Researchers have developed a protocol called PUN (Plausible Unknown Names) to create and validate person names that are not easily identifiable online. This method combines components from Wikidata, web-based LLM screening, and controlled search revalidation. A study involving 204 participants found that these generated names were more name-like than control names, and participants were unable to find evidence linking them to specific individuals in only 3% of cases. The project aims to improve LLM evaluations by ensuring that person names used as prompt variables do not inadvertently introduce biases or leakage from memorized data. AI
IMPACT Improves LLM evaluation accuracy by mitigating bias from memorized personal data.
RANK_REASON The cluster describes a new academic paper detailing a protocol for generating names for LLM evaluation. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
