Salvato in:
| Autori principali: | , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | https://arxiv.org/abs/2604.16916 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866918453189804032 |
|---|---|
| author | Chen, Yuheng Wu, Zhiyu Cheng, Bowen Takahashi, Tetsuro |
| author_facet | Chen, Yuheng Wu, Zhiyu Cheng, Bowen Takahashi, Tetsuro |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_16916 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | When Choices Become Risks: Safety Failures of Large Language Models under Multiple-Choice Constraints Chen, Yuheng Wu, Zhiyu Cheng, Bowen Takahashi, Tetsuro Computation and Language 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. |
| title | When Choices Become Risks: Safety Failures of Large Language Models under Multiple-Choice Constraints |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2604.16916 |