Mechanisms by which proprietary reasoning budgets reduce incoherence

Investigate and elucidate the mechanisms through which proprietary "reasoning budgets" in frontier large language models reduce incoherence in model outputs, given that their implementation details are undisclosed.

Background

The authors empirically find that increasing reasoning budgets can slightly reduce incoherence, though this effect is overshadowed by natural variation in reasoning length. However, because the implementation details of these budgets are not public, the causal mechanisms remain unknown.

They hypothesize that improvements may arise from enhanced backtracking and error-correction, and relate this to ensembling—which demonstrably reduces variance—but emphasize the need for a concrete mechanistic account.