Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
Researchers introduce procedural graphs, a framework that lets large language model agents autonomously evolve their execution plans. By representing tasks as interconnected nodes, agents can iteratively refine and reorganize steps, improving efficiency and adaptability. The approach reduces manual prompt engineering, supports dynamic problem solving, and demonstrates promising results on complex reasoning benchmarks across various domains, including finance, healthcare, and automation.