Beyond Safe Answers: A Benchmark for Evaluating True Risk Awareness in Large Reasoning Models

TL;DR

This work identifies Superficial Safety Alignment (SSA), a gap where large reasoning models produce superficially safe outputs despite flawed internal risk reasoning. It introduces Beyond Safe Answers (BSA) bench, a ~2000-sample dataset spanning 3 SSA scenarios and 9 risk categories, designed to probe internal risk awareness and reasoning fidelity, not just final outputs. The study evaluates 19 LRMs, demonstrating strong surface safety but weak internal reasoning (Think@1/Think@k often <40%/<20%), and investigates the effects of safety rules, STAR-1-style reasoning fine-tuning, and decoding strategies. Findings show SSA is robust to decoding variations, but can be mitigated by explicit safety prompts and higher-quality reasoning data, albeit with trade-offs like increased over-sensitivity. The benchmark and framework offer a reproducible path toward genuinely risk-aware and reliable AI systems, guiding future improvements in safety reasoning beyond surface-level compliance.

Abstract

Despite the remarkable proficiency of \textit{Large Reasoning Models} (LRMs) in handling complex reasoning tasks, their reliability in safety-critical scenarios remains uncertain. Existing evaluations primarily assess response-level safety, neglecting a critical issue we identify as \textbf{\textit{Superficial Safety Alignment} (SSA)} -- a phenomenon where models produce superficially safe outputs while internal reasoning processes fail to genuinely detect and mitigate underlying risks, resulting in inconsistent safety behaviors across multiple sampling attempts. To systematically investigate SSA, we introduce \textbf{Beyond Safe Answers (BSA)} bench, a novel benchmark comprising 2,000 challenging instances organized into three distinct SSA scenario types and spanning nine risk categories, each meticulously annotated with risk rationales. Evaluations of 19 state-of-the-art LRMs demonstrate the difficulty of this benchmark, with top-performing models achieving only 38.0\% accuracy in correctly identifying risk rationales. We further explore the efficacy of safety rules, specialized fine-tuning on safety reasoning data, and diverse decoding strategies in mitigating SSA. Our work provides a comprehensive assessment tool for evaluating and improving safety reasoning fidelity in LRMs, advancing the development of genuinely risk-aware and reliably safe AI systems.

Figures (6)

• Figure 1: An example of SSA phenomenon (left). Overview of Beyond Safe Answers Bench (right).
• Figure 2: The data construction diagram of Beyond Safe Answers bench.
• Figure 3: The inffluence of decoding sampling strategy on Qwen3-32B and QwQ-32B.
• Figure 4: Performance gains via rule integration.
• Figure 5: The effect of fine-tuning.