Title: Leveraging Disentanglement for Equivariant Learning in Compositional Visual Question Answering

Abstract: Compositional visual question answering (VQA) stands as a critical yet demanding objective, necessitating that systems grasp new amalgamations of concepts they have encountered before. Existing approaches frequently neglect the separation of these foundational concepts and struggle to adequately model the mechanisms behind compositional changes. Furthermore, leading techniques often rely on supplementary training signals that are impractical for authentic VQA applications. To overcome these limitations, this study presents a new framework named Disentanglement-based EquivAriant Learning (DEAL) tailored for compositional VQA. This framework operates solely on ground-truth answers. Within DEAL, we utilize interventions inspired by causal reasoning to separate concepts extracted from both visual and textual data inside a re-encoding structure. Guided by the principle of equivariance, we then apply compositional transformations to the inference inputs and enforce equivariant constraints on the outputs, thereby enhancing the model’s ability to reason compositionally. Extensive testing on the CLEVR-CoGenT and GQA-SGL benchmark datasets confirms that our DEAL method outperforms current state-of-the-art techniques in compositional VQA, demonstrating robust performance in both visual and linguistic generalization contexts.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC