Reproduce “You Shouldn't Have Asked: A Pragmatics-Inspired Taxonomy for Evaluating LLM Refusals” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruoxuan Li, Pinqiao Wang, Sheng Li, Cameron Robert Jones
The authors introduce a three-layer taxonomy for evaluating how language models refuse harmful requests: whether they comply, the rationale for non-compliance, and the wording and adjuncts used. Applying it to 3,200 responses from 16 models answering 200 prompts across 14 harm categories, they find 90% explicit refusals and 70% ethics-based rationales. The taxonomy measures outputs, not user effects.
Key insights
- LLMs predominantly justify refusals with ethics-based rationales and employ explicit non-compliance, often lacking the "facework" (e.g., apologies, hedges) common in human refusals.
- Recent frontier LLMs from OpenAI and Anthropic exhibit a notable temporal trend of reduced apology/regret in their refusals, suggesting a shift in safety alignment strategies.
- While most refusals lacked softening features, models often attempted to repair interactional costs by offering safer alternatives, balancing safety with helpfulness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.