STAR: Stage-Wise Attention-Guided Token Reduction for Efficient Large Vision-Language Models Inference

Although large vision-language models (LVLMs) leverage rich visual token representations to achieve strong performance on multimodal tasks, these tokens also introduce significant computational overhead during inference. Existing training-free token pruning methods typically adopt a single-stage strategy, focusing either on visual self-attention or visual-textual cross-attention. However, such localized perspectives often overlook the broader information flow across the model, leading to substantial performance degradation, especially under high pruning ratios. In this work, we propose STAR (Stage-wise Attention-guided token Reduction), a training-free, plug-and-play framework that approaches token pruning from a global perspective. Instead of pruning at a single point, STAR performs attention-guided reduction in two complementary stages: an early-stage pruning based on visual self-attention to remove redundant low-level features, and a later-stage pruning guided by cross-modal attention to discard task-irrelevant tokens. This holistic approach allows STAR to significantly reduce computational cost while better preserving task-critical information. Extensive experiments across multiple LVLM architectures and benchmarks show that STAR achieves strong acceleration while maintaining comparable, and in some cases even improved performance.

Mitigating Hallucinations via Inter-Layer Consistency Aggregation in Large Vision-Language Models

Despite the impressive capabilities of Large Vision-Language Models (LVLMs), they remain susceptible to hallucinations-generating content that is inconsistent with the input image. Existing training-free hallucination mitigation methods often suffer from unstable performance and high sensitivity to hyperparameter settings, limiting their practicality and broader adoption. In this paper, we propose a novel decoding mechanism, Decoding with Inter-layer Consistency via Layer Aggregation (DCLA), which requires no retraining, fine-tuning, or access to external knowledge bases. Specifically, our approach constructs a dynamic semantic reference by aggregating representations from previous layers, and corrects semantically deviated layers to enforce inter-layer consistency. The method allows DCLA to robustly mitigate hallucinations across multiple LVLMs. Experiments on hallucination benchmarks such as MME and POPE demonstrate that DCLA effectively reduces hallucinations while enhancing the reliability and performance of LVLMs.

Enhancing Layer Attention Efficiency through Pruning Redundant Retrievals

Growing evidence suggests that layer attention mechanisms, which enhance interaction among layers in deep neural networks, have significantly advanced network architectures. However, existing layer attention methods suffer from redundancy, as attention weights learned by adjacent layers often become highly similar. This redundancy causes multiple layers to extract nearly identical features, reducing the model's representational capacity and increasing training time. To address this issue, we propose a novel approach to quantify redundancy by leveraging the Kullback-Leibler (KL) divergence between adjacent layers. Additionally, we introduce an Enhanced Beta Quantile Mapping (EBQM) method that accurately identifies and skips redundant layers, thereby maintaining model stability. Our proposed Efficient Layer Attention (ELA) architecture, improves both training efficiency and overall performance, achieving a 30\% reduction in training time while enhancing performance in tasks such as image classification and object detection.