On the Reliability of Cue Conflict and Beyond

The article introduces REFINED-BIAS, a new dataset and evaluation framework designed to enhance the reliability of diagnosing shape-texture biases in neural networks. It addresses limitations of existing cue-conflict benchmarks by constructing balanced cue pairs based on explicit definitions and measuring cue-specific sensitivity across the full label space with a ranking-based metric. This framework allows for more accurate cross-model comparisons and clearer empirical insights into bias mechanisms, which is crucial for practitioners aiming to interpret and improve model performance in visual tasks.