Beyond "I Don't Know": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty

TL;DR

This paper introduces UA-Bench, a benchmark for evaluating LLMs' ability to distinguish between data uncertainty and model uncertainty, revealing current models' limitations and proposing a method to improve uncertainty attribution.

Contribution

The paper presents UA-Bench for explicit uncertainty attribution evaluation and a reinforcement learning strategy to enhance LLMs' uncertainty discrimination capabilities.

Findings

State-of-the-art LLMs struggle to differentiate data and model uncertainty.

High accuracy does not guarantee strong uncertainty attribution.

The proposed method improves uncertainty attribution without sacrificing accuracy.

Abstract

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.