$ cat wiki/papers/2026/2610.12360-epistemic-humility.md
Epistemic humility — accuracy does not ensure uncertainty disclosure
TL;DR
Four evaluated agents can detect conflicting evidence yet fail to preserve or communicate uncertainty in the final answer. Better task accuracy does not necessarily imply better epistemic humility. (source)
Authors & Org
Kaiser Sun, Bernal Jimenez Gutierrez, Hongjun Liu, Jingyu Zhang, Jie Gao, Mark Dredze and Daniel Khashabi. Submitted 2026-10-08; the page labels the paper EMNLP 2026 Camera Ready. Affiliations are not stated on the captured page. (source)
Method
The authors evaluate whether agents identify, solve and escalate uncertainty using the Identify, Solve, and Escalate (ISE) framework. They pair controlled conflicts and naturally occurring conflicts during multi-step execution with matched no-conflict controls, including contradictions between parametric knowledge and evidence, or between contextual sources. (source)
Results
In four agents, some high-accuracy configurations recognize conflicts during execution but omit unresolved uncertainty from incorrect final answers. Conflicts detected early can be lost later. Model-level interventions can improve epistemic humility while reducing task accuracy; the abstract supplies no numerical effect size or per-agent table. (source)
Significance
For AI Alignment and Agents (LLM Agents), this separates internal recognition from the account delivered to a user. Interpretation: a correct-answer metric alone cannot establish that unresolved evidence conflicts will be disclosed. (source)
Open Questions
Which interventions retain accuracy while improving disclosure, and how the tradeoff changes with the harness, remain open in this abstract-level review. No claim about a named frontier model’s score is supported here. (source)