Know Thy Reasoner: Not All Language Models Explore Alike

A new study argues that the best strategy for scaling compute on large language model reasoning tasks depends on a model's internal diversity profile, a finding that challenges one-size-fits-all approaches to inference-time search. The paper, posted to the arXiv preprint repository on April 12, 2026, and revised on June 15, formalizes a framework that decomposes reasoning uncertainty to determine when depth-based refinement of a single solution outperforms generating many parallel samples, a trade-off the authors call breadth versus depth [1][2]. The central claim is that the optimal strategy is dictated by the model's diversity regime: the spread of probability mass across different solution approaches [2]. The researchers validated their framework across three model families at both inference and training stages [1][2]. They found that low-diversity aligned models benefit from depth-based refinement using lightweight intrinsic signals, while high-diversity base models are often harmed by the same technique and instead require breadth or stronger external signals to compensate [2]. The submission history shows the initial manuscript was 57 KB, growing to 174 KB in the revised version [1]. The work was led by Moulik Choraria, according to the paper's metadata [1]. The preprint appeared on arXiv, an open-access repository that hosts electronic preprints across mathematics, physics, computer science, and related fields and has been operating since August 1991 [10]. As of late 2024, the repository receives approximately 24,000 new articles per month [10]. The paper's abstract page includes integration with several community-built discovery tools through the arXivLabs framework, a program launched in 2020 that allows third-party collaborators to develop experimental features directly on the site [9]. These tools include the Bibliographic Explorer for navigating citation trees and the CORE Recommender for surfacing related open-access papers [8][9]. arXivLabs collaborators operate under guidelines that require adherence to arXiv's values of openness, community, excellence, and user data privacy, with access limited to minimal and anonymized user data [9].