# Visual Token Pruning

Vision-language-action policies process hundreds of visual tokens per step, which drives up inference latency. APS-VLA keeps the backbone frozen and uses a lightweight scorer conditioned on the current observation, instruction and proprioception to pick tokens, training the retained representation both to reproduce policy outputs and to predict the subsequent visual change. On LIBERO with 128 tokens retained, mean success is 95.75%, 1.25 points below the unpruned 97.00%, and inference latency falls to 68.75 ms.

Vision-language-action policies process hundreds of visual tokens per step, which drives up inference latency. APS-VLA keeps the backbone frozen and uses a lightweight scorer conditioned on the current observation, instruction and proprioception to pick tokens, training the retained representation both to reproduce policy outputs and to predict the subsequent visual change. On LIBERO with 128 tokens retained, mean success is 95.75%, 1.25 points below the unpruned 97.00%, and inference latency falls to 68.75 ms.

Large vision-language models process hundreds to thousands of visual tokens per image, and inference cost rises with them. Earlier training-free pruning methods ranked all visual tokens by one importance criterion, overlooking that tokens take on different roles as encoding proceeds. BtsPruner locates the role transition from changes in attention concentration, then picks globally referenced representative tokens from later layers and recursively retrieves local detail tokens from earlier layers, preserving both roles under a fixed budget; on LLaVA-1.5-7B it keeps only 32 of 576 visual tokens (5.6%) and retains 91.6% of the dense model's performance.

Large vision-language models process hundreds to thousands of visual tokens per image, and inference cost rises with them. Earlier training-free pruning methods ranked all visual tokens by one importance criterion, overlooking that tokens take on different roles as encoding proceeds. BtsPruner locates the role transition from changes in attention concentration, then picks globally referenced representative tokens from later layers and recursively retrieves local detail tokens from earlier layers, preserving both roles under a fixed budget; on LLaVA-1.5-7B it keeps only 32 of 576 visual tokens (5.6%) and retains 91.6% of the dense model's performance.

Visual token selection in multimodal LLMs now propagates scores over a visual graph and corrects for neighborhood degree: across four backbones, ten benchmarks and seven pruning ratios, LLaVA-NeXT retains 98.7% accuracy while keeping 30% of visual tokens, with a net prefill speedup.

Visual token selection in multimodal LLMs now propagates scores over a visual graph and corrects for neighborhood degree: across four backbones, ten benchmarks and seven pruning ratios, LLaVA-NeXT retains 98.7% accuracy while keeping 30% of visual tokens, with a net prefill speedup.

FlashGaze is a training-free method that reduces spatiotemporal redundancy before ViT encoding by using pixel-space differences as a proxy for information loss and Quadtree Dynamic Programming to jointly decide patch dropping, merging, and keeping under a fixed budget; on Qwen3-VL-8B and NVILA-8B-Video it achieves higher accuracy and lower TTFT across multiple video understanding benchmarks, and at a 20% retained ratio preserves 98% of the full-input LongVideoBench accuracy.

FlashGaze is a training-free method that reduces spatiotemporal redundancy before ViT encoding by using pixel-space differences as a proxy for information loss and Quadtree Dynamic Programming to jointly decide patch dropping, merging, and keeping under a fixed budget; on Qwen3-VL-8B and NVILA-8B-Video it achieves higher accuracy and lower TTFT across multiple video understanding benchmarks, and at a 20% retained ratio preserves 98% of the full-input LongVideoBench accuracy.