Knowledge-Centric Agents Improve AI Workflow Generation

Generating workflows in visual creation systems, such as ComfyUI, demands not only syntactic correctness but also expert-level reasoning for modular compositions. Current large language model (LLM) approaches often treat this as a straightforward text-to-JSON conversion, which frequently results in structurally brittle outputs and lacks the experiential knowledge needed for effective design. This research argues that successful workflow generation requires explicitly modeling knowledge, including its structure, hierarchy, and reasoning dynamics. To address this, a knowledge-centric framework is introduced that learns to invert, inject, and infer knowledge across multiple abstraction levels. The process begins with knowledge inversion, distilling hierarchical representations—from full pseudo-codes and skeletons to high-level strategies—from extensive collections of real-world workflows. Subsequently, knowledge injection is performed through supervised fine-tuning, training the model to reason from task descriptions to strategies and then from strategies to executable structures. During inference, the model employs reversible reasoning to synthesize executable workflows, enhanced by self-refinement to ensure structural coherence. Extensive experiments demonstrate that this method yields workflows with greater node diversity, more coherent structures, and higher execution success rates compared to existing systems, establishing a new foundation for knowledge-driven, agentic workflow generation.