Stress Testing Generalization: How Minor Modifications Undermine Large Language Model Performance

This paper investigates the fragility of Large Language Models (LLMs) in generalizing to novel inputs, specifically focusing on minor perturbations in well-established benchmarks (e.g., slight changes in question format or distractor length). Despite high benchmark scores, LLMs exhibit significant accuracy drops and unexpected biases (e.g., preference for longer distractors) when faced with these minor but content-preserving modifications. For example, Qwen 2.5 1.5B's MMLU score rises from 60 to 89 and drops from 89 to 36 when option lengths are changed without altering the question. Even GPT-4 experiences a 25-point accuracy loss when question types are changed, with a 6-point drop across all three modification categories. These analyses suggest that LLMs rely heavily on superficial cues rather than forming robust, abstract representations that generalize across formats, lexical variations, and irrelevant content shifts. This work aligns with the ACL 2025 theme track on the Generalization of NLP models, proposing a "Generalization Stress Test" to assess performance shifts under controlled perturbations. The study calls for reevaluating benchmarks and developing more reliable evaluation methodologies to capture LLM generalization abilities better.

centroIDA: Cross-Domain Class Discrepancy Minimization Based on Accumulative Class-Centroids for Imbalanced Domain Adaptation

Unsupervised Domain Adaptation (UDA) approaches address the covariate shift problem by minimizing the distribution discrepancy between the source and target domains, assuming that the label distribution is invariant across domains. However, in the imbalanced domain adaptation (IDA) scenario, covariate and long-tailed label shifts both exist across domains. To tackle the IDA problem, some current research focus on minimizing the distribution discrepancies of each corresponding class between source and target domains. Such methods rely much on the reliable pseudo labels' selection and the feature distributions estimation for target domain, and the minority classes with limited numbers makes the estimations more uncertainty, which influences the model's performance. In this paper, we propose a cross-domain class discrepancy minimization method based on accumulative class-centroids for IDA (centroIDA). Firstly, class-based re-sampling strategy is used to obtain an unbiased classifier on source domain. Secondly, the accumulative class-centroids alignment loss is proposed for iterative class-centroids alignment across domains. Finally, class-wise feature alignment loss is used to optimize the feature representation for a robust classification boundary. A series of experiments have proved that our method outperforms other SOTA methods on IDA problem, especially with the increasing degree of label shift.