Retrieval‑augmented QA can still fail even when the correct answer appears in the retrieved evidence. The failure arises because a competing answer may satisfy only part of the question’s requirements but is mentioned frequently in the retrieved passages, making it seem the best‑supported choice. To test the direct impact of retrieval bias, the authors increased the proportion of passages favoring the competing answer; the model then became more likely to select it, demonstrating that candidate‑level support alone is insufficient for a correct decision. To address this, they propose Joint Constraint Aggregation (JCA), which evaluates all answer candidates against the same set of question‑derived constraints using the retrieved evidence. Across three evaluation settings and three instruction‑tuned backbones, JCA consistently improves overall accuracy, with a larger advantage when retrieval is skewed toward a competing answer. These findings suggest that when partial matches and retrieval prominence pull models toward an incorrect answer, comparing candidates against the full set of question requirements makes selection more reliable.
Review: JCA effectively mitigates retrieval bias and improves answer selection reliability.