Reasoning or Fluency? Dissecting Probabilistic Confidence in Best-of-N Selection
This work addresses a critical problem for AI researchers and practitioners in ensuring reliable reasoning assessment in language models, revealing that current metrics may be misleading, though it is incremental in proposing an alternative.
The study challenged the assumption that probabilistic confidence metrics reflect reasoning quality in Best-of-N selection by showing they are largely insensitive to logical structure, with selection accuracy degrading only marginally under disruptions to inter-step causal dependencies. It proposed a contrastive causality metric that improved output selection fidelity over existing probability-based approaches.
Probabilistic confidence metrics are increasingly adopted as proxies for reasoning quality in Best-of-N selection, under the assumption that higher confidence reflects higher reasoning fidelity. In this work, we challenge this assumption by investigating whether these metrics truly capture inter-step causal dependencies necessary for valid reasoning. We introduce three classes of inter-step causality perturbations that systematically disrupt dependencies between reasoning steps while preserving local fluency. Surprisingly, across diverse model families and reasoning benchmarks, we find that selection accuracy degrades only marginally under these disruptions. Even severe interventions, such as applying hard attention masks that directly prevent the model from attending to prior reasoning steps, do not substantially reduce selection performance. These findings provide strong evidence that current probabilistic metrics are largely insensitive to logical structure, and primarily capture surface-level fluency or in-distribution priors instead. Motivated by this gap, we propose a contrastive causality metric that explicitly isolates inter-step causal dependencies, and demonstrate that it yields more faithful output selection than existing probability-based approaches.