Fairness is Not Flat: Geometric Phase Transitions Against Shortcut Learning

Shortcut learning denotes the widespread situation in which a classifier exploits spurious correlations rather than diagnostic features. Existing mitigation strategies mostly aim to learn shortcut-invariant representations; their empirical success is limited and they cannot be applied to classifiers using frozen foundation model encoders. We propose to reframe shortcut learning as fundamentally a calibration problem: unconstrained learning implicitly calibrates each shortcut group to its training set disease prevalence, rendering the resulting classifier necessarily over-confident in one group and under-confident in the other. Building on this insight, we prevalence-equalize calibration between shortcut groups through two encoder-agnostic methods, an in-processing regularizer and a post-hoc prevalence-equalized recalibration step. Across chest-drain-pneumothorax benchmarks on CheXpert and SIIM-ACR, spanning fine-tuned CNNs and frozen foundation-model backbones, both methods substantially outperform all baselines. Post-hoc recalibration of a standard ERM-trained DenseNet raises misaligned-group AUROC from 0.23 to 0.73, indicating that shortcut reliance degrades the classification head rather than the underlying representation. Besides two new state-of-the-art shortcut mitigation approaches, our findings more fundamentally connect shortcut learning to calibration theory and algorithmic fairness.

This work presents Fairness Pruning, a lightweight structural intervention method designed for the management and future mitigation of demographic bias in large language models (LLMs). As a foundational empirical validation of this method, this work focuses on causal bias localization. Using minimally contrastive prompt pairs and inference-time activation capture, the method identifies neurons that react differentially when processing demographic attributes in GLU architectures, evaluating the signal at the down_proj input. Empirical evaluation was conducted on models of up to 3 billion parameters (Llama-3.2 family and Salamandra-2B), combining standardized benchmark evaluation with qualitative text generation experiments. Results demonstrate that zeroing the identified neurons alters how the model responds to associated demographic variables. However, rather than producing flat mitigation, the intervention causes bidirectional bias destabilization: because BiasScore is unsigned, candidate sets mix neurons that push toward and against the stereotype, and the net effect on aggregate bias depends on which sign dominates. The intervention is extremely surgical: zeroing at most 40 neurons in Llama-3.2-1B (less than 0.031% of total MLP width) achieves a mean retention of 99.49% in reasoning and general knowledge capabilities. These findings empirically confirm that demographic bias processing and model capabilities operate on dissociable circuits, establishing the methodological foundations for transitioning from blind zeroing toward directional behavior modulation.

Shortcut mitigation strategies commonly rely on training data annotations, group-balanced held-out data or the presence of all groups, i.e., all combinations of (spurious) attributes and classes, in the training data. However, these requirements are rarely met in practice. We instead propose a method for targeted model analysis to identify a small set of instances in which the model relies on spurious attributes. Using that set and following ``this feature should not be used for prediction'' reasoning, we identify highly relevant neurons in an intermediate layer and regularize their impact. This ensures that models learn to depend on informative features rather than being right for the wrong reasons, thereby improving robustness without requiring additional balanced held-out data or annotations.