Agentic Workflows Generation Based on Meta-Cognitive Chain-of-Thought Guided Monte Carlo Tree Search

Large language models(LLMs) demonstrate exceptional performance in understanding and generating language but face challenges in tackling complex decision-making tasks. Researchers have introduced LLM-based agent workflows to address real-world complex problems. However, these workflows rely heavily on manual design, which is not only time-intensive but also requires significant expertise and lacks adaptability to diverse scenarios and dynamic environments. This paper introduces a novel approach for automatic generation of agentic workflows powered by LLMs, leveraging a meta-cognitive based Chain-of-Thought (MC-CoT) guidance integrated with Monte Carlo Tree Search (MCTS). By combining MC-CoT for structured reasoning with MCTS for optimization, the framework dynamically adjusts workflows, delivering superior performance across benchmarks such as GSM8K and MATH. Experimental results demonstrate the system's ability to efficiently navigate complex problem spaces, making it a promising foundation for advancing automated workflows in industrial, reasoning, and computational domains.