Polite but Misaligned: Evaluating LLM Politeness Judgments Against Human Pragmatic Norms
Bridging the gap between linguistic accuracy and human-centric social pragmatics in language models.
Despite strong performance on standard benchmarks, it remains unclear whether large language models (LLMs) evaluate social pragmatics in ways that align with human judgments. We evaluate LLM politeness judgments using two English-language datasets with complementary annotation fo...
ArXiv CS.CL
Source
Generated insight by OnAirToday AI
As LLMs integrate into customer-facing roles, their ability to navigate complex social nuances like politeness is becoming as critical as their raw processing power. This research exposes a vital gap in current alignment techniques, highlighting that logical competence does not inherently translate to social intelligence.
Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.