Efficient VLM Inference

VLM: Vision-Language Model

Follow this topic Ask about this topic

The field studies how to make vision-language models answer faster and use less memory and energy, especially for high-resolution images, long videos, and edge devices. Common approaches prune or merge visual tokens and frames, adapt resolution or retrieve regions on demand, reuse or compress KV caches, and use quantization, efficient kernels, routing, or speculative decoding. Methods are evaluated on tasks such as VQAv2, TextVQA, OCRBench, VideoMME, and LongVideoBench, measuring accuracy alongside time-to-first-token, throughput, and memory use. The central tension is preserving fine-grained evidence and reliable grounding while reducing computation: fewer tokens do not always mean lower end-to-end latency, and generating long answers can dominate cost on some devices.

Latest papers

487 papers66 papers in the last four weeks, up 50% on the four weeks before. 0.4% of all new papers.

Mar 11, 2026

Taking Shortcuts for Categorical VQA Using Super Neurons

Sparse Attention Vectors (SAVs) have emerged as an excellent training-free alternative to supervised finetuning or low-rank adaptation to improve the performance of Vision Language Models (VLMs). At their heart, SAVs select a few accurate attention heads for a task of interest and use them as classifiers, rather than relying on the model's prediction. In a similar spirit, we find that directly probing the raw activations of the VLM, in the form of scalar values, is sufficient to yield accurate classifiers on diverse visually grounded downstream tasks. Shifting focus from attention vectors to scalar activations dramatically increases the search space for accurate parameters, allowing us to find more discriminative neurons immediately from the first generated token. We call such activations Super Neurons (SNs). In this probing setting, we discover that enough SNs appear in the shallower layers of the large language model to allow for extreme early exiting from the first layer of the model at the first generated token. Compared to the original network, SNs robustly improve the classification performance while achieving a speedup of up to 5.10x.

Pierre Musacchio, Jaeyi Jeong, Dahun Kim +1

Seoul National University · EPFL · Google Deepmind

Efficient VLM Inference Visual Question Answering Representation Probing

Mar 11, 2026

Language-Guided Token Compression with Reinforcement Learning in Large Vision-Language Models

Large Vision-Language Models (LVLMs) incur substantial inference costs due to the processing of a vast number of visual tokens. Existing methods typically struggle to model progressive visual token reduction as a multi-step decision process with sequential dependencies and often rely on hand-engineered scoring rules that lack adaptive optimization for complex reasoning trajectories. To overcome these limitations, we propose TPRL, a reinforcement learning framework that learns adaptive pruning trajectories through language-guided sequential optimization tied directly to end-task performance. We formulate visual token pruning as a sequential decision process with explicit state transitions and employ a self-supervised autoencoder to compress visual tokens into a compact state representation for efficient policy learning. The pruning policy is initialized through learning from demonstrations and subsequently fine-tuned using Proximal Policy Optimization (PPO) to jointly optimize task accuracy and computational efficiency. Our experimental results demonstrate that TPRL removes up to 66.7% of visual tokens and achieves up to a 54.2% reduction in FLOPs during inference while maintaining a near-lossless average accuracy drop of only 0.7%. Code is released at mypink{https://github.com/MagicVicCoder/TPRL}.

Sihan Cao, Jianwei Zhang, Pengcheng Zheng +7

School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China · School of Robotics and Advanced Manufacture, Harbin Institute of Technology, Shenzhen, China · College of Information and Control Engineering, Xi’an University of Architecture and Technology, Xi’an, China

Efficient VLM Inference Visual Token Pruning LVLM Pruning Visual Token Compression RL for VLMs

Mar 10, 2026

Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity

Vision-language models (VLMs) face significant computational inefficiencies caused by excessive generation of visual tokens. While prior work shows that a large fraction of visual tokens are redundant, existing compression methods struggle to balance importance preservation and information diversity. To address this, we propose PruneSID, a training-free Synergistic Importance-Diversity approach featuring a two-stage pipeline: (1) Principal Semantic Components Analysis (PSCA) for clustering tokens into semantically coherent groups, ensuring comprehensive concept coverage, and (2) Intra-group Non-Maximum Suppression (NMS) for pruning redundant tokens while preserving key representative tokens within each group. Additionally, PruneSID incorporates an information-aware dynamic compression ratio mechanism that optimizes token compression rates based on image complexity, enabling more effective average information preservation across diverse scenes. Extensive experiments demonstrate state-of-the-art performance, achieving 96.3% accuracy on LLaVA-1.5 with only 11.1% token retention, and 92.8% accuracy at extreme compression rates (5.6%) on LLaVA-NeXT, outperforming prior methods by 2.5% with 7.8

\times

faster prefilling speed compared to the original model. Our framework generalizes across diverse VLMs and both image and video modalities, showcasing strong cross-modal versatility. Code is available at https://github.com/ZhengyaoFang/PruneSID.

Zhengyao Fang, Pengyuan Lyu, Chengquan Zhang +3

Harbin Institute of Technology, Shenzhen · Independent Researcher · Peng Cheng Laboratory

Visual Token Pruning Visual Token Compression Efficient VLM Inference Video Token Compression

Mar 7, 2026

The Model Knows Which Tokens Matter: Automatic Token Selection via Noise Gating

Visual tokens dominate inference cost in vision-language models (VLMs), yet many carry redundant information. Existing pruning methods alleviate this but typically rely on attention magnitude or similarity scores. We reformulate visual token pruning as capacity constrained communication: given a fixed budget K, the model must allocate limited bandwidth to maximally preserve visual information. We propose AutoSelect, which attaches a lightweight Scorer and Denoiser to a frozen VLM and trains with only the standard next token prediction loss, without auxiliary objectives or extra annotations. During training, a variance preserving noise gate modulates each token's information flow according to its predicted importance so that gradients propagate through all tokens; a diagonal attention Denoiser then recovers the perturbed representations. At inference, only the Scorer and a hard top-K selection remain, adding negligible latency. On ten VLM benchmarks, AutoSelect retains 96.5% of full model accuracy while accelerating LLM prefill by 2.85x with only 0.69 ms overhead, and transfers to different VLM backbones without architecture-specific tuning. Code is available at https://github.com/MedHK23/AutoSelect.

Landi He, Xiaoyu Yang, Lijian Xu

Shenzhen University of Advanced Technology

Efficient VLM Inference Visual Token Pruning Vision-Language Models Gated Attention

Mar 7, 2026

VLN-Cache: Enabling Token Caching for VLN Models with Visual/Semantic Dynamics Awareness

Vision-and-Language Navigation (VLN) increasingly relies on large vision-language models, but their inference cost conflicts with real-time deployment. Token caching is a promising training-free strategy that avoids redundant computation by reusing stable visual tokens across frames. However, existing methods assume a static camera and fixed semantic focus, assumptions that VLN fundamentally violates. We identify two failure modes: (1) visual dynamics, where viewpoint shift displaces token positions across frames, causing position-wise matching to pair misaligned content; (2) semantic dynamics, where token relevance shifts across task stages as navigation progresses, making cached states stale. We propose VLN-Cache, a visual-dynamic-aware and semantic-dynamic-aware caching framework that introduces view-aligned remapping to recover geometric correspondences and a task-relevance saliency filter to veto reuse at semantic transitions. A layer-adaptive entropy policy further balances the per-layer reuse budget. Experiments on the R2R-CE simulation benchmark show up to 1.52x speedup while maintaining competitive navigation success rates.

Zihao Zheng, Zhihao Mao, Xingyue Zhou +9

School of Computer Science, Peking University, Beijing, China. · School of Computer Science, China University of Geosciences (Wuhan), Wuhan, China. · School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. +1

Efficient VLM Inference Vision-Language Navigation

Mar 6, 2026

History-Conditioned Spatio-Temporal Visual Token Pruning for Efficient Vision-Language Navigation

Vision-Language Navigation (VLN) enables robots to follow natural-language instructions in visually grounded environments, serving as a key capability for embodied robotic systems. Recent Vision-Language-Action (VLA) models have demonstrated strong navigation performance, but their high computational cost introduces latency that limits real-time deployment. We propose a training-free spatio-temporal vision token pruning framework tailored to VLA-based VLN. We apply spatial token selection to the current view, alongside spatio-temporal compression for historical memories, enabling efficient long-horizon inference while reducing redundant computation. Leveraging attention-based token importance and query-guided spatio-temporal filtering, the proposed approach preserves navigation-relevant information without retraining or modifying pretrained models, allowing plug-and-play integration into existing VLA systems. Through experiments on standard VLN benchmarks, we confirm that our method significantly outperforms existing pruning strategies. It successfully preserves superior navigation accuracy under extreme pruning scenarios, all while maintaining the highly competitive inference efficiency. Real-world deployment on a Unitree Go2 quadruped robot further validates reliable and low-latency instruction-following navigation under practical robotic constraints. We hope this work helps bridge the gap between large-scale multimodal modeling and efficient, real-time embodied deployment in robotic navigation systems. Project Page: https://wqtwjt1996.github.io/publications/2026-vln.html

Qitong Wang, Yijun Liang, Ming Li +2

The authors are with the Department of Computer and Information Sciences at the University of Delaware, Newark, DE 19711, USA. · The authors are with the University of Maryland’s Department of Computer Science, College Park, MD 20742, USA. · The author is with the Mohamed bin Zayed University of Artificial Intelligence, Masdar City, Abu Dhabi 7909, UAE.

Efficient VLM Inference Token Pruning Vision-Language Navigation Vision-Language-Action Models Visual Token Pruning

Mar 6, 2026

HiPP-Prune: Hierarchical Preference-Conditioned Structured Pruning for Vision-Language Models

Pruning vision-language models (VLMs) for efficient deployment is challenging because compression can affect not only task utility but also visual grounding, often amplifying object hallucinations even at the same sparsity level. We present HiPP-Prune, a hierarchical preference-conditioned structured pruning framework that treats pruning as conditional resource allocation under multiple objectives. HiPP-Prune makes plan-level decisions: a single policy invocation outputs a global pruning blueprint by factorizing decisions into an overall sparsity budget and a layer-wise allocation, enabling queryable trade-offs via a user-specified preference vector. To account for VLM-specific failure modes, our policy state integrates a visual sensitivity signal derived from attention flow between vision tokens and language hidden states, discouraging over-pruning of vision-critical layers that facilitate cross-modal fusion. We optimize pruning plans with plan-level Group Relative Policy Optimization (GRPO) under a multi-objective return that combines task utility, hallucination robustness (POPE), compression, and a synaptic-flow-inspired stability proxy to reduce unproductive exploration in high-sparsity regimes. Experiments on LLaVA with POPE and ScienceQA demonstrate that HiPP-Prune discovers diverse non-dominated pruning plans and provides controllable robustness--utility trade-offs under matched sparsity budgets.

Lincen Bai, Hedi Tabia, Raul Santos-Rodriguez

Université Paris-Saclay · University of Bristol

LVLM Pruning Neural Network Pruning Efficient VLM Inference Structured Pruning Object Hallucination in VLMs

Mar 6, 2026

Energy-Driven Adaptive Visual Token Pruning for Efficient Vision-Language Models

Visual token reduction is critical for accelerating Vision-Language Models (VLMs), since visual inputs are represented as token sequences that introduce substantial computational overhead in the LLM backbone. However, most pruning pipelines treat efficiency primarily as a token selection problem and retain a fixed visual token budget across inputs, overlooking the substantial variation in image information density. We propose E-AdaPrune, an energy driven adaptive pruning framework that determines an image specific token budget from the singular value spectrum of the visual feature matrix and passes this budget to existing token selectors. By preserving a certain proportion of spectral energy, our method allocates more tokens to information dense scenes while assigning fewer tokens to redundant scenes, without introducing additional learnable parameters. We evaluate E-AdaPrune across four VLM backbones, three token selectors, and nine benchmarks under matched average token budgets. Results show that E-AdaPrune removes a substantial amount of redundant computation from simple cases and converts the saved budget into larger gains on information rich cases. Notably, on SQA

^\mathrm{I}

with Qwen2.5-VL-3B, E-AdaPrune uses 35.8% fewer tokens for simple cases with only a 0.52% relative performance decrease. The saved budget is redirected to hard cases, which receive 52.5% more tokens and achieve a 1.94% relative performance improvement.

Jialuo He, Huangxun Chen

The Hong Kong University of Science and Technology (Guangzhou)

Efficient VLM Inference Visual Token Pruning Large Vision-Language Models

Mar 4, 2026

FocusGraph: Graph-Structured Frame Selection for Embodied Long Video Question Answering

Understanding long videos is crucial for embodied intelligent agents, as their performance depends on effectively accumulating and using long-horizon perceptual memories. Multimodal large language models (MLLMs) are increasingly used for long-video understanding, but their performance degrades and inference time increases as more frames are provided. Therefore, selecting informative keyframes is essential for efficient question answering over long videos. In this work, we develop FocusGraph, a framework for keyframe selection in egocentric long-video question answering. It includes a lightweight Scene-Graph LLM Selector that identifies query-relevant clips from compact graph-based captions, avoiding the need to process raw frame sequences at question time. From these clips, we extract keyframes using Patch-wise Sparse-Flow Retention (PSFR), an offline program-evolved method with no learned parameters at inference time, before passing them to an MLLM for answer generation. FocusGraph achieves state-of-the-art performance on FindingDory and HourVideo while reducing question-time inference cost compared with existing approaches.

Tatiana Zemskova, Solomon Andryushenko, Ilya Obrubov +4

AXXX · MIRAI · Yandex +1

Efficient VLM Inference Egocentric Video QA Embodied QA Egocentric Video Understanding Video Frame Selection Long-Video QA+1

Mar 4, 2026

EvoPrune: Early-Stage Visual Token Pruning for Efficient MLLMs

Multimodal Large Language Models (MLLMs) have shown strong performance in vision-language tasks, but their inference efficiency is severely limited by the exponential growth of visual tokens in complex scenarios such as high-resolution images and videos. Existing visual token pruning methods mainly operate after visual encoding, overlooking the substantial computational cost incurred during the encoding stage. To address this issue, we propose EvoPrune, an early-stage visual token pruning method for MLLMs that performs pruning directly during visual encoding. Specifically, EvoPrune employs a layer-wise pruning strategy guided by token similarity, diversity, and attention-based importance to retain the most informative visual tokens at selected encoding layers. Extensive experiments on image and video benchmarks validate the effectiveness of EvoPrune. In particular, on the VideoMME dataset, EvoPrune achieves 2

\times

inference speedup with less than 1% performance degradation, demonstrating its potential for latency-sensitive MLLM deployment.

Yuhao Chen, Bin Shan, Xin Ye +1

ByteDance, Beijing, China

Efficient VLM Inference Visual Token Pruning Large Vision-Language Models

Mar 3, 2026

Benchmarking Compact VLMs for Clip-Level Surveillance Anomaly Detection Under Weak Supervision

CCTV safety monitoring demands anomaly detectors combine reliable clip-level accuracy with predictable per-clip latency despite weak supervision. This work investigates compact vision-language models (VLMs) as practical detectors for this regime. A unified evaluation protocol standardizes preprocessing, prompting, dataset splits, metrics, and runtime settings to compare parameter-efficiently adapted compact VLMs against training-free VLM pipelines and weakly supervised baselines. Evaluation spans accuracy, precision, recall, F1, ROC-AUC, and average per-clip latency to jointly quantify detection quality and efficiency. With parameter-efficient adaptation, compact VLMs achieve performance on par with, and in several cases exceeding, established approaches while retaining competitive per-clip latency. Adaptation further reduces prompt sensitivity, producing more consistent behavior across prompt regimes under the shared protocol. These results show that parameter-efficient fine-tuning enables compact VLMs to serve as dependable clip-level anomaly detectors, yielding a favorable accuracy-efficiency trade-off within a transparent and consistent experimental setup.

Kirill Borodin, Kirill Kondrashov, Nikita Vasiliev +4

Faculty of Information Technology, Moscow Technical University of Communication and Informatics, Moscow 111024, Russia

Vision-Language Models Weakly Supervised Video Anomaly Detection VLM Adaptation Efficient VLM Inference Parameter-Efficient Fine-Tuning

Mar 3, 2026

Think-as-You-See: Streaming Chain-of-Thought Reasoning for Large Vision-Language Models

Large Vision Language Models (LVLMs) exhibit strong Chain-of-Thought (CoT) capabilities, yet most existing paradigms assume full-video availability before inference, a batch-style process misaligned with real-world video streams where information arrives sequentially. Motivated by the streaming nature of video data, we investigate two streaming reasoning paradigms for LVLMs. The first, an interleaved paradigm, alternates between receiving frames and producing partial reasoning but remains constrained by strictly ordered cache updates. To better match streaming inputs, we propose Think-as-You-See (TaYS), a unified framework enabling true concurrent reasoning. TaYS integrates parallelized CoT generation, stream-constrained training, and stream-parallel inference. It further employs temporally aligned reasoning units, streaming attention masks and positional encodings, and a dual KV-cache that decouples visual encoding from textual reasoning. We evaluate all paradigms on the Qwen2.5-VL family across representative video CoT tasks, including event dynamics analysis, causal reasoning, and thematic understanding. Experiments show that TaYS consistently outperforms both batch and interleaved baselines, improving reasoning performance while substantially reducing time-to-first-token (TTFT) and overall reasoning delay. These results demonstrate the effectiveness of data-aligned streaming reasoning in enabling efficient and responsive video understanding for LVLMs. We release our code at https://github.com/EIT-NLP/StreamingLLM/tree/main/TaYS

Jialiang Zhang, Junlong Tong, Junyan Lin +4

Institute of Digital Twin, Eastern Institute of Technology, Ningbo · Ocean University of China · Shanghai Jiao Tong University +3

VLM Reasoning Streaming Inference Streaming Video Understanding Large Vision-Language Models Efficient VLM Inference

Feb 27, 2026

Stateful Token Reduction for Long-Video Hybrid VLMs

Token reduction accelerates long-video vision--language models (VLMs), but existing methods target Transformers, where reduction is treated as token pruning. We study token reduction in hybrid Mamba--Transformer VLMs and find that it is stateful: Mamba layers maintain a recurrent state that accumulates information from earlier tokens, allowing discarded tokens to persist, so reduction behaves more like compression than dropping.We support this view with a representation-based probing method measuring how much information from discarded tokens is retained, and analyze layer-wise sparsity and cross-layer importance stability. Our findings show importance is sparse within layers but unstable across layers, making aggressive early pruning unreliable while hybrids remain robust to later reduction.Motivated by this, we propose a hybrid-aware token reduction framework with a low-to-high progressive schedule and a unified query-conditioned importance score for attention and Mamba layers. For Mamba, excluding the position-dependent decay from the recurrence produces a stronger selection signal. Across long-video benchmarks, our method achieves

3.8{\times}

--

4.2{\times}

prefilling speedups at a 25% token budget while maintaining near-baseline accuracy and improving with light finetuning. Hybrid models benefit from aggressive reduction, improving both efficiency and accuracy, whereas Transformers exhibit the standard trade-off. Our method also outperforms prior baselines on the same hybrid backbone and combines effectively with visual redundancy reduction methods.

Jindong Jiang, Amala Sanjay Deshmukh, Kateryna Chumachenko +7

NVIDIA

Hybrid Mamba-Transformer Architectures Efficient VLM Inference Video-Language Models Multimodal Token Compression Visual Token Compression

Feb 27, 2026

HiDrop: Hierarchical Vision Token Reduction in MLLMs via Late Injection, Concave Pyramid Pruning, and Early Exit

The quadratic computational cost of processing vision tokens in Multimodal Large Language Models (MLLMs) hinders their widespread adoption. While progressive vision token pruning offers a promising solution, current methods misinterpret shallow layer functions and use rigid schedules, which fail to unlock the full efficiency potential. To address these issues, we propose HiDrop, a framework that aligns token pruning with the true hierarchical function of MLLM layers. HiDrop features two key innovations: (1) Late Injection, which bypasses passive shallow layers to introduce visual tokens exactly where active fusion begins; and (2) Concave Pyramid Pruning with an Early Exit mechanism to dynamically adjust pruning rates across middle and deep layers. This process is optimized via an inter-layer similarity measure and a differentiable top-k operator. To ensure practical efficiency, HiDrop further incorporates persistent positional encoding, FlashAttention-compatible token selection, and parallel decoupling of vision computation to eliminate hidden overhead associated with dynamic token reduction. Extensive experiments show that HiDrop compresses about 90% visual tokens while matching the original performance and accelerating training by 1.72 times. Our work not only sets a new state-of-the-art for efficient MLLM training and inference but also provides valuable insights into the hierarchical nature of multimodal fusion. The code is released at https://github.com/EIT-NLP/HiDrop.

Hao Wu, Yingqi Fan, Jinyang Dai +3

Institute of Digital Twin, Eastern Institute of Technology, Ningbo · Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative · University of Science and Technology of China +2

Efficient VLM Inference Large Vision-Language Models Multimodal Token Compression Visual Token Pruning

Feb 26, 2026

RT-NeuS: Towards Real-Time Neuro-Symbolic Video Understanding via Adaptive Temporal Verification

Long-form video question answering (LVQA) requires answering natural-language queries about videos spanning minutes to hours, demanding temporal reasoning across thousands of frames. Standard vision-language models (VLMs) struggle with this task: their fixed frame budgets force aggressive downsampling that misses the temporal structure that complex queries depend on. Neuro-symbolic approaches address this by decomposing queries into atomic propositions, translating them into temporal logic specifications, and applying formal model checking to retrieve segments satisfying the specification. This yields up to 10% higher accuracy on temporally complex benchmarks, with interpretability and formal guarantees. However, constructing the video automaton requires grounding every proposition at every frame window via VLM calls, resulting in up to 130x slower inference than standard VLM prompting. We present RT-NeuS, a framework that preserves the accuracy and formal guarantees of temporal-logic-guided LVQA while closing this latency gap. RT-NeuS introduces coarse-to-fine adaptive sampling to identify the small set of query-relevant, visually distinct frames, and batched proposition detection with KV-cache reuse to evaluate all propositions per window in a single forward pass. We derive latency upper bounds as a function of video length, proposition count, and sampling density. Experiments on LongVideoBench, Video-MME, and MLVU reduce inference latency by up to 13x on a single NVIDIA H200 GPU, while matching or exceeding prior neuro-symbolic accuracy.

Shawn Liang, Sahil Shah, Chengwei Zhou +5

Case Western Reserve University · The University of Texas at Austin

Long-Video QA Temporal Video Understanding Neuro-Symbolic Reasoning Efficient VLM Inference Query-Aware Video Frame Selection

Feb 26, 2026

Spatio-Temporal Token Pruning for Efficient High-Resolution GUI Agents

Pure-vision GUI agents provide universal interaction capabilities but suffer from severe efficiency bottlenecks due to the massive spatiotemporal redundancy inherent in high-resolution screenshots and historical trajectories. We identify two critical misalignments in existing compression paradigms: the temporal mismatch, where uniform history encoding diverges from the agent's "fading memory" attention pattern, and the spatial topology conflict, where unstructured pruning compromises the grid integrity required for precise coordinate grounding, inducing spatial hallucinations. To address these challenges, we introduce GUIPruner, a training-free framework tailored for high-resolution GUI navigation. It synergizes Temporal-Adaptive Resolution (TAR), which eliminates historical redundancy via decay-based resizing, and Stratified Structure-aware Pruning (SSP), which prioritizes interactive foregrounds and semantic anchors while safeguarding global layout. Extensive evaluations across diverse benchmarks demonstrate that GUIPruner consistently achieves state-of-the-art performance, effectively preventing the collapse observed in large-scale models under high compression. Notably, on Qwen2-VL-2B, our method delivers a 3.4x reduction in FLOPs and a 3.3x speedup in vision encoding latency while retaining over 94% of the original performance, enabling real-time, high-precision navigation with minimal resource consumption.

Zhou Xu, Bowen Zhou, Qi Wang +2

Tsinghua University, Shenzhen, China · Xidian University, Xi’an, China · The Chinese University of Hong Kong, Hong Kong, China

GUI Agents Visual Token Pruning Visual Token Compression Efficient VLM Inference

Feb 26, 2026

LookWise: Knowing When and Where to Look for Fine-Grained Visual Reasoning in Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) are shifting towards "Thinking with Images" by actively exploring image details. While effective, large-scale training is computationally expensive, which has spurred growing interest in lightweight, training-free solutions. However, existing training-free methods suffer from two flaws: perceptual redundancy from indiscriminate cropping, which increases computational cost and introduces noise; and a drift between semantic intent and spatial attention, which prevents accurate localization of user-focused regions. To address these challenges, we propose LookWise, a framework for adaptive visual reasoning. LookWise follows a two-stage pipeline: a confidence-based module decides when to look more carefully, and a semantic-guided localization module determines where to look. This design enables MLLMs to adaptively acquire fine-grained visual evidence without additional training. Experiments on fine-grained and high-resolution visual reasoning benchmarks show that LookWise consistently improves accuracy over strong baselines while achieving an approximately

4.0\times

inference speedup over the search-based method ZoomEye, demonstrating robust cross-model generalization.

Yuxiang Shen, Hailong Huang, Zhenkun Gao +6

the Institute of Intelligent Machines, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei 230031, China, and also with the University of Science and Technology of China, Hefei 230026, China · Zhejiang University, Hangzhou 310027, China · East China Normal University, Shanghai 200241, China +3

VLM Reasoning Large Vision-Language Models Fine-Grained Visual Perception Efficient VLM Inference

Feb 24, 2026

TinyVLM: Zero-Shot Object Detection on Microcontrollers via Vision-Language Distillation with Matryoshka Embeddings

Zero-shot object detection enables recognising novel objects without task-specific training, but current approaches rely on large vision language models (VLMs) like CLIP that require hundreds of megabytes of memory - far exceeding the constraints of micro controller units (MCUs). We present TinyVLM, the first framework enabling zero-shot object detection on resource-constrained MCUs with less than 1MB of memory. Our approach introduces three key innovations: (1) a decoupled architecture that separates visual inference from text encoding, allowing precomputed class embeddings to be stored in flash memory; (2) Matryoshka distillation that trains nested embeddings at multiple dimensions (16-256), enabling flexible accuracy-memory trade-offs; and (3) quantized embedding storage that reduces class prototype memory by 4x with minimal accuracy loss. Trained on Conceptual Captions 3M (CC3M), TinyVLM achieves competitive zero-shot accuracy on COCO, Flowers102, and Food101 while requiring only 285KB of RAM and 892KB of flash memory for the deployed vision encoder. We demonstrate real-time inference at 26 FPS on STM32H7 and over 1,000 FPS on MAX78000 with its CNN accelerator, enabling practical zero-shot detection on edge devices for the first time.

Bibin Wilson

Independent Researcher

Object Detection Open-Vocabulary Object Detection Efficient VLM Inference Tiny ML VLM Distillation

Feb 23, 2026

ApET: Approximation-Error Guided Token Compression for Efficient VLMs

Recent Vision-Language Models (VLMs) have demonstrated remarkable multimodal understanding capabilities, yet the redundant visual tokens incur prohibitive computational overhead and degrade inference efficiency. Prior studies typically relies on [CLS] attention or text-vision cross-attention to identify and discard redundant visual tokens. Despite promising results, such solutions are prone to introduce positional bias and, more critically, are incompatible with efficient attention kernels such as FlashAttention, limiting their practical deployment for VLM acceleration. In this paper, we step away from attention dependencies and revisit visual token compression from an information-theoretic perspective, aiming to maximally preserve visual information without any attention involvement. We present ApET, an Approximation-Error guided Token compression framework. ApET first reconstructs the original visual tokens with a small set of basis tokens via linear approximation, then leverages the approximation error to identify and drop the least informative tokens. Extensive experiments across multiple VLMs and benchmarks demonstrate that ApET retains 95.2% of the original performance on image-understanding tasks and even attains 100.4% on video-understanding tasks, while compressing the token budgets by 88.9% and 87.5%, respectively. Thanks to its attention-free design, ApET seamlessly integrates with FlashAttention, enabling further inference acceleration and making VLM deployment more practical. Code is available at https://github.com/MaQianKun0/ApET.

Qiankun Ma, Ziyao Zhang, Haofei Wang +3

Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Peng Cheng Laboratory · University of Chinese Academy of Sciences +2

Visual Token Compression Efficient VLM Inference Visual Token Pruning Video Token Compression

Feb 23, 2026

QuickGrasp: Responsive Video-Language Querying Service via Accelerated Tokenization and Edge-Augmented Inference

Video-language models (VLMs) are reshaping video querying services, bringing unified solutions to complex perception and reasoning tasks. However, deploying large VLMs in real-world systems remains challenging due to their high resource demands, and remote-based deployment often results in unacceptable response delays. Although small, locally deployable VLMs offer faster responses, they unavoidably fall short in accuracy. To reconcile this trade-off, we propose QuickGrasp, a responsive, quality of service (QoS)-aware system that bridges this gap through a local-first architecture with on-demand edge augmentation. Built upon the highly modular architecture of VLMs, QuickGrasp shares the vision representation across model variants to avoid redundant computation. To maximize system-wide efficiency, QuickGrasp introduces three key designs: accelerated video tokenization, query-adaptive edge augmentation, and delay-aware, accuracy-preserving vision token density configuration. We implement a prototype of QuickGrasp and evaluate it across multiple video understanding benchmarks. The results show that QuickGrasp matches the accuracy of large VLMs while achieving up to a 12.8x reduction in response delay. QuickGrasp represents a key advancement toward building responsive video querying services for open-world understanding that fully leverage the capabilities of VLMs.

Miao Zhang, Ruixiao Zhang, Jianxin Shi +3

School of Computing Science, Simon Fraser University, Canada · Computer Science Department, University of Illinois Urbana-Champaign, United States · College of Cryptology and Cyber Science, Nankai University, China +1

Efficient VLM Inference VLM Inference Video-Language Models Visual Tokenization

Feb 22, 2026

Seeing Farther and Smarter: Value-Guided Multi-Path Reflection for VLM Policy Optimization

Solving complex, long-horizon robotic manipulation tasks requires a deep understanding of physical interactions, reasoning about their long-term consequences, and precise high-level planning. Vision-Language Models (VLMs) offer a general perceive-reason-act framework for this goal. However, previous approaches using reflective planning to guide VLMs in correcting actions encounter significant limitations. These methods rely on inefficient and often inaccurate implicit learning of state-values from noisy foresight predictions, evaluate only a single greedy future, and suffer from substantial inference latency. To address these limitations, we propose a novel test-time computation framework that decouples state evaluation from action generation. This provides a more direct and fine-grained supervisory signal for robust decision-making. Our method explicitly models the advantage of an action plan, quantified by its reduction in distance to the goal, and uses a scalable critic to estimate. To address the stochastic nature of single-trajectory evaluation, we employ beam search to explore multiple future paths and aggregate them during decoding to model their expected long-term returns, leading to more robust action generation. Additionally, we introduce a lightweight, confidence-based trigger that allows for early exit when direct predictions are reliable, invoking reflection only when necessary. Extensive experiments on diverse, unseen multi-stage robotic manipulation tasks demonstrate a 24.6% improvement in success rate over state-of-the-art baselines, while significantly reducing inference time by 56.5%.

Yanting Yang, Shenyuan Gao, Qingwen Bu +2

Rutgers University · The Hong Kong University of Science and Technology · The University of Hong Kong

VLMs for Robotics Large Language Model-Based Robot Planning Long-Horizon Robotic Manipulation Efficient VLM Inference

Feb 22, 2026

OTPrune: Distribution-Aligned Visual Token Pruning via Optimal Transport

Multi-modal large language models (MLLMs) achieve strong visual-language reasoning but suffer from high inference cost due to redundant visual tokens. Recent work explores visual token pruning to accelerate inference, while existing pruning methods overlook the underlying distributional structure of visual representations. We propose OTPrune, a training-free framework that formulates pruning as distribution alignment via optimal transport (OT). By minimizing the 2-Wasserstein distance between the full and pruned token distributions, OTPrune preserves both local diversity and global representativeness while reducing inference cost. Moreover, we derive a tractable submodular objective that enables efficient optimization, and theoretically prove its monotonicity and submodularity, providing a principled foundation for stable and efficient pruning. We further provide a comprehensive analysis that explains how distributional alignment contributes to stable and semantically faithful pruning. Comprehensive experiments on wider benchmarks demonstrate that OTPrune achieves superior performance-efficiency tradeoffs compared to state-of-the-art methods. The code is available at https://github.com/xiwenc1/OTPrune.

Xiwen Chen, Wenhui Zhu, Gen Li +9

Morgan Stanley · Clemson University · Arizona State University +4

Visual Token Pruning Efficient VLM Inference LVLM Pruning

Feb 21, 2026

DUET-VLM: Dual stage Unified Efficient Token reduction for VLM Training and Inference

Vision-language models (VLMs) have achieved remarkable multimodal understanding and reasoning capabilities, yet remain computationally expensive due to dense visual tokenization. Existing efficiency approaches either merge redundant visual tokens or drop them progressively in language backbone, often trading accuracy for speed. In this work, we propose DUET-VLM, a versatile plug-and-play dual compression framework that consists of (a) vision-only redundancy aware compression of vision encoder's output into information-preserving tokens, followed by (b) layer-wise, salient text-guided dropping of visual tokens within the language backbone to progressively prune less informative tokens. This coordinated token management enables aggressive compression while retaining critical semantics. On LLaVA-1.5-7B, our approach maintains over 99% of baseline accuracy with 67% fewer tokens, and still retains >97% even at 89% reduction. With this dual-stage compression during training, it achieves 99.7% accuracy at 67% and 97.6% at 89%, surpassing prior SoTA visual token reduction methods across multiple benchmarks. When integrated into Video-LLaVA-7B, it even surpasses the baseline -- achieving >100% accuracy with a substantial 53.1% token reduction and retaining 97.6% accuracy under an extreme 93.4% setting. These results highlight end-to-end training with DUET-VLM, enabling robust adaptation to reduced visual (image/video) input without sacrificing accuracy, producing compact yet semantically rich representations within the same computational budget. Our code is available at https://github.com/AMD-AGI/DUET-VLM.

Aditya Kumar Singh, Hitesh Kandala, Pratik Prabhanjan Brahma +2

Advanced Micro Devices, Inc. (AMD)

Efficient VLM Inference Multimodal Token Compression Visual Token Compression Visual Token Pruning Video Token Compression

Feb 18, 2026

Designing Production-Scale OCR for India: Multilingual and Domain-Specific Systems

Designing Optical Character Recognition (OCR) systems for India requires balancing linguistic diversity, document heterogeneity, and deployment constraints. In this paper, we study two training strategies for building multilingual OCR systems with Vision-Language Models through the Chitrapathak series. We first follow a popular multimodal approach, pairing a generic vision encoder with a strong multilingual language model and training the system end-to-end for OCR. Alternatively, we explore fine-tuning an existing OCR model, despite not being trained for the target languages. Through extensive evaluation on multilingual Indic OCR benchmarks and deployment-oriented metrics, we find that the second strategy consistently achieves better accuracy-latency trade-offs. Chitrapathak-2 achieves 3-6x speedup over its predecessor with being state-of-the-art (SOTA) in Telugu (6.69 char ANLS) and second best in the rest. In addition, we present Parichay, an independent OCR model series designed specifically for 9 Indian government documents to extract structured key fields, achieving 89.8% Exact Match score with a faster inference. Together, these systems achieve SOTA performance and provide practical guidance for building production-scale OCR pipelines in the Indian context.

Ali Faraz, Raja Kolla, Ashish Kulkarni +1

Krutrim AI, Bangalore, India

Multilingual OCR Optical Character Recognition Document Information Extraction Low-Resource Language Processing Efficient VLM Inference

Feb 13, 2026

HSD: Training-Free Acceleration for Document Parsing Vision-Language Models with Hierarchical Speculative Decoding

Document parsing is a fundamental task in multimodal understanding, supporting a wide range of downstream applications such as information extraction and intelligent document analysis. Benefiting from strong semantic modeling and robust generalization, VLM-based end-to-end approaches have emerged as the mainstream paradigm in recent years. However, these models often suffer from substantial inference latency, as they must autoregressively generate long, full-page sequences when processing long-form documents. While recent hybrid methods mitigate this issue via region-level parallel decoding with VLMs, independent region decoding loses full-page context and might weaken global coherence. To address this issue, we propose Hierarchical Speculative Decoding (HSD), a two-stage local-to-global framework for document parsing. HSD first employs a lightweight pipeline drafter to predict region partitions and generate coarse drafts for each region. The first stage verifies the generated region-level drafts in parallel for efficiency, while the second stage further performs page-level verification on these refined outputs to preserve full-page coherence. Experimental results show that HSD achieves a near-lossless 2.7x speedup with HunyuanOCR on OmniDocBench v1.5 and up to 7.04x speedup on long-document parsing tasks, demonstrating the effectiveness of the proposed method. The code is available at https://github.com/whlscut/HSD.

Wenhui Liao, Hongliang Li, Pengyu Xie +15

South China University of Technology, Guangzhou, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China · Nanjing University, Nanjing, China +1

Efficient VLM Inference Speculative Decoding Document Parsing Document Image Understanding

Feb 6, 2026

Think Proprioceptively: State-Grounded Visual Token Selection for VLA Policies

Vision-language-action (VLA) models typically inject proprioception only as a late conditioning signal, preventing robot state from grounding instruction understanding or directing visual attention. We introduce ThinkProprio, which discretizes proprioception into VLM-vocabulary tokens and uses them jointly with the instruction to gate visual patches before VLM computation, steering the model toward action-relevant evidence while discarding redundant tokens early. We find that proprioception added as a passive conditioning signal leaves performance essentially unchanged; its value emerges when token-form state acts as an active query that, with the instruction, selects which visual patches the VLM processes. Systematic ablations show that VLM-vocabulary tokens outperform learned projectors as the state encoding, and that retaining only about \SI{12}{\percent} of the visual tokens surpasses on CALVIN ABC

\to

D. Across CALVIN, LIBERO, and real-world manipulation, ThinkProprio reduces end-to-end inference latency while improving the matched full-token baseline.

Fangyuan Wang, Peng Zhou, Jiaming Qi +4

The Hong Kong Polytechnic University · Great Bay University · Northeast Forestry University +2

Efficient VLM Inference Visuomotor Policy Learning Efficient VLA Models Vision-Language-Action Models VLMs for Robotics Visual Token Pruning

Feb 6, 2026

SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs

Despite recent successes, test-time scaling -- i.e., dynamically expanding the token budget during inference as needed -- remains brittle for vision-language models (VLMs). Unstructured visual reasoning chains entangle perception and reasoning, leading to long, disorganized contexts where small perceptual mistakes may cascade into completely wrong answers. Reasoning also requires expensive reinforcement learning with hand-crafted rewards. Here, we introduce SPARC (Separating Perception And Reasoning Circuits), a modular framework that explicitly decouples visual perception from reasoning. Inspired by sequential sensory-to-cognitive processing in the brain, SPARC implements a two-stage pipeline where the model first performs explicit visual search to localize question-relevant regions, then conditions its reasoning on those regions to produce the final answer. This separation enables independent test-time scaling with asymmetric compute allocation (e.g., prioritizing perceptual processing under distribution shift), and supports selective optimization (e.g., improving the perceptual stage alone when it is the bottleneck for end-to-end performance). It also accommodates compressed contexts by running global search at lower image resolutions and allocating high-resolution processing only to selected regions, thereby reducing visual token count and compute. SPARC outperforms monolithic baselines and strong visual-grounding approaches across challenging visual reasoning tasks, such as improving Qwen3VL 4B on the

V^*

VQA benchmark by 6.7 points and surpassing "thinking with images" by 4.6 points in an OOD setting with a

200\times

lower token budget.

Niccolo Avogaro, Nayanika Debnath, Li Mi +6

ETH Zürich · IBM Research · MIT-IBM Watson AI Lab.

Visual Question Answering Efficient VLM Inference Test-Time Scaling for VLMs Large Vision-Language Models VLM Reasoning Visually Grounded Reasoning+1

Jan 30, 2026

Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs

Vision-Language Models (VLMs) achieve strong multimodal performance but are costly to deploy, and post-training quantization often causes significant accuracy loss. Despite its potential, quantization-aware training for VLMs remains underexplored. We propose GRACE, a framework unifying knowledge distillation and QAT under the Information Bottleneck principle: quantization constrains information capacity while distillation guides what to preserve within this budget. Treating the teacher as a proxy for task-relevant information, we introduce confidence-gated decoupled distillation to filter unreliable supervision, relational centered kernel alignment to transfer visual token structures, and an adaptive controller via Lagrangian relaxation to balance fidelity against capacity constraints. Across extensive benchmarks on LLaVA and Qwen families, our INT4 models consistently outperform FP16 baselines (e.g., LLaVA-1.5-7B: 70.1 vs. 66.8 on SQA; Qwen2-VL-2B: 76.9 vs. 72.6 on MMBench), nearly matching teacher performance. Using real INT4 kernel, we achieve 3

\times

throughput with 54% memory reduction. This principled framework significantly outperforms existing quantization methods, making GRACE a compelling solution for resource-constrained deployment. Code and data are available at: https://github.com/ForeverBlue816/GRACE.

Yanlong Chen, Amirhossein Habibian, Luca Benini +1

Department of Information Technology and Electrical Engineering, ETH Zurich, Zurich, Switzerland · School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore · Qualcomm AI Research, Amsterdam, the Netherlands +1

Cross-Modal Knowledge Distillation Vision-Language Models VLM Quantization Quantization-Aware Training Efficient VLM Inference VLM Distillation+1

Jan 29, 2026

DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation

Manipulating dynamic objects remains an open challenge for Vision-Language-Action (VLA) models. Although recent VLAs generalize well in static manipulation, dynamic scenes introduce a latency-induced perception-execution mismatch: object states continue to evolve during inference, making actions predicted from past observations stale at execution time. We present DynamicVLA, a latency-aware VLA model for dynamic object manipulation. It combines a compact 0.4B architecture and convolutional vision encoder for efficient multimodal inference with a continuous inference schedule that overlaps reasoning and execution for non-blocking control. Latent-aware Action Streaming then discards latency-invalid action prefixes and executes only the temporally valid suffix of each predicted chunk, preserving action-time alignment under dynamic object motion. To fill the missing foundation of dynamic manipulation data, we introduce the Dynamic Object Manipulation (DOM) benchmark, built with an automated collection pipeline that gathers 200K synthetic episodes across 2.8K scenes and 206 objects, and enables fast collection of 2K real-world episodes without teleoperation. Extensive evaluations in simulation and on real robots show that DynamicVLA improves dynamic manipulation success under changing object motion, perception-heavy instructions, and unseen motion patterns.

Haozhe Xie, Beichen Wen, Jiarui Zheng +4

Inference-Time Optimization Efficient VLM Inference Efficient VLA Model Inference Vision-Language-Action Models Robot Manipulation Benchmarks

Jan 28, 2026

Youtu-Parsing: Perception, Structuring and Recognition via High-Parallelism Decoding

This paper presents Youtu-Parsing, an efficient and versatile document parsing model designed for high-performance content extraction. The architecture employs a native Vision Transformer (ViT) featuring a dynamic-resolution visual encoder to extract shared document features, coupled with a prompt-guided Youtu-LLM-2B language model for layout analysis and region-prompted decoding. Leveraging this decoupled and feature-reusable framework, we introduce a high-parallelism decoding strategy comprising two core components: token parallelism and query parallelism. The token parallelism strategy concurrently generates up to 64 candidate tokens per inference step, which are subsequently validated through a verification mechanism. This approach yields a 5--11x speedup over traditional autoregressive decoding and is particularly well-suited for highly structured scenarios, such as table recognition. To further exploit the advantages of region-prompted decoding, the query parallelism strategy enables simultaneous content prediction for multiple bounding boxes (up to five), providing an additional 2x acceleration while maintaining output quality equivalent to standard decoding. Youtu-Parsing encompasses a diverse range of document elements, including text, formulas, tables, charts, seals, and hierarchical structures. Furthermore, the model exhibits strong robustness when handling rare characters, multilingual text, and handwritten content. Extensive evaluations demonstrate that Youtu-Parsing achieves state-of-the-art (SOTA) performance on both the OmniDocBench and olmOCR-bench benchmarks. Overall, Youtu-Parsing demonstrates significant experimental value and practical utility for large-scale document intelligence applications.

Haoyu Cao, Kun Yin, Yunfei Wu +16

Youtu-Parsing Team*

Efficient VLM Inference Document Parsing Parallel Decoding Document Image Understanding