We introduce a representation-level counterfactual framework to mitigate co-occurrence bias in vision–language models without retraining or prompt engineering.