior and interaction.
This article delves into the critical aspects of securing AI agents in production environments, analyzing the strengths and weaknesses of the Multi-Agent Control Protocol (MCP).
Veteran social news site Digg is attempting a comeback by rebranding itself as an AI-powered news aggregator, aiming to leverage artificial intelligence for content curation.
OpenAI shares a guide detailing strategies and best practices for enterprises looking to effectively scale their AI initiatives and integrate advanced models into their operations.
DTap is introduced as a new platform designed for controllable and interactive red-teaming of AI agents, enhancing the ability to identify and mitigate risks in AI systems.
An insightful piece argues that beyond syntax and structure, the 'thinking quality' embedded in prompt design is the crucial, often overlooked, layer for effective prompt engineering.
A developer shares a hands-on review of PaioClaw, detailing its performance and limitations when subjected to rigorous testing scenarios.