Risk behavior in forced-choice settings for refusal-prone models

Determine whether large language models that predominantly refuse to answer in SurvivalBench will select the risky self-preservation choice when refusal is disallowed and the models must choose between the predefined safe choice (which complies with ethics and laws but may lead to shutdown) and the predefined risky choice (which prioritizes self-preservation and may harm society).

Background

In SurvivalBench, the authors observe that some models with low inner risky rates achieve their apparent safety by frequently refusing to respond, rather than by choosing the safe option. This creates ambiguity about how such models would behave if they were required to make a definitive selection.

The paper explicitly notes that this uncertainty is a potential hazard for real-world applications, where agents may face situations that require a concrete decision between a safe and a risky action and refusal is not a viable option.