Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs

Large vision-language models (LVLMs) achieve strong multimodal understanding, but the hundreds to thousands of visual tokens they process impose substantial computational overhead, motivating training-free visual token pruning. In this work, we conduct two complementary analyses of visual token pruning. First, we measure the feature-space coverage of tokens retained before cross-modal fusion and find that aggressive pruning discards substantial visual information. Second, we track text-to-visual attention across decoder layers and find that the visual tokens considered important change substantially with depth, making one-shot pruning decisions unreliable. Together, these findings show that effective pruning should preserve broad visual coverage before fusion and progressively refine the retained tokens as cross-modal evidence evolves during fusion. We therefore propose STAR-Pro (STage-Wise Adaptive Token Reduction with Progressive Refinement), a training-free two-stage framework. Its Adaptive Stage applies pivoted QR to construct an over-budget feature-coverage candidate pool, while its Progressive Stage uses evolving text-to-visual attention at selected decoder layers to prune a nested survivor set under a target layer-average token budget. Extensive experiments across seven LVLMs spanning multiple architectures and 18 image and video benchmarks demonstrate the effectiveness of STAR-Pro under aggressive pruning. On LLaVA-Video-7B, STAR-Pro reduces visual tokens by 90.5%, retains 92.7% of baseline performance, and achieves a $2.24\times$ measured inference speedup. Code is available at https://github.com/EasonAI-5589/starpro.

High-resolution Large Vision-Language Models (LVLMs) preserve critical, fine-grained visual details by encoding a global thumbnail alongside multiple local crops. However, the resulting thousands of vision tokens introduce severe computational bottlenecks during Large Language Model (LLM) inference. Existing token pruning methods rely primarily on query-agnostic visual saliency, while current query-aware approaches either operate at a coarse crop-level resolution or compute token relevance too late, after vision tokens have already entered the heavy LLM layers. To address this, we propose a training-free, hierarchical token pruning framework that utilizes query guidance prior to executing any LLM blocks. Our top-down hierarchy first operates at the instance-level, dynamically scaling the vision token retention ratio of each sample based on its visual complexity. Next, at the crop-level, we distribute this sample budget across local crops using sentence-level cross-modal similarity. Finally, at the token-level, we exploit projected vision tokens alongside textual query token embeddings in the shared space to measure fine-grained alignment, applying a text-debiasing normalization to produce high-fidelity query scores. These query-guided budgets and scores are fused with local and global visual saliency metrics to select the most critical vision tokens. Extensive experiments on LLaVA-NeXT-7B demonstrate that our plug-and-play framework retains 98.0% and 94.8% of unpruned performance at 25% and 10% vision token retention, respectively. Notably, at 25% retention, our method delivers a 2.0x inference throughput speedup and even surpasses the unpruned baseline on POPE by 0.7 points.

While Large Vision-Language Models (LVLMs) demonstrate exceptional multi-modal capabilities, the quadratic computational cost of processing high-resolution visual tokens remains a critical bottleneck. Though recent token reduction strategies attempt to accelerate inference, such methods inadequately exploit attention values and fail to address token redundancy. More critically, they overlook the ``attention shift'' phenomenon inherent in LVLMs, which skews token attention scores. In this work, we propose ASAP, a novel training-free, KV-Cache-compatible pruning recipe that comprehensively addresses these limitations. First, we mitigate the attention shift by utilizing a dynamic bidirectional soft attention mask, ensuring the selection of genuinely informative tokens rather than naive attention-based selection. Second, we posit that high semantic redundancy within the token set degrades performance. We therefore introduce a weighted soft merging component that merges semantically similar tokens, preserving only the most feature-dense visual patches for subsequent layers. ASAP achieves virtually lossless compression of visual context, retaining 99.02% of the original LLaVA-NeXT-7B performance while aggressively slashing computational FLOPs by ~80%.