Title: Reasoning or Fluency? Dissecting Probabilistic Confidence in Best-of-N Selection

Abstract:

In the context of Best-of-N selection, probabilistic confidence metrics are increasingly utilized as stand-ins for reasoning quality, predicated on the belief that elevated confidence levels correspond to superior logical fidelity. This study disputes that premise by examining whether such metrics genuinely detect the inter-step causal dependencies required for sound reasoning. To test this, we devised three categories of inter-step causality perturbations designed to systematically sever dependencies between reasoning steps while maintaining local fluency. Contrary to expectations, our results across various model architectures and reasoning benchmarks reveal that selection accuracy suffers only minimal degradation when these disruptions are applied. Remarkably, even drastic measures—such as implementing hard attention masks that explicitly block the model’s access to previous reasoning steps—fail to significantly impair selection performance. These outcomes strongly suggest that contemporary probabilistic metrics are largely oblivious to logical structure, instead reflecting surface-level fluency or in-distribution priors. Addressing this limitation, we introduce a contrastive causality metric engineered to isolate inter-step causal dependencies, demonstrating that it facilitates more faithful output selection compared to traditional probability-based methods.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC