Abstract
Contemporary large language models (LLMs) exhibit strong capabilities for a wide range of tasks, yet their performance on complex, multi-step reasoning problems often depends critically on the prompting method and the internal reasoning strategy employed. We introduce Reasoning Router, a modular framework that dynamically selects and orchestrates among multiple reasoning strategies — such as Chain-of-Thought, Tree-of-Thought, debate-style synthesis, reflective iteration, and viewpoint aggregation — based on features of the input problem and intermediate reasoning state. The system implements an adaptive router that predicts which strategy or combination of strategies is likely to yield the most accurate and efficient solution, and then executes the chosen strategy using a graph-based orchestration layer that preserves state and supports multi-agent interactions. We evaluate Reasoning Router on standard mathematical and logical reasoning benchmarks and on curated problem categories emphasizing compositionality, ambiguity, and multi-perspective judgment. The results show consistent improvements in solution accuracy, robustness to adversarial phrasing, and interpretability of intermediate steps compared to single-strategy baselines. We analyze the router's selection behavior and perform ablation studies that highlight the complementary strengths of different strategies.
The source code is available at
github.com/Pro-GenAI/Reasoning-Router.
Keywords: Large Language Models, LLMs, Artificial Intelligence, Generative AI, reasoning, Chain-of-Thought, CoT, multi-agent orchestration