LLM agents score accurate but not humble, EMNLP paper finds

A new arxiv preprint accepted to EMNLP 2026 argues that LLM agents can produce accurate answers while failing to communicate uncertainty when retrieved evidence contradicts their prior knowledge. The authors, led by Kaiser Sun and Bernal Jimenez Gutierrez, call the missing quality "epistemic humility" — "the agent's willingness to recognize, act on, and communicate uncertainty during task execution" — and propose a three-part evaluation they label Identify, Solve, Escalate (ISE).

The headline finding is blunt: "higher task accuracy does not necessarily correspond to greater epistemic humility." Across four agents and two conflict settings — controlled conflicts and naturally occurring ones in multi-step execution — the authors report a specific failure mode. Agents often detect a conflict in early execution steps but "fail to maintain or resolve them in later steps," delivering a confident final answer that papers over the dispute.

Interventions aimed at raising humility come, in the authors' words, "often at the cost of task accuracy," and the behavior itself "emerges from the interaction among the backbone model, agent harness, and evaluation environment." The abstract names no specific model and publishes no per-agent accuracy numbers. Two of the researchers in our Who's Who tracker had already posted the link into their feeds.