PACE tests whether an assistant will refuse a reasonable-sounding request because of something it has to retrieve about you first

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2609.03293 (2026-09-03) points out that conflict and safety detection work mostly assumes the conflicting factor is stated in the prompt, while real cases hide it in a knowledge base. PACE pairs persona-grounded user requests with egocentric KB facts and events that make an otherwise sensible request inappropriate, so the model has to retrieve the constraint before it can decide to refuse. The implicit retrieval setting breaks the direct association between request and conflicting knowledge, which is exactly where current models fail. They also propose PaceMaker, a multi-agent framework with specialized agents for the retrieval and judgment steps.