Reasoning Topology Evolution: Automated Discovery of Effective Reasoning DAG Structures for LLM Agents

This paper introduces Reasoning Topology Evolution (RTE), a framework that applies evolutionary algorithms to discover reasoning graph structures for large language model (LLM) agents. While Chain-of-Thought, Tree-of-Thought, and Graph-of-Thought use fixed human-designed topologies, RTE evolves the topology itself. Reasoning strategies are encoded as directed acyclic graphs (Reasoning Genomes) where nodes represent reasoning operations and edges represent information flow. The evolutionary search starts from only linear chains and random DAGs with no hand-designed multi-path seeds. Across five independent runs on Qwen-2.5-1.5B-Instruct, evolution discovers topologies achieving 0.720 accuracy on a 50-problem held-out set, significantly outperforming linear Chain-of-Thought (0.420, p < 0.001) and random DAGs (0.360, p < 0.001), while matching hand-designed Tree-of-Thought (0.720). The fitness landscape is multimodal: different runs discover structurally distinct yet comparably effective topologies. Code, results, and all figures are included.