Despite their powerful capabilities, Multimodal Large Language Models (MLLMs) suffer from considerable computational overhead due to their reliance on massive visual tokens. Recent studies have explored token pruning to alleviate this problem, which typically uses text-vision cross-attention or [CLS] attention to assess and discard redundant visual tokens. In this work, we identify a critical limitation of such attention-first pruning approaches, i.e., they tend to preserve semantically similar tokens, resulting in pronounced performance drops under high pruning ratios. To this end, we propose HoloV, a simple yet effective, plug-and-play visual token pruning framework for efficient inference.

Large models achieve strong performance on Vision-and-Language Navigation (VLN) tasks, but are costly to run in resource-limited environments. Token pruning offers appealing tradeoffs for efficiency with minimal performance loss by reducing model input size, but prior work overlooks VLN-specific challenges. For example, information loss from pruning can effectively increase computational cost due to longer walks. Thus, the inability to identify uninformative tokens undermines the supposed efficiency gains from pruning. To address this, we propose Navigation-Aware Pruning (NAP), which uses navigation-specific traits to simplify the pruning process by pre-filtering tokens into foreground and background. For example, image views are filtered based on whether the agent can navigate in that direction. We also extract navigation-relevant instructions using a Large Language Model. After filtering, we focus pruning on background tokens, minimizing information loss. To further help avoid increases in navigation length, we discourage backtracking by removing low-importance navigation nodes. Experiments on standard VLN benchmarks show NAP significantly outperforms prior work, preserving higher success rates while saving more than 50% FLOPS.

Existing visual token pruning methods target prompt alignment and visual preservation with static strategies, overlooking the varying relative importance of these objectives across tasks, which leads to inconsistent performance. To address this, we derive the first closed-form error bound for visual token pruning based on the Hausdorff distance, uniformly characterizing the contributions of both objectives. Moreover, leveraging $ε$-covering theory, we reveal an intrinsic trade-off between these objectives and quantify their optimal attainment levels under a fixed budget. To practically handle this trade-off, we propose Multi-Objective Balanced Covering (MoB), which reformulates visual token pruning as a bi-objective covering problem. In this framework, the attainment trade-off reduces to budget allocation via greedy radius trading. MoB offers a provable performance bound and linear scalability with respect to the number of input visual tokens, enabling adaptation to challenging pruning scenarios. Extensive experiments show that MoB preserves 96.4% of performance for LLaVA-1.5-7B using only 11.1% of the original visual tokens and accelerates LLaVA-Next-7B by 1.3-1.5$\times$ with negligible performance loss. Additionally, evaluations on Qwen2-VL and Video-LLaVA confirm that MoB integrates seamlessly into advanced MLLMs and diverse vision-language tasks.

Vision-Language Models (VLMs) demand substantial computational resources during inference, largely due to the extensive visual input tokens for representing visual information. Previous studies have noted that visual tokens tend to receive less attention than text tokens, suggesting their lower importance during inference and potential for pruning. However, their methods encounter several challenges: reliance on greedy heuristic criteria for token importance and incompatibility with FlashAttention and KV cache. To address these issues, we introduce \textbf{TopV}, a compatible \textbf{TO}ken \textbf{P}runing with inference Time Optimization for fast and low-memory \textbf{V}LM, achieving efficient pruning without additional training or fine-tuning. Instead of relying on attention scores, we formulate token pruning as an optimization problem, accurately identifying important visual tokens while remaining compatible with FlashAttention. Additionally, since we only perform this pruning once during the prefilling stage, it effectively reduces KV cache size. Our optimization framework incorporates a visual-aware cost function considering factors such as Feature Similarity, Relative Spatial Distance, and Absolute Central Distance, to measure the importance of each source visual token, enabling effective pruning of low-importance tokens. Extensive experiments demonstrate that our method outperforms previous token pruning methods, validating the effectiveness and efficiency of our approach.

Vision tokens in multimodal large language models often dominate huge computational overhead due to their excessive length compared to linguistic modality. Abundant recent methods aim to solve this problem with token pruning, which first defines an importance criterion for tokens and then prunes the unimportant vision tokens during inference. However, in this paper, we show that the importance is not an ideal indicator to decide whether a token should be pruned. Surprisingly, it usually results in inferior performance than random token pruning and leading to incompatibility to efficient attention computation operators.Instead, we propose DART (Duplication-Aware Reduction of Tokens), which prunes tokens based on its duplication with other tokens, leading to significant and training-free acceleration. Concretely, DART selects a small subset of pivot tokens and then retains the tokens with low duplication to the pivots, ensuring minimal information loss during token pruning. Experiments demonstrate that DART can prune 88.9% vision tokens while maintaining comparable performance, leading to a 1.99$\times$ and 2.99$\times$ speed-up in total time and prefilling stage, respectively, with good compatibility to efficient attention operators. Our codes are available at https://github.com/ZichenWen1/DART.

Multimodal large language models (MLLMs) have shown remarkable performance for cross-modal understanding and generation, yet still suffer from severe inference costs. Recently, abundant works have been proposed to solve this problem with token pruning, which identifies the redundant tokens in MLLMs and then prunes them to reduce the computation and KV storage costs, leading to significant acceleration without training. While these methods claim efficiency gains, critical questions about their fundamental design and evaluation remain unanswered: Why do many existing approaches underperform even compared to naive random token selection? Are attention-based scoring sufficient for reliably identifying redundant tokens? Is language information really helpful during token pruning? What makes a good trade-off between token importance and duplication? Are current evaluation protocols comprehensive and unbiased? The ignorance of previous research on these problems hinders the long-term development of token pruning. In this paper, we answer these questions one by one, providing insights into the design of future token pruning methods.

Vision-Language Models (VLMs) have achieved notable success in multimodal tasks but face practical limitations due to the quadratic complexity of decoder attention mechanisms and autoregressive generation. Existing methods like FASTV and VTW have achieved notable results in reducing redundant visual tokens, but these approaches focus on pruning tokens in a single forward pass without systematically analyzing the redundancy of visual tokens throughout the entire generation process. In this paper, we introduce a dynamic pruning strategy tailored for VLMs, namedDynamic Rate (DyRate), which progressively adjusts the compression rate during generation. Our analysis of the distribution of attention reveals that the importance of visual tokens decreases throughout the generation process, inspiring us to adopt a more aggressive compression rate. By integrating a lightweight predictor based on attention distribution, our approach enables flexible adjustment of pruning rates based on the attention distribution. Our experimental results demonstrate that our method not only reduces computational demands but also maintains the quality of responses.

In large vision-language models (LVLMs), images serve as inputs that carry a wealth of information. As the idiom "A picture is worth a thousand words" implies, representing a single image in current LVLMs can require hundreds or even thousands of tokens. This results in significant computational costs, which grow quadratically as input image resolution increases, thereby severely impacting the efficiency of both training and inference. Previous approaches have attempted to reduce the number of image tokens either before or within the early layers of LVLMs. However, these strategies inevitably result in the loss of crucial image information, ultimately diminishing model performance. To address this challenge, we conduct an empirical study revealing that all visual tokens are necessary for LVLMs in the shallow layers, and token redundancy progressively increases in the deeper layers of the model. To this end, we propose PyramidDrop, a visual redundancy reduction strategy for LVLMs to boost their efficiency in both training and inference with neglectable performance loss. Specifically, we partition the LVLM into several stages and drop part of the image tokens at the end of each stage with a pre-defined ratio, creating pyramid-like visual tokens across model layers. The dropping is based on a lightweight similarity calculation with a negligible time overhead. Extensive experiments demonstrate that PyramidDrop can achieve a 40% training time and 55% inference FLOPs acceleration of LLaVA-NeXT with comparable performance. Besides, the PyramidDrop could also serve as a plug-and-play strategy for inference acceleration without training, with better performance and lower inference cost than counterparts. We hope that the insights and approach introduced by PyramidDrop will inspire future research to further investigate the role of image tokens in LVLMs.

Recent advancements in vision-language models (VLMs) have expanded their potential for real-world applications, enabling these models to perform complex reasoning on images. In the widely used fully autoregressive transformer-based models like LLaVA, projected visual tokens are prepended to textual tokens. Oftentimes, visual tokens are significantly more than prompt tokens, resulting in increased computational overhead during both training and inference. In this paper, we propose Visual Compact Token Registers (Victor), a method that reduces the number of visual tokens by summarizing them into a smaller set of register tokens. Victor adds a few learnable register tokens after the visual tokens and summarizes the visual information into these registers using the first few layers in the language tower of VLMs. After these few layers, all visual tokens are discarded, significantly improving computational efficiency for both training and inference. Notably, our method is easy to implement and requires a small number of new trainable parameters with minimal impact on model performance. In our experiment, with merely 8 visual registers--about 1% of the original tokens--Victor shows less than a 4% accuracy drop while reducing the total training time by 43% and boosting the inference throughput by 3.3 . Vision-language models (VLMs) have attracted considerable attention for their capability to process visual and textual information, enabling various real-world applications, such as image captioning, visual question answering, and multimodal reasoning (OpenAI, 2023; Liu et al., 2024c). For example, GPT-4V (OpenAI, 2023) demonstrates the potential of these models in helping visually impaired individuals to "see" the world through cell phone cameras.

In this study, we identify the inefficient attention phenomena in Large Vision-Language Models (LVLMs), notably within prominent models like LLaVA-1.5, QwenVL-Chat and Video-LLaVA. We find out that the attention computation over visual tokens is of extreme inefficiency in the deep layers of popular LVLMs, suggesting a need for a sparser approach compared to textual data handling. To this end, we introduce FastV, a versatile plug-and-play method designed to optimize computational efficiency by learning adaptive attention patterns in early layers and pruning visual tokens in subsequent ones. Our evaluations demonstrate FastV's ability to dramatically reduce computational costs (e.g., a 45 reduction in FLOPs for LLaVA-1.5-13B) without sacrificing performance in a wide range of image and video understanding tasks. The computational efficiency and performance trade-off of FastV are highly customizable and pareto-efficient. It can compress the FLOPs of a 13B-parameter model to achieve a lower budget than that of a 7B-parameter model, while still maintaining superior performance. We believe FastV has practical values for deployment of LVLMs in edge devices and commercial models. Code is released at https://github.com/pkunlp-icler/FastV.