Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Sun, Yike, Xu, Mingkun, You, Mu, He, Zhongzhi, Shen, Henghua, Tan, Zehan, Wong, Derek F., Fang, Tao
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2605.19848
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866910252148981760
author Sun, Yike
Xu, Mingkun
You, Mu
He, Zhongzhi
Shen, Henghua
Tan, Zehan
Wong, Derek F.
Fang, Tao
author_facet Sun, Yike
Xu, Mingkun
You, Mu
He, Zhongzhi
Shen, Henghua
Tan, Zehan
Wong, Derek F.
Fang, Tao
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2605_19848
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle CLIF: Concept-Level Influence Functions for Transparent Bottleneck Models
Sun, Yike
Xu, Mingkun
You, Mu
He, Zhongzhi
Shen, Henghua
Tan, Zehan
Wong, Derek F.
Fang, Tao
Computation and Language
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.
title CLIF: Concept-Level Influence Functions for Transparent Bottleneck Models
topic Computation and Language
url https://arxiv.org/abs/2605.19848