Decoupled Similarity for Task-Aware Token Pruning in Large Vision-Language Models

Vision-Language Models (VLMs) have recently demonstrated remarkable capabilities in visual understanding and reasoning, but they also impose significant computational burdens due to long visual sequence inputs. Recent works address this issue by pruning unimportant visual tokens, achieving substantial computational reduction while maintaining model performance. The core of token pruning lies in determining token importance, with current approaches primarily relying on attention scores from vision encoders or Large Language Models (LLMs). In this paper, we analyze the effectiveness of attention mechanisms in both vision encoders and LLMs. We find that vision encoders suffer from attention sink, leading to poor focus on informative foreground regions, while in LLMs, although prior studies have identified attention bias toward token positions, text-to-vision attention demonstrates resistance to this bias and enables effective pruning guidance in middle layers. Based on these observations, we propose LearnPruner, a two-stage token pruning framework that first removes redundant vision tokens via a learnable pruning module after the vision encoder, then retains only task-relevant tokens in the LLM's middle layer. Experimental results show that our LearnPruner can preserve approximately 95% of the original performance while using only 5.5% of vision tokens, and achieve 3.2$\times$ inference acceleration, demonstrating a superior accuracy-efficiency trade-off.

Vision-Language Models (VLMs) have demonstrated remarkable multimodal capabilities but suffer from prohibitive computational overhead due to the massive number of visual tokens. While token pruning is a promising solution, existing methods still suffer from two major drawbacks: 1) unreliable pruning criteria that fail to accurately localize task-specific regions due to inherent attention biases or query-agnostic designs, and 2) suboptimal pruning strategies that employ rigid, non-adaptive schedules, overlooking the dynamic evolution of token redundancy across different model depths. To address these challenges, we propose LearnPrunerV2, which extends the previous two-stage heuristic framework (LearnPruner) into a fully learnable architecture. Specifically, we integrate lightweight learnable modules into every layer of the VLM backbone, which estimate token importance by fusing intrinsic visual saliency with extrinsic query relevance captured via a shadow token. By uncovering a consistent staircase redundancy pattern across VLM layers, we further introduce Budget-Deterministic Pruning (BDP) to establish a hardware-friendly schedule with fixed per-layer budgets. Extensive experimental results across various benchmarks and VLMs demonstrate that LearnPrunerV2 consistently outperform previous state-of-the-art methods. Notably, LearnPrunerV2 preserves 99.8% of the original performance when pruning 77.8% of visual tokens on LLaVA-v1.5, achieving a superior accuracy-efficiency trade-off.

Visual token pruning has been widely studied as a practical approach to reducing the computational cost of large vision-language models. However, it struggles to preserve essential visual information, which can lead to substantial performance degradation. In particular, image-based token selection can overlook task-relevant details, while text-guided token selection may fail to capture the text--visual relationships needed for complex reasoning. We find that applying textual guidance too early can limit its ability to identify answer-relevant visual regions, whereas text-to-visual attention becomes more informative at intermediate decoder depths. This finding motivates our training-free method, which separates early vision-guided pruning from deferred text-guided reselection. We first prune visual tokens using vision-encoder attention, retain additional candidates until the decoder midpoint, and then use text-to-visual attention to determine the final visual-token set. Across eight benchmarks and three models, our method outperforms the best-performing baselines by an average of 11.10 and 16.84 percentage points in performance recovery at 80% and 90% pruning, respectively, with comparable or lower LLM-prefill latency than most baselines. The source code is publicly available at https://github.com/kmc3661/DeFT