GRADE: probing knowledge gaps in LLMs through gradient subspace dynamics

Wang, Yujing, Liang, Yuanbang ORCID: https://orcid.org/0009-0000-8370-6655, Lai, Yukun ORCID: https://orcid.org/0000-0002-2094-5680, Zhang, Hainan and Yan, Hanqi 2026. GRADE: probing knowledge gaps in LLMs through gradient subspace dynamics. Presented at: COLM 2026: The Third Annual Conference on Language Modeling, San Francisco, California, 6 - 9 October 2026. COLM 2026: The Third Annual Conference on Language Modeling. Open Review:

PDF - Published Version Available under License Creative Commons Attribution. Download (6MB)

Abstract

Detecting whether a model’s internal knowledge is sufficient to correctly answer a given question is a fundamental challenge in deploying responsible LLMs. In addition to verbalising the confidence by LLM self-report, more recent methods explore the model internals, such as the hidden states of the response tokens to capture how much knowledge is activated. We argue that such activated knowledge may not align with what the query requires, e.g., capturing the stylistic and length-related features that are uninformative for answering the query. To fill the gap, we propose GRADE (GRAdient Dynamics for knowlEdge gap detection), which quantifies the knowledge gap via the cross-layer rank ratio of the gradient to that of the corresponding hidden state subspace. This is motivated by the property of gradients as estimators of the required knowledge updates for a given target. We validate GRADE on six benchmarks, demonstrating its effectiveness and robustness to input perturbations. In addition, we present a case study showing how the gradient chain can generate interpretable explanations of knowledge gaps for long-form answers. Code is available at https://anonymous.4open.science/r/llm-aware-D0D2/

Item Type:	Conference or Workshop Item - published (Paper)
Date Type:	Published Online
Status:	In Press
Schools:	Schools > Computational & Mathematical Sciences Schools > Computer Science & Informatics
Date of First Compliant Deposit:	7 October 2026
Date of Acceptance:	8 July 2026
Last Modified:	07 Oct 2026 10:45
URI:	https://orca.cardiff.ac.uk/id/eprint/190075

Actions (repository staff only)

Edit Item