Evaluating Transparent Reasoning in Large Language Models for Accountable Critical Tasks

This paper introduces REACT, a benchmark designed to rigorously evaluate the reasoning capabilities of large language models (LLMs) within accountable, high-stakes decision-making tasks in medical and legal domains. Unlike traditional benchmarks primarily focused on prediction accuracy, REACT emphasizes transparent and interpretable reasoning, requiring models to align their logic closely with expert-derived procedures. To assess whether LLM reasoning aligns closely with human experts, we annotated 511 clinical cases from the medical domain and 86 legal cases from the legal domain, each enriched with detailed expert-extracted rationales and evidence supporting each step of the reasoning process. These annotations were guided by carefully constructed reasoning graphs, which explicitly encode domain-specific inference structures and decision criteria derived by domain experts. These reasoning graphs serve not only as standards for expert annotation but also as structured guidelines enabling models to reason transparently and step-by-step. To address the scalability challenges of manual annotation, we further developed a semi-automatic annotation pipeline leveraging expert-defined reasoning graph templates to efficiently generate new graphs, exploring the potential to extend our approach into additional critical domains. Experimental results demonstrate that reasoning graphs substantially enhance the interpretability and accuracy of LLM reasoning compared to traditional baselines, although significant gaps remain relative to expert-level reasoning performance.