Beyond "I Don't Know": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty
For developers and users of LLMs, this work provides a benchmark and method to improve models' ability to attribute uncertainty, enabling better downstream decisions like requesting clarification or invoking tools.
LLMs often refuse to answer with a generic 'I don't know', failing to distinguish data uncertainty from model uncertainty. The authors introduce UA-Bench, a benchmark of 3,500+ questions, and find that even top LLMs struggle to discriminate these uncertainties; they propose a reinforcement learning method that improves uncertainty attribution on Qwen3-4B and Qwen3-8B while maintaining accuracy.
Reliable Large Language Models (LLMs) should abstain when confidence is insufficient. However, prior studies often treat refusal as a generic "I don't know'', failing to distinguish input-level ambiguity (data uncertainty) from capability limitations (model uncertainty). This lack of distinction limits downstream action decisions like requesting clarification or invoking external tools. In this work, we introduce UA-Bench, a benchmark of over 3,500 questions drawn from six datasets spanning knowledge-intensive and reasoning-intensive tasks, designed to evaluate explicit uncertainty attribution. An evaluation of 18 frontier LLMs shows that even state-of-the-art models struggle to reliably discriminate between data uncertainty and model uncertainty, and that high answer accuracy does not necessarily imply strong uncertainty attribution ability. To narrow this gap, we propose a lightweight data synthesis and reinforcement learning strategy. Experiments on both Qwen3-4B-Instruct-2507 and Qwen3-8B in thinking mode show that the proposed method improves uncertainty attribution while preserving answer accuracy. Our code and data are publicly available now.