Short-circuiting Shortcuts: Mechanistic Investigation of Shortcuts in Text Classification

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Eshuijs, Leon, Wang, Shihan, Fokkens, Antske
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918015077974016
author Eshuijs, Leon
Wang, Shihan
Fokkens, Antske
author_facet Eshuijs, Leon
Wang, Shihan
Fokkens, Antske
contents Reliance on spurious correlations (shortcuts) has been shown to underlie many of the successes of language models. Previous work focused on identifying the input elements that impact prediction. We investigate how shortcuts are actually processed within the model's decision-making mechanism. We use actor names in movie reviews as controllable shortcuts with known impact on the outcome. We use mechanistic interpretability methods and identify specific attention heads that focus on shortcuts. These heads gear the model towards a label before processing the complete input, effectively making premature decisions that bypass contextual analysis. Based on these findings, we introduce Head-based Token Attribution (HTA), which traces intermediate decisions back to input tokens. We show that HTA is effective in detecting shortcuts in LLMs and enables targeted mitigation by selectively deactivating shortcut-related attention heads.
format Preprint
id arxiv_https___arxiv_org_abs_2505_06032
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Short-circuiting Shortcuts: Mechanistic Investigation of Shortcuts in Text Classification
Eshuijs, Leon
Wang, Shihan
Fokkens, Antske
Machine Learning
Computation and Language
Reliance on spurious correlations (shortcuts) has been shown to underlie many of the successes of language models. Previous work focused on identifying the input elements that impact prediction. We investigate how shortcuts are actually processed within the model's decision-making mechanism. We use actor names in movie reviews as controllable shortcuts with known impact on the outcome. We use mechanistic interpretability methods and identify specific attention heads that focus on shortcuts. These heads gear the model towards a label before processing the complete input, effectively making premature decisions that bypass contextual analysis. Based on these findings, we introduce Head-based Token Attribution (HTA), which traces intermediate decisions back to input tokens. We show that HTA is effective in detecting shortcuts in LLMs and enables targeted mitigation by selectively deactivating shortcut-related attention heads.
title Short-circuiting Shortcuts: Mechanistic Investigation of Shortcuts in Text Classification
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2505.06032