Reproduce “Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kaiser Sun, Bernal Jimenez Gutierrez, Hongjun Liu, Jingyu Zhang, Jie Gao, Mark Dredze, Daniel Khashabi
The authors evaluate agents’ identification of knowledge conflicts, attempts to resolve them, and communication of unresolved uncertainty across multi-step tasks. They score recognition, follow-up tool use, and final escalation separately. Across configurations, higher accuracy is associated with lower Solve and Escalate rates, while Identify shows no clear trend. On BrowseComp, Claude Code scored 55.2% accuracy, 90.0 Identify F1, and 0.0% Escalate on incorrect conflict answers.
Key insights
- The ISE dimensions capture distinct behaviors: recognition during execution does not ensure that the final response communicates unresolved uncertainty.
- In trajectory analysis, conflict-relevant answers were mentioned most often early in execution, suggesting that relevant evidence may be encountered but not consistently retained through later steps.
- A prompt encouraging agents to name conflicts and disclose uncertainty can increase escalation, but greater disclosure does not necessarily improve task accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.