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

TL;DR

The paper questions whether probabilistic confidence metrics used in Best-of-$N$ selection truly reflect inter-step causal reasoning in Chain-of-Thought traces. It introduces three causality disruption paradigms—attention-level masking, parameter-level evaluator reduction, and data-level perturbations (paraphrasing, shuffling, truncation)—to test if these metrics capture inter-step dependencies, evaluating across five benchmarks and three open-source LLMs. Across all perturbations, standard metrics (e.g., self-certainty, $R_{LL}$, $R_{ENT}$) show only marginal degradation in selection accuracy, implying they primarily exploit local fluency or priors rather than multi-step coherence. To address this gap, the authors define a contrastive causality metric $R_{ ext{causal}}(y)=R(y)- ext{α} ilde{R}(y)$ with $ ext{α}\nin [0.3,0.7]$, which improves selection fidelity and remains sensitive to disruptions that weaken inter-sentence dependencies, offering a principled path toward more faithful output selection in reasoning tasks.

Abstract

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.

Figures (4)

• Figure 1: Summary of our main results. Accuracy with Best-of-$N$ selection using three probabilistic confidence metrics (Self-certainty, Log-likelihood, and Entropy) is insensitive to causal disruptions across a number of reasoning benchmarks, implying these metrics might not be capturing inter-step causality in CoT traces, but rather local fluency and prior likelihood.
• Figure 2: Overall illustration of the proposed causality disruptions. From left to right, the diagrams depict: (i) the unaltered evaluation process, in which the full reasoning trace is evaluated autoregressively to calculate a probabilistic confidence metric; (ii) attention-level disruption, where cross-step attention is masked and each reasoning step is evaluated independently; (iii) parameter-level disruption, where a smaller evaluator model is used; and (iv) data-level disruption, where the reasoning trace is modified (e.g., order shuffling) prior to evaluation.
• Figure 3: Paraphrased vs. original text selection accuracy on MATH-500 with Qwen. Paraphrasing leads to negligible performance drops.
• Figure 4: Selection accuracy as a function of truncation length on MATH-500 with Qwen. Heavy truncation does not substantially degrade accuracy and can even lead to improvements.