UC Berkeley

Modern foundation models have made rapid progress in language, vision, and reasoning. However, many real-world tasks remain difficult because they require more than mapping a single input to a single output: they require decomposing ambiguous goals, coordinating multiple specialized systems, grounding intermediate representations in the visual world, and allocating computation adaptively across subtasks. This thesis studies how task decomposition with multi-agent systems can make generative and reasoning models more controllable, scalable, and effective. I approach this question through two complementary settings. First, I show how large language models can serve as planners and controllers for visual generation, translating complex text prompts into structured spatial and temporal representations, using feedback to correct generation errors, and enabling finer control over images and videos. This visual part also includes Describe Anything, which produces detailed localized descriptions for image and video regions and can support high-quality synthetic data generation for controllable image and video synthesis. Second, I extend decomposition to language-model reasoning, introducing systems for adaptive parallel reasoning and efficient threaded inference. Across these contributions, a common pattern emerges: when complex tasks are represented as interacting subproblems, models can use language, visual grounding, and parallel computation to solve tasks that are difficult for monolithic generation. Together, the methods in this dissertation advance a view of AI systems as coordinated collections of agents and modules that can plan, verify, specialize, and reason in parallel.