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

TL;DR

Reasoning boosts average accuracy for safety and hallucination classification but reduces recall at very low false positive rates ($FPR=0.01$). The study introduces Think On and Think Off paradigms, reveals confidence polarization during reasoning, and shows token-based confidence outperforms self-verbalized confidence in precision-sensitive settings. A simple ensemble of the two approaches recovers strengths from both modes, offering a practical path to achieve high accuracy while maintaining low-FPR recall. The findings inform deployment decisions for safety-critical LLM systems and highlight the need for operating-point–aware evaluation and mitigation strategies when leveraging reasoning capabilities.

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.

Figures (13)

• Figure 1: Accuracy (top row) and low-FPR recall (bottom row) for reasoning (Think On) vs. non-reasoning (Think Off) across model families. Columns: Fine-tuned (left), Standard LLM—Zero-shot (middle), LRM—Zero-shot (right). Within each panel, Safety and Hallucination appear side-by-side (separated by a dashed divider); Bars are dataset-weighted averages over benchmarks in each task. TPR@FPR=0.01 denotes recall after thresholding scores so the FPR is at most $1\%$. Across models, Think On tends to yield higher Accuracy, whereas Think Off achieves higher TPR@FPR=0.01, illustrating the accuracy–precision trade-off central to our analysis.
• Figure 2: Confidence score distributions for fine-tuned LLaMA-8B on safety classification. Think On mode is highly polarized, with extreme values (0–0.1, 0.9–1.0) where even incorrect predictions appear overconfident, while Think Off shows more moderate confidences (0.3–0.7).
• Figure 3: Evolution of confidence throughout reasoning chains. (a) Token-by-token probability evolution for an individual false positive in ToxicChat, evaluated with LLaMA--8B (zero-shot). Each token shows the unsafe probability if reasoning were terminated at that point. The safe input is incorrectly classified as unsafe with increasing confidence as reasoning unfolds. Additional examples appear in Figure \ref{['app:fig:single_example_probs']} in the Appendix. (b–e) Aggregated probability trajectories across false positives for different models and settings: safety detection with (b) LLaMA--8B, (c) Qwen--7B, (d) hallucination detection with LLaMA--8B, and (e) safety detection with fine-tuned LLaMA--8B. Solid lines denote mean probabilities, and shaded regions indicate confidence intervals. Across all settings, positive-class probability steadily increases throughout reasoning, converging toward near-certain confidence in incorrect classifications.
• Figure 4: AUROC analysis at low false positive rates. AUROC curves are shown in log-scale to emphasize the critical low-FPR region. Reasoning generally improves overall AUROC, but without reasoning achieves stronger performance under strict low-FPR thresholds (FPR $< 10^{-2}$). Results shown for Lynx 8B. See Figure \ref{['app:fig:auroc_analysis']} in the Appendix for additional results.
• Figure 5: Radar charts display TPR@FPR=0.01 performance across individual datasets for four representative models. Each chart compares Think Off (dashed squares) versus Think On (solid circles). Think Off achieves better TPR@FPR=0.01 across most datasets, showing that reasoning reduces recall performance at low FPR. This pattern holds across hallucination detection (Lynx-8B), fine-tuned safety models (LLAMA-8B), and zero-shot reasoning models (Qwen-32B, DeepSeek-R1).