Large language models used in clinical classification can be swayed by spurious features, according to a new preprint on arXiv. The authors report that swapping a single gender token in a patient's clinical notes altered gender-irrelevant diagnoses in up to 53.5% of cases. This suggests the models are picking up on superficial cues rather than robust clinical evidence.

The paper, titled "Trajectory-Guided Steering: Mitigating Spurious Feature Bias in Clinical Classification," proposes a correction approach called trajectory-guided steering. The abstract notes that existing inference-time corrections are either incomplete or introduce other issues, though the specific limitations are not detailed in the available excerpt.

Because the abstract is truncated, the full mechanics of the proposed method and its evaluation are not yet visible. The core finding, however, underscores a persistent risk: even small, irrelevant changes to input text can shift model predictions in high-stakes medical settings.