CatchUp: Depth-Aware Re-Admission of Visual Tokens for Efficient Large Vision-Language Models

Eunju Park; Uiwon Hwang

Home/Archive/arXiv

arXiv · 2610.23324

CatchUp: Depth-Aware Re-Admission of Visual Tokens for Efficient Large Vision-Language Models

Eunju Park·Uiwon Hwang

View original source ↗Read source document ↗Export metadata ↓

Abstract

Large Vision-Language Models (LVLMs) repeatedly process hundreds of visual tokens across decoder layers, incurring substantial computation and memory costs. Existing inference-time visual token pruning methods often irreversibly remove low-ranked tokens, although the set of visual tokens receiving high query-conditioned attention can change across decoder depth. Recoverable routing preserves skipped tokens for later reconsideration, but re-admits them with stale representations after bypassing intermediate transformations. We further find that replacing stale re-admitted states with their current-depth oracle states from a full-token reference trajectory substantially improves grounding performance. We therefore propose CatchUp, a training-free framework for depth-aware visual-token re-admission. CatchUp repeatedly re-evaluates query-conditioned visual relevance, processes only selected visual tokens through each decoder interval, and preserves deferred tokens for later re-admission. Upon re-admission, CatchUp approximates missed representation evolution using reference-to-current transition signals from semantically and spatially related visual tokens. Across three referring-expression grounding datasets and six multimodal benchmarks, CatchUp achieves the highest average relative retention in both benchmark groups at 64-, 128-, and 192-token budgets. At a 64-token budget with 88.9% fewer active visual tokens, CatchUp retains 92.17% of the full model's performance on average across six multimodal benchmarks while reducing end-to-end latency by 8.0%. Code is available at https://github.com/pej0918/CatchUp

Explore related subjects

cs.CV

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Eunju Park, Uiwon Hwang. 2026-10-16. CatchUp: Depth-Aware Re-Admission of Visual Tokens for Efficient Large Vision-Language Models. https://arxiv.org/abs/2610.23324

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

MLLM-Guided Autonomous Object Detection for Open-World Understanding

Object detection typically relies on predefined category vocabularies, while open-world and open-vocabulary approaches remain limited by generic unknown labels or externally specified detection concepts. We propose LAOD, an MLLM-Guided Autonomous Object Detection framework in which a Multimodal Large Language Model (MLLM) generates an image-specific detection vocabulary that is grounded by an open-vocabulary detector. This decomposition separates semantic concept discovery from spatial localization and removes the need to manually enumerate target categories at inference time. We evaluate LAOD using Class-Agnostic Average Precision (CAAP), Semantic Naming Average Precision (SNAP), and Semantic Localization Average Precision (SemLoc-AP), which measure localization, semantic naming, and their correct association, respectively. Experiments on COCO-Val, LVIS-minival, and COCO-OOD across multiple MLLMs and detectors demonstrate strong performance across common, long-tailed, and out-of-distribution objects, while showing that grounding quality remains a major determinant of end-to-end performance. LAOD also substantially outperforms both direct MLLM-based object detection and dedicated models trained specifically for open-ended detection. Overall, the results support a modular formulation of autonomous object detection in which MLLMs determine what to detect and specialized open-vocabulary detectors determine where those concepts occur.

cs.CV↗

arXiv2026

Shape Distribution Matters: Shape-Specific Mixture-of-Experts for Amodal Segmentation Under Diverse Occlusions

Amodal segmentation aims to predict complete object masks covering both visible and occluded regions, serving as a critical foundation for robust multimedia scene understanding and interactive media editing. This task poses significant challenges in complex visual media due to severe occlusions and extreme shape variation, ranging from rigid furniture to highly deformable clothing. Existing one-size-fits-all approaches struggle to reason about diverse amodal shapes in multimedia content due to limited representation capacity. While a mixture-of-experts framework is a natural solution, naively applying it without considering the underlying shape distribution leads to mismatched routing and redundant experts. To address these issues, we introduce ShapeMoE, a shape-specific sparse mixture-of-experts framework designed for complex visual scene parsing. The key idea is to learn a latent shape distribution space and dynamically route each object to a lightweight expert tailored to its geometric characteristics. Specifically, ShapeMoE encodes each object into a compact Gaussian embedding that captures key shape features. A Shape-Aware Sparse Router then maps the object to the most suitable expert, enabling precise shape-aware routing. Each expert is lightweight and specializes in recovering occluded regions for specific shape patterns. ShapeMoE is designed for shape-specific expert specialization while maintaining computational efficiency. Extensive experiments on three challenging benchmarks demonstrate that ShapeMoE consistently outperforms state-of-the-art methods. Beyond advancing fundamental visual reasoning, ShapeMoE provides a robust and efficient visual prior for downstream multimedia applications, particularly in occlusion-heavy immersive environments.

cs.CV↗

arXiv2026

Generative Spatiotemporal Data Augmentation

We explore spatiotemporal data augmentation using video foundation models to diversify both camera viewpoints and scene dynamics. Unlike existing approaches based on simple geometric transforms or appearance perturbations, our method leverages off-the-shelf video diffusion models to generate realistic 3D spatial and temporal variations from a given image dataset. Incorporating these synthesized video clips as supplemental training data yields consistent performance gains in low-data settings, such as UAV-captured imagery where annotations are scarce. Beyond empirical improvements, we provide practical guidelines for (i) choosing an appropriate spatiotemporal generative setup, (ii) transferring annotations to synthetic frames, and (iii) addressing disocclusion - regions newly revealed and unlabeled in generated views. Experiments on COCO subsets and UAV-captured datasets show that, when applied judiciously, spatiotemporal augmentation broadens the data distribution along axes underrepresented by traditional and prior generative methods, offering an effective lever for improving model performance in data-scarce regimes.

cs.CV↗