ResPrune: Text-Conditioned Subspace Reconstruction for Visual Token Pruning in Large Vision-Language Models

Large Vision-Language Models (LVLMs) rely on dense visual tokens to capture fine-grained visual information, but processing all these tokens incurs substantial computational and memory overhead during inference. To address this issue, we propose ResPrune, a training-free visual token pruning framework that enables efficient LVLM inference by selecting a compact yet informative subset of visual tokens. ResPrune formulates visual token pruning as a subspace reconstruction problem and employs a greedy subspace expansion strategy guided by residual energy, allowing it to preserve the geometric structure of the original visual token space. To further incorporate cross modal alignment, the selection process is conditioned on textual relevance, encouraging the retention of tokens that are both informative and instruction-relevant. The proposed method is lightweight and model-agnostic, and can be seamlessly integrated into existing LVLM pipelines without retraining or architectural modifications. Extensive experiments on multiple LVLM backbones, including LLaVA-1.5, LLaVA-NeXT, and Qwen2.5-VL, demonstrate that ResPrune consistently outperforms existing pruning approaches across a wide range of benchmarks, while achieving effective reductions in computation, memory consumption, and inference latency.

DynamicVGGT: Learning Dynamic Point Maps for 4D Scene Reconstruction in Autonomous Driving

Dynamic scene reconstruction in autonomous driving remains a fundamental challenge due to significant temporal variations, moving objects, and complex scene dynamics. Existing feed-forward 3D models have demonstrated strong performance in static reconstruction but still struggle to capture dynamic motion. To address these limitations, we propose DynamicVGGT, a unified feed-forward framework that extends VGGT from static 3D perception to dynamic 4D reconstruction. Our goal is to model point motion within feed-forward 3D models in a dynamic and temporally coherent manner. To this end, we jointly predict the current and future point maps within a shared reference coordinate system, allowing the model to implicitly learn dynamic point representations through temporal correspondence. To efficiently capture temporal dependencies, we introduce a Motion-aware Temporal Attention (MTA) module that learns motion continuity. Furthermore, we design a Dynamic 3D Gaussian Splatting Head that explicitly models point motion by predicting Gaussian velocities using learnable motion tokens under scene flow supervision. It refines dynamic geometry through continuous 3D Gaussian optimization. Extensive experiments on autonomous driving datasets demonstrate that DynamicVGGT significantly outperforms existing methods in reconstruction accuracy, achieving robust feed-forward 4D dynamic scene reconstruction under complex driving scenarios.

Training-Free Pyramid Token Pruning for Efficient Large Vision-Language Models via Region, Token, and Instruction-Guided Importance

Large Vision-Language Models (LVLMs) have significantly advanced multimodal understanding but still struggle with efficiently processing high-resolution images. Recent approaches partition high-resolution images into multiple sub-images, dramatically increasing the number of visual tokens and causing exponential computational overhead during inference. To address these limitations, we propose a training-free token pruning strategy, Pyramid Token Pruning (PTP), that integrates bottom-up visual saliency at both region and token levels with top-down instruction-guided importance. Inspired by human visual attention mechanisms, PTP selectively retains more tokens from visually salient regions and further leverages textual instructions to pinpoint tokens most relevant to specific multimodal tasks. Extensive experiments across 13 diverse benchmarks demonstrate that our method substantially reduces computational overhead and inference latency with minimal performance loss.

HERO: Rethinking Visual Token Early Dropping in High-Resolution Large Vision-Language Models

By cropping high-resolution images into local tiles and encoding them independently, High-Resolution Large Vision-Language Models (HR-LVLMs) have demonstrated remarkable fine-grained visual understanding capabilities. However, this divide-and-conquer paradigm significantly increases the number of visual tokens, resulting in substantial computational and memory overhead. To better understand and address this challenge, we empirically investigate visual token utilization in HR-LVLMs and uncover three key findings: (1) the local tiles have varying importance, jointly determined by visual saliency and task relevance; (2) the CLS token in CLIP-based vision encoders exhibits a two-stage attention pattern across layers, with each stage attending to different types of visual tokens; (3) the visual tokens emphasized at different stages encode information at varying levels of granularity, playing complementary roles within LVLMs. Building on these insights, we propose HERO, a High-resolution visual token early dropping framework that integrates content-adaptive token budget allocation with function-aware token selection. By accurately estimating tile-level importance and selectively retaining visual tokens with complementary roles, HERO achieves superior efficiency-accuracy trade-offs across diverse benchmarks and model scales, all in a training-free manner. This study provides both empirical insights and practical solutions toward efficient inference in HR-LVLMs.

Global Semantic-Guided Sub-image Feature Weight Allocation in High-Resolution Large Vision-Language Models

As the demand for high-resolution image processing in Large Vision-Language Models (LVLMs) grows, sub-image partitioning has become a popular approach for mitigating visual information loss associated with fixed-resolution processing. However, existing partitioning methods uniformly process sub-images, resulting in suboptimal image understanding. In this work, we reveal that the sub-images with higher semantic relevance to the entire image encapsulate richer visual information for preserving the model's visual understanding ability. Therefore, we propose the Global Semantic-guided Weight Allocator (GSWA) module, which dynamically allocates weights to sub-images based on their relative information density, emulating human visual attention mechanisms. This approach enables the model to focus on more informative regions, overcoming the limitations of uniform treatment. We integrate GSWA into the InternVL2-2B framework to create SleighVL, a lightweight yet high-performing model. Extensive experiments demonstrate that SleighVL outperforms models with comparable parameters and remains competitive with larger models. Our work provides a promising direction for more efficient and contextually aware high-resolution image processing in LVLMs, advancing multimodal system development.

A Spatial-Temporal Dual-Mode Mixed Flow Network for Panoramic Video Salient Object Detection

Salient object detection (SOD) in panoramic video is still in the initial exploration stage. The indirect application of 2D video SOD method to the detection of salient objects in panoramic video has many unmet challenges, such as low detection accuracy, high model complexity, and poor generalization performance. To overcome these hurdles, we design an Inter-Layer Attention (ILA) module, an Inter-Layer weight (ILW) module, and a Bi-Modal Attention (BMA) module. Based on these modules, we propose a Spatial-Temporal Dual-Mode Mixed Flow Network (STDMMF-Net) that exploits the spatial flow of panoramic video and the corresponding optical flow for SOD. First, the ILA module calculates the attention between adjacent level features of consecutive frames of panoramic video to improve the accuracy of extracting salient object features from the spatial flow. Then, the ILW module quantifies the salient object information contained in the features of each level to improve the fusion efficiency of the features of each level in the mixed flow. Finally, the BMA module improves the detection accuracy of STDMMF-Net. A large number of subjective and objective experimental results testify that the proposed method demonstrates better detection accuracy than the state-of-the-art (SOTA) methods. Moreover, the comprehensive performance of the proposed method is better in terms of memory required for model inference, testing time, complexity, and generalization performance.