Long-Tailed Distribution-Aware Router For Mixture-of-Experts in Large Vision-Language Model

The mixture-of-experts (MoE), which replaces dense models with sparse architectures, has gained attention in large vision-language models (LVLMs) for achieving comparable performance with fewer activated parameters. Existing MoE frameworks for LVLMs focus on token-to-expert routing (TER), encouraging different experts to specialize in processing distinct tokens. However, these frameworks often rely on the load balancing mechanism, overlooking the inherent distributional differences between vision and language. To this end, we propose a Long-Tailed Distribution-aware Router (LTDR) for vision-language TER, tackling two challenges: (1) Distribution-aware router for modality-specific routing. We observe that language TER follows a uniform distribution, whereas vision TER exhibits a long-tailed distribution. This discrepancy necessitates distinct routing strategies tailored to each modality. (2) Enhancing expert activation for vision tail tokens. Recognizing the importance of vision tail tokens, we introduce an oversampling-like strategy by increasing the number of activated experts for these tokens. Experiments on extensive benchmarks validate the effectiveness of our approach.

Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model

The Mixture-of-Experts (MoE) has gained increasing attention in the study of Large Vision-Language Models (LVLMs). It uses a sparse model to replace the dense model, achieving comparable performance while activating fewer parameters during inference, thus significantly reducing the inference cost. Existing MoE methods in LVLMs encourage different experts to handle different tokens, and thus they employ a router to predict the routing for each token. However, the predictions are based solely on sample features and do not truly reveal the optimization direction of tokens. This can lead to severe optimization conflicts between different tokens within an expert. To address this problem, this paper proposes a novel method based on token-level gradient analysis. Specifically, we first use token-level gradients to identify conflicting tokens in experts. Then, we add a specialized loss tailored to eliminate conflicts among tokens within each expert. Our method can serve as a plug-in for diverse Large Vision-Language Models, and extensive experimental results demonstrate the effectiveness of our method. The code will be publicly available at https://github.com/longrongyang/STGC.

Bridging Cross-task Protocol Inconsistency for Distillation in Dense Object Detection

Knowledge distillation (KD) has shown potential for learning compact models in dense object detection. However, the commonly used softmax-based distillation ignores the absolute classification scores for individual categories. Thus, the optimum of the distillation loss does not necessarily lead to the optimal student classification scores for dense object detectors. This cross-task protocol inconsistency is critical, especially for dense object detectors, since the foreground categories are extremely imbalanced. To address the issue of protocol differences between distillation and classification, we propose a novel distillation method with cross-task consistent protocols, tailored for the dense object detection. For classification distillation, we address the cross-task protocol inconsistency problem by formulating the classification logit maps in both teacher and student models as multiple binary-classification maps and applying a binary-classification distillation loss to each map. For localization distillation, we design an IoU-based Localization Distillation Loss that is free from specific network structures and can be compared with existing localization distillation losses. Our proposed method is simple but effective, and experimental results demonstrate its superiority over existing methods. Code is available at https://github.com/TinyTigerPan/BCKD.