Answers are grounded in this paper and cite their sources.
This framework quantifies shortcut learning in transformer models, uncovering universal patterns in biases and accuracy.
We present a novel framework for quantifying shortcut learning in transformer-based language models through strategically designed minimal pairs. The proposed approach introduces the Shortcut Reliance Score (SRS), a theoretically grounded metric that identifies when models rely on spurious correlations versus robust features. Through comprehensive analysis of 1,250 minimal pairs across five datasets (HANS, PAWS, WinoBias, BoolQ, IMDB), we uncover a striking discovery: word-order processing errors (PAWS) and gender-occupation biases (WinoBias) exhibit 0.85 correlation, revealing that social biases emerge from syntactic shortcuts rather than biased training data. This finding fundamentally reframes debiasing from a data problem to an architectural one. Layer-wise analysis reveals monotonic vulnerability increase (0.34 to 0.62) with critical transitions at layers 3, 7, and 10, aligning with known linguistic processing stages. Cross-dataset analysis demonstrates that shortcuts form a transferable hierarchy, with correlations reaching 0.85 across seemingly unrelated tasks. The proposed method achieves 89% effectiveness in shortcut detection compared to 32% for random baselines, while requiring only CPU resources. These findings challenge the prevailing assumption that different benchmarks test independent capabilities, instead revealing that model failures stem from universal architectural patterns that manifest differently across domains.
No takes yet. Share an insight, caveat, or question.
May Stow May Stow (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: