PACE: Towards Surfacing Hidden Conflicts in User Requests
The paper tests whether assistants can spot user-specific conflicts that are only visible after retrieving hidden context.
PACE is a dataset built around personas, user requests, and egocentric knowledge-base facts that can make an otherwise reasonable request inappropriate. The authors also propose PaceMaker, a multi-agent retrieval framework for finding the conflict-relevant evidence through query reformulation, graph traversal, and filtering. In their experiments, PaceMaker beats existing approaches on both evidence retrieval and conflict-decision accuracy. HF Daily Papers' note
PACE is a dataset built around personas, user requests, and egocentric knowledge-base facts that can make an otherwise reasonable request inappropriate. The authors also propose PaceMaker, a multi-agent retrieval framework for finding the conflict-relevant evidence through query reformulation, graph traversal, and filtering. In their experiments, PaceMaker beats existing approaches on both evidence retrieval and conflict-decision accuracy. HF Daily Papers' note
score 4