Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict

When retrieved evidence contradicts an agent's prior beliefs, does it revise its answer, acknowledge uncertainty, or persist with an incorrect conclusion? Existing evaluations of agentic systems focus primarily on task success, offering limited insight into how agents handle such conflicts. We propo