Interventional Contrastive Learning Through Dual Knowledge Distillation on Bongard Problems

BONGARD-LOGO is a visual reasoning benchmark for evaluating few-shot concept learning from positive and negative examples. Its problems require models to learn both common visual knowledge and task-specific knowledge across different problem types. However, joint training may introduce interference between different types of specialized knowledge, while separate training weakens shared representation learning. To address this issue, we propose dual knowledge distillation-interventional contrastive learning (DKD-ICL). Specifically, we construct a structural causal model (SCM) in which task-specific knowledge from nontarget problem types is treated as a confounding factor. Based on this analysis, DKD-ICL combines contrastive learning with dual-stage knowledge distillation (KD) to disentangle shared and task-specific features. Experimental results show that DKD-ICL achieves 97.5% on basic shape (BA), 89.7% on free-form shape (FF), 72.0% on combinatorial abstract shape (CM), and 73.8% on novel abstract shape (NV), significantly outperforming existing methods and demonstrating its effectiveness in terms of mitigating cross-task interference while preserving shared semantic knowledge.