Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict
The paper separates getting the answer right from knowing when the agent should admit conflict or uncertainty.
The authors test “epistemic humility” through whether agents identify conflict, solve it, and escalate unresolved uncertainty. They use cases where retrieved evidence clashes with the model’s prior knowledge or where contextual sources disagree. Across four agents, higher accuracy does not reliably mean more humility: some agents notice conflicts but still give wrong final answers without flagging uncertainty. The paper also finds interventions can improve humility, often with an accuracy tradeoff. HF Daily Papers' note
The authors test “epistemic humility” through whether agents identify conflict, solve it, and escalate unresolved uncertainty. They use cases where retrieved evidence clashes with the model’s prior knowledge or where contextual sources disagree. Across four agents, higher accuracy does not reliably mean more humility: some agents notice conflicts but still give wrong final answers without flagging uncertainty. The paper also finds interventions can improve humility, often with an accuracy tradeoff. HF Daily Papers' note
score 3