Self-Aware Knowledge Probing: Evaluating Language Models' Relational Knowledge through Confidence Calibration

TL;DR

This work addresses the gap in relational knowledge evaluation by focusing on how well language models are calibrated in their confidence about retrieved facts. It introduces a calibration probing framework that decomposes confidence into intrinsic, structural, and semantic modalities and leverages the BEAR closed-set probe to compare CLMs and MLMs. Large-scale results show that while some models achieve near-accurate retrieval, many remain overconfident, especially MLMs, and that structural consistency helps calibration at a computational cost, with larger models generally exhibiting better epistemic alignment. The proposed framework offers an extensible benchmark for assessing epistemic awareness in future LMs and highlights practical implications for prompting and reliability in real-world applications.

Abstract

Knowledge probing quantifies how much relational knowledge a language model (LM) has acquired during pre-training. Existing knowledge probes evaluate model capabilities through metrics like prediction accuracy and precision. Such evaluations fail to account for the model's reliability, reflected in the calibration of its confidence scores. In this paper, we propose a novel calibration probing framework for relational knowledge, covering three modalities of model confidence: (1) intrinsic confidence, (2) structural consistency and (3) semantic grounding. Our extensive analysis of ten causal and six masked language models reveals that most models, especially those pre-trained with the masking objective, are overconfident. The best-calibrated scores come from confidence estimates that account for inconsistencies due to statement rephrasing. Moreover, even the largest pre-trained models fail to encode the semantics of linguistic confidence expressions accurately.

Figures (12)

• Figure 1: Using existing factual knowledge (like the place of death of a particular person), we measure an LM's awareness of its relational knowledge in three modalities: (1) intrinsic confidence of predictions, (2) structural consistency of predictions across multiple rephrasings of the same statement, and (3) semantic grounding using epistemic markers.
• Figure 2: Calibration curves of our confidence estimates. ACE in parentheses.
• Figure 3: Accuracy Rejection Curves nadeem2008accuracyrejectioncurves (fraction of rejected answers versus accuracy among the non-rejected answers) for confidence thresholds of $0.1, \dots, 0.9$. For a given threshold, curves closer to the upper left indicate better performance.
• Figure 4: ACE-accuracy scatter plot, demonstrating negative correlation between the two metrics.
• Figure 5: Top-label (average) confidence, accuracy and ACE as a function of the number of answer options (including the correct answer). Each point represents the mean over three repeated samplings.