Conditional PAC-Efficient Routing in Large Language Models

A recent study on arXiv examines distribution-free risk management for routing models in large language systems, emphasizing the balance between computational efficiency and dependability. The researchers define pointwise conditional efficiency with a probably approximately correct (PAC) assurance, revealing that it necessitates an almost unattainable router: for nearly every input where the rapid model surpasses the target loss, the algorithm must assign at least one minus the specified error level probability to the expert. To tackle this issue, they propose a limited conditional framework based on a predetermined set of conditioning families, accompanied by a specific router. This router ensures finite-sample conditional validity and, under certain conditions, approaches optimal expert usage. The key takeaway is that the conditioning degree influences the coexistence of distribution-free reliability with computational efficiency: while pointwise control is excessive, structured setwise control is viable. The paper falls under Statistics and Machine Learning and was submitted on December 3, 2025 (arXiv:2512.03057).