AI Cutting Edge
Stagione 1
Agent S family of computer-use agents (CUAs)
di Sujith
These sources chronicle the evolution of the Agent S family of computer-use agents (CUAs), detailing improvements in performance and methodology across several iterations. Agent S introduced hierarchical planning, memory, and a constrained action space to tackle complex desktop tasks, achieving a state-of-the-art (SoTA) result of 20.6% on the OSWorld benchmark. Agent S2 significantly advanced this by implementing Proactive Hierarchical Planning and a Mixture-of-Grounding strategy, pushing the SoTA to 48.8%. Finally, Agent S3 incorporated a native coding agent and a wide-scaling framework called Behavior Best-of-N (bBoN), which generates multiple rollouts and uses concise "behavior narratives" for principled trajectory selection, ultimately achieving a new SoTA of 69.9% and approaching human-level performance.
Lavida-O, a novel unified Masked Diffusion Model (MDM) designed for both multi-modal understanding and generation tasks, which surpasses previous MDMs in capability and performance. Unlike earlier models that were limited to simple image tasks or low-resolution output, Lavida-O leverages a single framework to handle complex tasks like high-resolution text-to-image synthesis (1024px), object grounding, and sophisticated image editing. A key innovation is the Elastic Mixture-of-Transformers (Elastic-MoT) architecture, which efficiently couples a smaller generation branch with a larger understanding branch, allowing for scalable training and flexible task-specific parameter loading. Furthermore, Lavida-O explicitly uses its understanding capabilities to enhance generation quality through planning and iterative self-reflection, establishing a new, highly efficient paradigm for integrated multimodal AI.