Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Zhang, Wenbo, Zhang, Lijinghua, Xiang, Liner, Cai, Hengrui
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866918494710267904
author Zhang, Wenbo
Zhang, Lijinghua
Xiang, Liner
Cai, Hengrui
author_facet Zhang, Wenbo
Zhang, Lijinghua
Xiang, Liner
Cai, Hengrui
contents Reasoning-capable large language models (LLMs) have recently been adopted as automated judges, but their benefits and costs in LLM-as-a-Judge settings remain unclear. Through controlled comparisons between reasoning and non-reasoning judges, we show that explicit reasoning substantially improves judgment accuracy on tasks requiring structured verification (e.g., math and coding), while offering limited or even negative gains on simpler evaluations and incurring significantly higher computational cost. These findings motivate that reasoning should be used selectively rather than universally, with awareness of possible distribution shift. We propose a Robust Adaptive Cost-Efficient Routing (RACER), which dynamically selects between reasoning and non-reasoning judges under a fixed budget by formulating routing as a constrained distributionally robust optimization problem. RACER explicitly accounts for distribution shift via a KL-divergence uncertainty set, admits an efficient primal--dual algorithm, and enjoys theoretical guarantees including uniqueness of the optimal policy and linear convergence. Extensive experiments show that RACER achieves superior accuracy--cost trade-offs under distribution shift.
format Preprint
id arxiv_https___arxiv_org_abs_2605_10805
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge
Zhang, Wenbo
Zhang, Lijinghua
Xiang, Liner
Cai, Hengrui
Artificial Intelligence
Computation and Language
Machine Learning
Reasoning-capable large language models (LLMs) have recently been adopted as automated judges, but their benefits and costs in LLM-as-a-Judge settings remain unclear. Through controlled comparisons between reasoning and non-reasoning judges, we show that explicit reasoning substantially improves judgment accuracy on tasks requiring structured verification (e.g., math and coding), while offering limited or even negative gains on simpler evaluations and incurring significantly higher computational cost. These findings motivate that reasoning should be used selectively rather than universally, with awareness of possible distribution shift. We propose a Robust Adaptive Cost-Efficient Routing (RACER), which dynamically selects between reasoning and non-reasoning judges under a fixed budget by formulating routing as a constrained distributionally robust optimization problem. RACER explicitly accounts for distribution shift via a KL-divergence uncertainty set, admits an efficient primal--dual algorithm, and enjoys theoretical guarantees including uniqueness of the optimal policy and linear convergence. Extensive experiments show that RACER achieves superior accuracy--cost trade-offs under distribution shift.
title Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge
topic Artificial Intelligence
Computation and Language
Machine Learning
url https://arxiv.org/abs/2605.10805