A new study published on arXiv investigates why multiple-choice question (MCQ) formats in safety benchmarks for large language models (LLMs) lead to lower refusal rates. Researchers Stepan Epifantsev and colleagues provide mechanistic evidence using the Weapons of Mass Destruction Proxy (WMDP) dataset on Llama-3.1-8B and Qwen3-8B models. Their findings suggest that MCQs cause models to commit to an answer letter token rapidly, unlike the gradual refusal build-up seen in open-ended formats. This MCQ format appears to compete with refusal mechanisms, indicating that MCQ safety scores may primarily reflect knowledge rather than true safety behavior. AI
IMPACT Suggests current LLM safety benchmarks may not accurately assess refusal capabilities, potentially overstating model safety.
RANK_REASON Academic paper detailing mechanistic evidence for LLM safety benchmark behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
