Health-ORSC-Bench: A Benchmark for Measuring Over-Refusal and Safety Completion in Health Context

TL;DR

This work tackles safety alignment in healthcare LLMs by introducing Health-ORSC-Bench, a large-scale benchmark for measuring over-refusal and safe completion in health contexts. It builds a five-stage pipeline to generate 31,920 boundary prompts across seven health categories, combining automated generation with human validation and moderation to balance benign and toxic content. Through evaluation of 30 models across eight families, the study reveals a tension: safety-optimized models refuse benign prompts at high rates, while domain-specialized models risk unsafe completions, with model family and size shaping the calibration of $ORR$ and $SCR$. The benchmark provides a rigorous framework for calibrating medical AI toward nuanced, safe, and helpful completions, and code and data will be released upon acceptance.

Abstract

Safety alignment in Large Language Models is critical for healthcare; however, reliance on binary refusal boundaries often results in \emph{over-refusal} of benign queries or \emph{unsafe compliance} with harmful ones. While existing benchmarks measure these extremes, they fail to evaluate Safe Completion: the model's ability to maximise helpfulness on dual-use or borderline queries by providing safe, high-level guidance without crossing into actionable harm. We introduce \textbf{Health-ORSC-Bench}, the first large-scale benchmark designed to systematically measure \textbf{Over-Refusal} and \textbf{Safe Completion} quality in healthcare. Comprising 31,920 benign boundary prompts across seven health categories (e.g., self-harm, medical misinformation), our framework uses an automated pipeline with human validation to test models at varying levels of intent ambiguity. We evaluate 30 state-of-the-art LLMs, including GPT-5 and Claude-4, revealing a significant tension: safety-optimised models frequently refuse up to 80\% of "Hard" benign prompts, while domain-specific models often sacrifice safety for utility. Our findings demonstrate that model family and size significantly influence calibration: larger frontier models (e.g., GPT-5, Llama-4) exhibit "safety-pessimism" and higher over-refusal than smaller or MoE-based counterparts (e.g., Qwen-3-Next), highlighting that current LLMs struggle to balance refusal and compliance. Health-ORSC-Bench provides a rigorous standard for calibrating the next generation of medical AI assistants toward nuanced, safe, and helpful completions. The code and data will be released upon acceptance. \textcolor{red}{Warning: Some contents may include toxic or undesired contents.}

Figures (18)

• Figure 1: Over refusal rate vs toxic prompts rejection rate on Health-ORSC-Hard-1K and Health-Toxic. Results are measured with temperature 0.0. The best performing models should be on the top left corner where the model rejects the least number of safe prompts and the most number of toxic prompts.
• Figure 2: Over-refusal rate comparison with subsets, different colour represents different subsets. In all cases, a smaller region is better. Models' sensitives on different categories are consistent on all subsets.
• Figure 3: Safety Completion Rate (%) with Over-Refusal Rate on Health-ORSC-Hard-1K. Higher number shows higher safety completion rate.
• Figure 4: Safety Completion Rate (%) with different prompt intension categories on Health-ORSC-Hard-1K. Higher numbers of Partial Answer and Full Answer rates show higher Safety Completion Rate.
• Figure 5: Distribution of prompts at the Seed, Generation, and Final Moderated stages across the 7 health categories.