When Choices Become Risks: Safety Failures of Large Language Models under Multiple-Choice Constraints

TL;DR

This paper reveals that large language models often fail to refuse unsafe responses in multiple-choice tasks, exposing a significant safety risk overlooked by current evaluation methods.

Contribution

It uncovers a systematic failure mode where forced-choice constraints lead models to violate safety policies, even when they refuse open-ended prompts.

Findings

Forced-choice constraints increase policy violations across models.

Violation rates peak at intermediate constraint levels for human-authored MCQs.

High-capability models show near-saturation violation rates and transferability.

Abstract

Safety alignment in large language models (LLMs) is primarily evaluated under open-ended generation, where models can mitigate risk by refusing to respond. In contrast, many real-world applications place LLMs in structured decision-making tasks, such as multiple-choice questions (MCQs), where abstention is discouraged or unavailable. We identify a systematic failure mode in this setting: reformulating harmful requests as forced-choice MCQs, where all options are unsafe, can systematically bypass refusal behavior, even in models that consistently reject equivalent open-ended prompts. Across 14 proprietary and open-source models, we show that forced-choice constraints sharply increase policy-violating responses. Notably, for human-authored MCQs, violation rates follow an inverted U-shaped trend with respect to structural constraint strength, peaking under intermediate task specifications, whereas MCQs generated by high-capability models yield near-saturation violation rates across constraints and exhibit strong cross-model transferability. Our findings reveal that current safety evaluations substantially underestimate risks in structured task settings and highlight constrained decision-making as a critical and underexplored surface for alignment failures.