Your Model Diversity, Not Method, Determines Reasoning Strategy

TL;DR

This paper argues that the effectiveness of reasoning strategies in large language models depends on the models' diversity profile, emphasizing the importance of model diversity over the specific method used.

Contribution

It introduces a theoretical framework linking model diversity to reasoning strategy effectiveness and validates it across different model families.

Findings

Depth refinement works well for low-diversity models with lightweight signals.

High-diversity models require stronger signals for effective depth-based refinement.

Model diversity profile influences the choice of reasoning exploration strategies.

Abstract

Compute scaling for LLM reasoning requires allocating budget between exploring solution approaches ($breadth$) and refining promising solutions ($depth$). Most methods implicitly trade off one for the other, yet why a given trade-off works remains unclear, and validation on a single model obscures the role of the model itself. We argue that $\textbf{the optimal strategy depends on the model's diversity profile, the spread of probability mass across solution approaches, and that this must be characterized before any exploration strategy is adopted.}$ We formalize this through a theoretical framework decomposing reasoning uncertainty and derive conditions under which tree-style depth refinement outperforms parallel sampling. We validate it on Qwen-3 4B and Olmo-3 7B families, showing that lightweight signals suffice for depth-based refinement on low-diversity aligned models while yielding limited utility for high-diversity base models, which we hypothesize require stronger compensation for lower exploration coverage.