Graph of Thoughts (GoT) in Agentic Workflows for Non-Linear Problem Solving

ABSTRACT

While Chain-of-Thought and Tree-of-Thoughts prompting greatly enhance LLM reasoning, they strictly enforce linear or hierarchical cognitive paths. We introduce Graph of Thoughts (GoT), a novel cognitive architecture for Agentic AI that models information generated by an LLM as an arbitrary graph. This allows agents to synergize multiple independent thoughts, loop back to refine previous states, and distribute sub-tasks across a swarm asynchronously. Evaluations on complex sorting and document merging tasks show that GoT-enabled agents outperform Tree-of-Thoughts by 62% while reducing overall token volume.