Reasoning's Razor: Reasoning Improves Accuracy but Can Hurt Recall at Critical Operating Points in Safety and Hallucination Detection

TL;DR

This study systematically evaluates how reasoning in large language models affects accuracy and recall in safety and hallucination detection tasks, revealing a trade-off where reasoning improves overall accuracy but can impair low-FPR performance.

Contribution

It provides the first comprehensive analysis of reasoning's impact on classification tasks under strict low-FPR conditions, highlighting when reasoning helps or hurts in safety-critical applications.

Findings

Reasoning improves overall accuracy but can reduce recall at low FPR thresholds.

Token-based scoring outperforms self-verbalized confidence in precision-sensitive settings.

Ensembling reasoning and non-reasoning modes recovers strengths of both approaches.

Abstract

Reasoning has become a central paradigm for large language models (LLMs), consistently boosting accuracy across diverse benchmarks. Yet its suitability for precision-sensitive tasks remains unclear. We present the first systematic study of reasoning for classification tasks under strict low false positive rate (FPR) regimes. Our analysis covers two tasks--safety detection and hallucination detection--evaluated in both fine-tuned and zero-shot settings, using standard LLMs and Large Reasoning Models (LRMs). Our results reveal a clear trade-off: Think On (reasoning-augmented) generation improves overall accuracy, but underperforms at the low-FPR thresholds essential for practical use. In contrast, Think Off (no reasoning during inference) dominates in these precision-sensitive regimes, with Think On surpassing only when higher FPRs are acceptable. In addition, we find token-based scoring substantially outperforms self-verbalized confidence for precision-sensitive deployments. Finally, a simple ensemble of the two modes recovers the strengths of each. Taken together, our findings position reasoning as a double-edged tool: beneficial for average accuracy, but often ill-suited for applications requiring strict precision.