EntroRouter: Learning Efficient Model Routing via Entropy Regulation

A new single-round routing framework called EntroRouter retains 98.3% of the strongest expert model’s accuracy while cutting computational costs by 48.25%, according to a preprint posted to arXiv on June 28, 2026 [1][2]. The paper, titled “EntroRouter: Learning Efficient Model Routing via Entropy Regulation,” addresses a structural failure mode the authors term Trust Region Collapse [2]. In multi-round routing frameworks that interleave reasoning and planning, strong pre-training priors under sparse supervision can cause capable expert models to be systematically suppressed, leading to degenerate local optima [2]. EntroRouter decouples reasoning from routing by treating entropy regulation as a core objective [1][2]. The framework initializes its policy through Soft Supervision, fitting a distribution of suitable models to establish a high-entropy prior for exploration, then stabilizes reinforcement learning with a Soft Anchor that uses offline capability estimates to control entropy contraction within a safe trust region [2]. arXiv, where the preprint appeared, is an open-access repository of electronic preprints that are moderated but not peer-reviewed [6]. Founded in 1991, the repository surpassed two million articles by the end of 2021 and currently receives about 24,000 submissions per month [6]. The EntroRouter paper was posted in the Computation and Language category [1]. The preprint’s abstract page also links to community-built tools through arXivLabs, a framework launched in 2020 that allows third-party collaborators to develop experimental features on the site [5]. These tools, which include bibliographic explorers and code finders, operate under guidelines that require partners to share arXiv’s values of openness, community, excellence, and user data privacy [5]. arXivLabs collaborators receive only minimal, anonymized user data and are prohibited from using it for any purpose beyond the correct functioning of their features [5]. The EntroRouter results arrive as the machine-learning community continues to scrutinize preprints for reproducibility. In 2023, a widely circulated arXiv preprint claiming room-temperature superconductivity in a material called LK-99 was later found to be unsupported after replication attempts identified non-superconducting ferromagnetic and diamagnetic causes for the initial observations [8]. The EntroRouter paper has not yet undergone peer review [6].