New paper proposes evaluating "epistemic humility" in AI agents

A paper by Sun et al. (arXiv 2610.12360v1) introduces epistemic humility as an eval dimension for AI agents: whether an agent detects conflicts between retrieved evidence and its prior beliefs, revises its answer, and flags uncertainty. Tests show high task accuracy does not guarantee such behavior.

• Three dimensions proposed: identify the conflict, revise the answer, escalate uncertainty
• Agents were tested under injected and naturally occurring source conflicts
• Key failure mode: agents detect a conflict mid-trajectory but output a confident final answer
• Authors recommend trajectory logging and explicit uncertainty flags in outputs