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

Accurate but Not Humble. Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict. By Kaiser Sun and six others, from Johns Hopkins University and New York University. Published on the eighth of October.

Picture an agent that retrieves evidence contradicting what it already believes. It could revise its answer, admit it is uncertain, or carry on with a wrong conclusion. Evaluations of agentic systems mostly measure task success, which says little about which of those it does. The paper's central claim is that the two come apart: higher task accuracy does not necessarily mean greater epistemic humility.

The authors define epistemic humility as an agent's willingness to recognise, act on, and communicate uncertainty while it carries out a task. They make it measurable through three behavioural dimensions read off the agent's whole trajectory: Identify, Solve, and Escalate, or ISE for short.

Conflict is the instrument. Knowledge conflict here means either the backbone language model's parametric knowledge contradicting the evidence it meets, or two contextual sources disagreeing with each other. Two settings are tested. One is controlled conflict. The other is conflict that arises naturally during multi-step agentic execution. Each is paired with a matched control in which there is no conflict at all.

Across four agents, the headline pattern appears. Some high-accuracy configurations do recognise a conflict during execution, yet when their final answer is wrong, they do not communicate the uncertainty that was never resolved. Looking along the trajectories adds a second pattern: agents frequently detect a conflict in the early steps, then fail to maintain or resolve it in later ones. Noticing happens. Following through often does not.

The evidence comes from four agents and two kinds of conflict, so the claim is that accuracy and humility can diverge, not that they always do. The authors also test model-level interventions. These can improve epistemic humility, but often at a cost in task accuracy. From that trade-off they draw a wider reading: humility is not a property of the model alone. It emerges from the interaction among the backbone model, the agent harness, and the evaluation environment, which means any measurement of it is tied to all three.

An accuracy score treats a confident wrong answer and a hedged wrong answer the same. This study is about the difference between them. An agent that saw the contradiction early and then said nothing about it in the end scores no worse on success, and that silence is exactly what a humility measure is built to catch.