Visual question answering (VQA) models are prone to a known bias problem: they overfit to regularities in frequent patterns rather than truly understanding the content. According to a new arXiv preprint, previous research on this issue has largely concentrated on language bias, where models rely on superficial textual correlations. The paper argues that this view is incomplete.
The authors propose a variational and decomposed method specifically designed to mitigate what they call bimodal bias. Rather than treating bias as a purely linguistic artifact, the method accounts for biases arising from both visual and textual inputs. The decomposition step separates these sources of bias, while the variational component helps the model learn more robust representations.
The work is described in arXiv:2610.22845, and the abstract suggests that addressing both modalities together is key to building more reliable VQA systems. The paper does not provide full experimental details in the abstract, but the framing points to a broader shift in how VQA bias is understood and tackled.