<p>While mixed-based data augmentation effectively enhances generalization, it frequently encounters the challenge of semantic ambiguity, where the synthesized visual content diverges from the assigned labels. Current solutions often depend on heavy external computations, limiting their practicality. To overcome this, we propose Adaptive Semantic Mixing (ASMix), a closed-loop framework that integrates attention-driven feedback to align labels with mixed input samples at virtually no extra computational cost. ASMix operates on two coupled levels. In the image space, we introduce an adaptive multi-scale mixing paradigm that asymmetrically integrates local-to-global contextual information, thereby preserving the structural integrity of discriminative regions. Crucially, in the label space, we establish a closed-loop feedback mechanism via Semantic-aware Label Assignment (SaLA). By leveraging the model’s intrinsic attention maps as a real-time perceptual feedback signal, SaLA dynamically rectifies the target label distribution to align with the network’s actual focus. This ensures that supervision signals remain semantically faithful to the augmented input. Extensive experiments demonstrate that this attention-driven feedback loop significantly enhances learning efficiency and generalization. Moreover, ASMix functions as a potent regularizer, effectively reducing model sensitivity to object scale variations.</p>