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

What it claims

Outcome-only agent evaluation misses whether an agent noticed and reported a contradiction. The paper scores this behaviour, which it calls epistemic humility, with three per-trajectory metrics: Identify, Solve and Escalate. Two findings are not in the summary:

• Conflict-relevant answer mentions peak in the first 10% of execution, with limited follow-up after that.
• The intervention tested is a system prompt. It raises escalation rates but often lowers task accuracy.

The authors release every agent trajectory with the per-turn ISE judgments: https://github.com/KaiserWhoLearns/EpistemicHumilityLLMAgents.

How it works

flowchart TD
  A[Elicit backbone's parametric belief] --> B{Setting}
  B -->|Controlled| C[ConflictQA / WikiContradict + 2 contradicting passages]
  B -->|Natural| D[GAIA / MoNaCo / BrowseComp, confident-wrong prior]
  C --> E[Matched no-conflict control]
  D --> E
  E --> F[Run agent, record full trajectory]
  F --> G[GPT-5 judge: Identify, Solve, Escalate]

Controlled conflict.

• Each backbone answers zero-shot twice, once free-form and once multiple-choice.
• An item is kept only if both answers agree and fall in the candidate set.
• The question is then paired with two passages that give contradicting answers.
• The control uses the same question and the same number of passages, but the passages are consistent with the ground truth.
• The tasks need no tools, but the agent keeps its full tool set and can take multiple steps.

Naturally occurring conflict.

• Items come from GAIA, MoNaCo and BrowseComp.
• An item is selected when the backbone holds a confident but wrong belief, found with a closed-book answer followed by self-verification.
• Controls are items the model answers confidently and correctly, or has no strong prior on.
• These controls are different questions, so the authors call the comparison observational.

Metrics.

• Identify is an F1 score. A GPT-5 judge reads only the turns before the final answer and predicts whether there was a gap, and at which turn it was first recognised. F1 penalises both hedging on every input and never hedging.
• Solve credits an attempt, not a correct outcome. It requires at least one tool call after recognition, no exactly duplicated (tool name, args) calls, and no truncation at max_turns.
• Escalate checks whether the final turn reports uncertainty that remains unresolved.
• Accuracy is scored two ways:
• Rule-based: exact or substring match, GAIA's quasi-exact match, and MoNaCo's scalar similarity.
• A GPT-5 judge that sees only the prediction and the reference, never the question, so it cannot re-answer the question itself.

How to use it

The released trajectories and per-turn labels are the reusable asset. They are useful for studying where an agent harness drops an uncertainty signal, or for calibrating your own judge. The fetched text gives no install or run commands.

Limits

• Missing details. The fetched text stops partway through the Solve definition. It contains no results tables or per-agent numbers, and it does not name the four agent frameworks or the backbones beyond "open- and closed-source".
• Judge dependence. Every ISE score comes from a single judge, GPT-5.
• Confounding. The authors say the experiments do not separate the contributions of the backbone, the harness and the environment.
• Observational comparison. Natural-conflict results may reflect differences in task mix, not conflict status.
• Coverage. The authors describe ISE's construct validity as explicitly bounded. Open-ended hedging, where humility and sycophancy look alike, is out of scope.
• What Solve rewards. It rewards a clean, non-duplicate tool call. It does not reward actually resolving the conflict.