Causal Dependency-Aware Unsupervised Routing for Large Reasoning Models

As Large Language Model (LLM) ecosystems grow, routing queries to the most suitable model in a diverse pool has become a critical strategy for building efficient and high-performing AI systems. A common approach is to train a supervised router; however, this requires vast, expensive human-annotated preference data and creates models that are notoriously brittle, failing to generalize when faced with inevitable distribution shifts in user queries. Consequently, developing robust, unsupervised routing methods that adapt without retraining is a crucial research frontier. This challenge is severely amplified by Large Reasoning Models (LRMs), which introduce a dual problem for any label-free method: their outputs have a causal thinking $\to$ answer structure that must be modeled, and a structural imbalance where long reasoning text can dominate the final answer signal. We introduce ReasoningRouter, a novel framework that resolves these issues with a length-balanced embedding strategy and a probabilistic model capturing the thinking-to-answer dependency. The proposed Causal Triangulation Property enables the label-free estimation of component qualities and their causal link. Beyond competitive routing accuracy, ReasoningRouter offers unprecedented insights into model behavior, enabling separate quality assessment of reasoning and answer components while maintaining computational efficiency.

Modern AI reasoning models are designed to think before they speak, but choosing the best model to answer a specific question without relying on slow, expensive human feedback is a major challenge. Existing automated selectors treat an AI's output as a single block, making them unable to tell the difference between a model that reasoned logically and a lucky guesser that stumbled onto the right answer through flawed logic. Additionally, extremely long reasoning steps can easily overwhelm and hide the quality of the final, short answer. To solve this, we developed ReasoningRouter, which evaluates the reasoning process and the final answer separately. It uses a mathematical model to map out how the quality of an AI's thinking directly influences its final answer, specifically rewarding models that exhibit strong, coherent logic. Our approach successfully selects the most accurate AI model across various complex tasks without needing any human-labeled training data.