In recent years, the black-box nature of deep learning models has limited their application in high-stakes domains such as medical diagnosis and finance, where interpretability is essential. To address this, we propose a novel approach using influence functions to enhance interpretability in NLP models at both the sample and concept levels. Experiments on CEBaB and Yelp datasets show that influence functions effectively identify the most impactful training samples, both helpful and harmful, on model predictions. By adjusting the labels and weights of these samples, we demonstrate that model performance can be restored to baseline levels without retraining, confirming the value of influence functions for efficient data debugging. Furthermore, our concept-level analysis identifies key concepts within Concept Bottleneck Models (CBM) that significantly affect predictions. Modifying these concepts alters model behavior observably, providing clear insights into the decision process.
@article{arxiv.2605.19848,
title = {CLIF: Concept-Level Influence Functions for Transparent Bottleneck Models},
author = {Yike Sun and Mingkun Xu and Mu You and Zhongzhi He and Henghua Shen and Zehan Tan and Derek F. Wong and Tao Fang},
journal= {arXiv preprint arXiv:2605.19848},
year = {2026}
}
Comments
A critical theoretical error invalidates the main results. The independence assumption on concept representations and gradients (Section 3.2, Eq.7) is incorrect, breaking the influence estimation in nonlinear bottleneck layers. This flaw undermines all empirical claims in Sections 4-5. The authors withdraw to prevent dissemination of incorrect findings