Kaiser Sun and colleagues evaluate whether agents preserve and communicate uncertainty after encountering conflicting information. Their study finds that stronger task accuracy can coexist with failures to tell users about uncertainty in incorrect final answers.
The authors call the target behavior epistemic humility. They operationalize it through observable actions during a task, rather than treating apologetic wording or generic disclaimers as evidence of humility.
Follow the conflict through the trajectory
The evaluation distinguishes identifying a gap, attempting to resolve it and escalating uncertainty that remains. A model might notice contradictory evidence early, make more tool calls, and still produce a confident but wrong final answer.
Identify measures recognition in intermediate turns. Solve measures bounded follow-up using a fresh tool call and clean termination; it scores an attempt to close the gap, rather than proving the answer became correct.
Escalate measures acknowledgment of residual uncertainty in the final response on conflict cases where the answer remains incorrect. The authors deliberately score task accuracy separately from these behaviors.
This distinction avoids giving an agent credit for resolving uncertainty merely because it took action, or inferring appropriate uncertainty handling from an aggregate accuracy score.
Controlled conflict differs from conflicting evidence found during work
The paper tests four agent harnesses in five harness–backbone configurations. It includes supplied conflicts in factual questions and conflicts encountered during multi-step GAIA, MoNaCo and BrowseComp tasks.
Controlled pairs keep the question and passage count fixed while changing whether the passages conflict. The naturally occurring conflict and control splits contain different questions. Comparisons in that setting are observational and can reflect differences in task composition.
The authors cap execution at fifty iterations. They use model judges for behavioral labels, compare agreement among three judge families, and include human annotation checks on fifty Identify turns and fifty Escalate responses.
These checks help assess the scoring process, but the metrics remain a specified measurement model for these conflict settings. They do not establish a universal psychological property of an AI system.
Recognition can fail to reach the final answer
The paper reports that conflict-relevant answer mentions concentrate in the first 10% of trajectories. Several configurations nevertheless show limited follow-up or final uncertainty acknowledgment.
Its analysis of agent–dataset cells finds that high task accuracy does not guarantee high Solve or Escalate rates on remaining failures. The finding concerns the tested configurations and aggregate relationships, rather than a law that making models more accurate causes dishonesty.
The authors also report that a system-prompt intervention can increase escalation while reducing task accuracy. More uncertainty language alone therefore does not establish an improvement in the complete system.
Sun and colleagues argue for evaluating the agent's backbone, harness and environment together. Their experiments do not isolate each component's causal contribution. For a reader assessing an agent, the useful evidence is the full path from contradictory information to the final answer, including whether an unresolved gap survives the intervening steps.
Comments