论文检索

Baoheng Zhang, Jiahui Liu, Gui Zhao, Weizhou Zhang, Yixuan Ma, Jun Jiang, Yingxian Chen, Wilton W.T. Fok, Xiaojuan Qi, Hayden Kwok-Hay So

Multimodal Large Language Models (MLLMs) perform strong vision-language reasoning under standard conditions but fail in extreme illumination, where RGB inputs lose irrevocable structure and semantics. We propose Event-MLLM, an event-enhanced model that performs all-light visual reasoning by dynamically fusing event streams with RGB frames. Two key components drive our approach: an Illumination Indicator -- a learnable signal derived from a DINOv2 branch that represents exposure degradation and adaptively modulates event-RGB fusion -- and an Illumination Correction Loss that aligns fused features with non-degraded (normal-light) semantics in the latent space, compensating for information lost in extreme lighting. We curate the first multi-illumination event-instruction corpus for MLLMs, with 2,241 event-RGB samples (around 6 QA pairs each) across diverse scenes and 17 brightness rates (0.05x - 20x), plus an instruct-following benchmark for reasoning, counting, and fine-grained recognition under extreme lighting. Experiments show that Event-MLLM markedly outperforms general-purpose, illumination-adaptive, and event-only baselines, setting a new state of the art in robust multimodal perception and reasoning under challenging illumination.

Qiankun Ma, Ziyao Zhang, Haofei Wang, Zhen Song, Jie Chen, Hairong Zheng

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 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.

Ruoxuan Zhang, Qiyun Zheng, Zhiyu Zhou, Ziqi Liao, Siyu Wu, Jian-Yu Jiang-Lin, Bin Wen, Hongxia Xie, Jianlong Fu, Wen-Huang Cheng

Theory of Mind (ToM) refers to the ability to infer others' mental states, such as beliefs, desires, and intentions. Current vision-language embodied agents lack ToM-based decision-making, and existing benchmarks focus solely on human mental states while ignoring the agent's own perspective, hindering coherent decision and action generation. To address this, we propose MindPower, a Robot-Centric framework integrating Perception, Mental Reasoning, Decision Making and Action. Given multimodal inputs, MindPower first perceives the environment and human states, then performs ToM Reasoning to model both self and others, and finally generates decisions and actions guided by inferred mental states. Furthermore, we introduce Mind-Reward, a novel optimization objective that encourages VLMs to produce consistent ToM Reasoning and behavior. Our model outperforms GPT-4o by 12.77% in decision making and 12.49% in action generation. Project page: https://zhangdaxia22.github.io/MindPower/.

Junhao Cheng, Liang Hou, Xin Tao, Jing Liao

While language models have become impactful in many real-world applications, video generation remains largely confined to entertainment. Motivated by video's inherent capacity to demonstrate physical-world information that is difficult to convey through language alone (e.g., imagine teaching someone to tie a tie using only text), we identify an underutilized opportunity to extend video as a new answer modality for Next-Event Prediction (NEP), formalized as Video-Next-Event Prediction (VNEP). While the established NEP task takes a video with a procedural or predictive question as input to predict the next event in text, VNEP requires dynamic video responses. This shift from telling to showing unlocks more intuitive and customized answers for procedural learning and creative exploration. However, this task remains challenging for existing models, as it demands an understanding of multimodal input, instruction-conditioned reasoning, and the generation of video with visual and semantic consistency. To address this, we introduce VANS, a model that leverages reinforcement learning to align a Vision-Language Model (VLM) with a Video Diffusion Model (VDM) for VNEP. The core of VANS is our proposed Joint-GRPO that orchestrates the VLM and VDM to function as a unit. Driven by a shared reward on their respective output, it optimizes the VLM to produce captions that are both accurate and friendly to visualize, while guiding the VDM to generate videos that are faithful to these captions and the input visual context. To enable this learning, we craft VANS-Data-100K, a dedicated dataset for the VNEP task. Experiments on procedural and predictive benchmarks demonstrate that VANS achieves state-of-the-art performance in both video event prediction and visualization.

Jiafei Song, Fengwei Zhou, Jin Qu, Wenjin Jason Li, Tong Wu, Gengjian Xue, Zhikang Zhao, Daomin Wei, Yichao Lu, Bailin Na

Recent Multimodal Large Language Models (MLLMs) have demonstrated strong performance on vision-language understanding tasks, yet their inference efficiency is often hampered by the large number of visual tokens, particularly in high-resolution or multi-image scenarios. To address this issue, we propose EvoComp, a visual token compression framework that significantly reduces token count while preserving task accuracy. EvoComp introduces a lightweight encoder-only transformer-based compressor that selects the most informative and non-redundant visual tokens by jointly considering visual and textual contexts. A core challenge lies in providing effective supervision for training the compressor. To this end, we design an evolutionary labeling strategy that searches for token subsets minimizing the MLLM's output loss, while enforcing semantic diversity through vocabulary-based token grouping. We further train the compressor using a tailored loss function combining the GHM loss to mitigate class and difficulty imbalance, and a cosine similarity regularization to encourage semantic separation between retained and discarded tokens. Extensive experiments across multiple vision-language benchmarks show that EvoComp outperforms existing methods based on attention or similarity heuristics. Notably, it retains 99.3% of the original accuracy under 3x token compression and delivers up to 1.6x speedup on mobile devices.

Aditya Kumar Singh, Hitesh Kandala, Pratik Prabhanjan Brahma, Zicheng Liu, Emad Barsoum

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.

Hao Wu, Xudong Wang, Jialiang Zhang, Junlong Tong, Xinghao Chen, Junyan Lin, Yunpu Ma, Xiaoyu Shen

One-stream Transformer-based trackers achieve advanced performance in visual object tracking suffer from significant computational overhead that hinders real-time deployment. While token pruning offers a path to efficiency, a critical limitation persists: no existing work performs pruning jointly across all three critical components--the search region, dynamic template, and static template. This isolation overlooks interdependencies, yielding suboptimal pruning and degraded accuracy. To address this, we introduce UTPTrack, a simple and Unified Token Pruning framework that, for the first time, jointly compresses all three components. UTPTrack employs an attention-guided, token type-aware strategy to holistically model redundancy, a design that seamlessly supports unified tracking across multi-modal and language-guided tasks within a single model. Comprehensive evaluations on 10 benchmarks demonstrate that UTPTrack achieves a new state-of-the-art in the accuracy-efficiency trade-off for pruning-based trackers, pruning 65.4% of vision tokens in RGB-based tracking and 67.5% in unified tracking while preserving 99.7% and 100.5% of baseline performance, respectively. This strong performance across both RGB and multimodal scenarios underlines its potential as a robust foundation for future research in efficient visual tracking.

Qihui Zhu, Tao Zhang, Yuchen Wang, Shuangwu Chen, Xiaobin Tan, Jian Yang, Yang Liu, Yinfei Pan

In multimodal large language models (MLLMs), the surge of visual tokens significantly increases the inference time and computational overhead, making them impractical for real-time or resource-constrained applications.Visual token pruning is a promising strategy for reducing the cost of MLLM inference by removing redundant visual tokens.Existing researches usually assume that all attention heads contribute equally to the visual interpretation.However, our study reveals that different heads may capture distinct visual semantics and inherently play distinct roles in visual processing.In light of this observation, we propose HAWK, a head importance-aware visual token pruning method that perceives the varying importance of attention heads in visual tasks to maximize the retention of crucial tokens.By leveraging head importance weights and text-guided attention to assess visual token significance, HAWK effectively retains task-relevant visual tokens while removing redundant ones.The proposed HAWK is entirely training-free and can be seamlessly applied to various MLLMs.Extensive experiments on multiple mainstream vision-language benchmarks demonstrate that HAWK achieves state-of-the-art accuracy. When applied to Qwen2.5-VL, HAWK retains 96.0% of the original accuracy after pruning 80.2% of the visual tokens. Additionally, it reduces end-to-end latency to 74.4% of the original and further decreases GPU memory usage across the tested models.

Keliang Liu, Zizhi Chen, Mingcheng Li, Jingqun Tang, Dingkang Yang, Lihua Zhang

Document understanding is a long-standing practical task. Vision-Language Models (VLMs) have gradually become a primary approach in this domain, demonstrating effective performance on single-page tasks. However, their effectiveness diminishes when handling long documents. In such scenarios, clues are often scattered across multiple pages and modalities, and redundancy from lengthy inputs can impair the model's judgment. While retrieval-augmented generation mitigates this issue by filtering for question-relevant content, the retrieved results still contain substantial redundancy. To address these limitations, we propose SLEUTH, a multi-agent framework. Concretely, SLEUTH orchestrates a retriever and four collaborative agents in a coarse-to-fine process. The framework identifies key textual and visual clues within the retrieved pages, filters for salient visual evidence such as tables and charts, and analyzes the query to devise a reasoning strategy. It ultimately synthesizes a distilled, evidence-dense multimodal context to generate the final prediction. SLEUTH is model-agnostic and scalable. When paired with advanced VLM backbones, it consistently improves performance on multiple long-document benchmarks, achieving SOTA results. Ablation studies verify each module's effectiveness and confirm the benefits of our hierarchical refinement paradigm.

Zhiwen Fan, Jian Zhang, Renjie Li, Junge Zhang, Runjin Chen, Hezhen Hu, Kevin Wang, Peihao Wang, Huaizhi Qu, Shijie Zhou 等

The rapid advancement of Large Multimodal Models (LMMs) for 2D images and videos has sparked interest in extending these models to 3D scenes, with the goal of human-like visual-spatial intelligence. However, achieving deep spatial understanding comparable to human capabilities remains challenging for both model design and data acquisition. Existing methods often rely on external depth sensors for geometry capture or off-the-shelf algorithms for pre-constructing 3D maps, which limits their scalability.In this work, we introduce VLM-3R, a framework for Vision-Language Models that couples 3D reconstructive instruction tuning with scalable training data curation and a new benchmark for temporal reasoning. Specifically, VLM-3R processes monocular video frames with a geometry encoder that derives implicit 3D tokens representing scene context (spatial tokens) and camera motion (view tokens). In parallel, we build a scalable data creation pipeline with over 200K 3D reconstructive instruction-tuning question-answer pairs. To evaluate temporal reasoning, we further introduce the Vision-Spatial-Temporal Intelligence benchmark (VSTI-Bench), which contains over 138.6K question-answer pairs across five distinct tasks focused on evolving spatial relationships. Extensive experiments show that VLM-3R supports robust visual-spatial reasoning and improves the understanding of temporal 3D context changes, enabling monocular 3D spatial assistance and embodied reasoning.

Tongtian Yue, Xuange Gao, Longteng Guo, Zijia Zhao, Zikang Liu, Jie Jiang, Hua Huang, Jing Liu

Recent advances in multimodal large language models (MLLMs) have shown impressive progress in integrating visual and linguistic understanding. However, most existing MLLMs inject visual information into the input space of large language models (LLMs), which substantially increases context length and computational overhead, while often disrupting pretrained linguistic priors by forcing the LLM to optimize on vision-dominant multimodal sequences. In this work, we propose a rotation-based vision injection paradigm that aligns visual information with the parameter space of LLMs. Visual semantics are encoded as rotation matrices and applied directly to the pretrained parameters. This parameter-space injection eliminates the need for long input sequences, thus avoiding the quadratic computational overhead inherent in input-space injection. Besides, it preserves the linguistic competence of the LLM by maintaining the intrinsic geometric structure of the pretrained parameters. Building upon this paradigm, we develop ROSE, a 7B MLLM that achieves fine-grained vision-language alignment with remarkable computational efficiency. Extensive experiments across 12 multimodal benchmarks show that ROSE delivers superior or competitive performance compared with leading models.At comparable accuracy, ROSE reduces FLOPs by 80.7% and inference latency by 56.4% relative to Qwen2.5-VL-7B, demonstrating the effectiveness and scalability. The code and model weights will be publicly released.

Guangzhi Xiong, Sanchit Sinha, Zhenghao He, Aidong Zhang

Vision-language models (VLMs) have achieved impressive performance across a wide range of multimodal reasoning tasks, but they often struggle to disentangle fine-grained visual attributes and reason about underlying causal relationships. In-context learning (ICL) offers a promising avenue for VLMs to adapt to new tasks, but its effectiveness critically depends on the selection of demonstration examples. Existing retrieval-augmented approaches typically rely on passive similarity-based retrieval, which tends to select correlated but non-causal examples, amplifying spurious associations and limiting model robustness. We introduce CIRCLES (Composed Image Retrieval for Causal Learning Example Selection), a novel framework that actively constructs demonstration sets by retrieving counterfactual-style examples through targeted, attribute-guided composed image retrieval. By incorporating counterfactual-style examples, CIRCLES enables VLMs to implicitly reason about the causal relations between attributes and outcomes, moving beyond superficial correlations and fostering more robust and grounded reasoning. Comprehensive experiments on four diverse datasets demonstrate that CIRCLES consistently outperforms existing methods across multiple architectures, especially on small-scale models, with pronounced gains under information scarcity. Furthermore, CIRCLES retrieves more diverse and causally informative examples, providing qualitative insights into how models leverage in-context demonstrations for improved reasoning. Our code is available at https://github.com/gzxiong/CIRCLES.

Enguang Wang, Qiang Wang, Yuanchen Wu, Ke Yan, Xinbin Yuan, Shouhong Ding, Xialei Liu, Ming-Ming Cheng

While Multimodal Large Language Models (MLLMs) excel at vision-language tasks, the cost of their language-driven training on internal visual foundational competence remains unclear. In this paper, we conduct a detailed diagnostic analysis to unveil a pervasive issue: visual representation degradation in MLLMs. Specifically, we find that compared to the initial visual features, the visual representation in the middle layers of LLM exhibits both a degradation in global function and patch structure. We attribute this phenomenon to a visual sacrifice driven by the singular text-generation objective, where the model compromises its visual fidelity to optimize for answer generation. We argue that a robust MLLM requires both strong cross-modal reasoning and core visual competence, and propose Predictive Regularization (PRe) to force degraded intermediate features to predict initial visual features, thereby maintaining the inherent visual attributes of the MLLM's internal representations. Extensive experiments confirm that mitigating this visual degradation effectively boosts vision-language performance, underscoring the critical importance of fostering robust internal visual representations within MLLMs for comprehensive multimodal understanding.

Mahesh Bhosale, Abdul Wasi, Shantam Srivastava, Shifa Latif, Tianyu Luan, Mingchen Gao, David Doermann, Xuan Gong

While powerful in image-conditioned generation, multimodal large language models (MLLMs) can display uneven performance across demographic groups, highlighting fairness risks. In safety-critical clinical settings, such disparities risk producing unequal diagnostic narratives and eroding trust in AI-assisted decision-making. While fairness has been studied extensively in vision-only and language-only models, its impact on MLLMs remains largely underexplored. To address these biases, we introduce FairLLaVA, a parameter-efficient fine-tuning method that mitigates group disparities in visual instruction tuning without compromising overall performance. By minimizing the mutual information between target attributes, FairLLaVA regularizes the model's representations to be demographic-invariant. The method can be incorporated as a lightweight plug-in, maintaining efficiency with low-rank adapter fine-tuning, and provides an architecture-agnostic approach to fair visual instruction following. Extensive experiments on large-scale chest radiology report generation and dermoscopy visual question answering benchmarks show that FairLLaVA consistently reduces inter-group disparities while improving both equity-scaled clinical performance and natural language generation quality across diverse medical imaging modalities. Code can be accessed at https://github.com/bhosalems/FairLLaVA.

Tengjin Weng, Wenhao Jiang, Jingyi Wang, Ming Li, Lin Ma, Zhong Ming

Multimodal large language models (MLLMs) have achieved remarkable performance across a wide range of vision-language tasks. However, their ability in low-level visual perception, particularly in detecting fine-grained visual discrepancies, remains underexplored and lacks systematic analysis.In this work, we introduce OddGridBench, a controllable benchmark for evaluating the visual discrepancy sensitivity of MLLMs. OddGridBench comprises over 1,400 grid-based images, where a single element differs from all others by one or multiple visual attributes such as color, size, rotation, or position.Experiments reveal that all evaluated MLLMs, including open-source families such as Qwen3-VL and InternVL3.5, and proprietary systems like Gemini-2.5-Pro and GPT-5, perform far below human levels in visual discrepancy detection. We further propose OddGrid-GRPO, a reinforcement learning framework that integrates curriculum learning and distance-aware reward. By progressively controlling the difficulty of training samples and incorporating spatial proximity constraints into the reward design, OddGrid-GRPO significantly enhances the model's fine-grained visual discrimination ability.We hope OddGridBench and OddGrid-GRPO will lay the groundwork for advancing perceptual grounding and visual discrepancy sensitivity in multimodal intelligence.All resources will be publicly released upon acceptance.

Linqing Zhong, Yi Liu, Yifei Wei, Ziyu Xiong, Si Liu, Guanghui Ren

Vision-Language-Action models have emerged as essential generalist robot policies for diverse manipulation tasks, conventionally relying on directly translating multimodal inputs into actions via Vision-Language Model embeddings. Recent advancements have introduced explicit intermediary reasoning--such as sub-task prediction (language) or goal image synthesis (vision)--to guide action generation. However, these intermediate reasoning are often indirect and inherently limited in their capacity to convey the full, granular information required for precise action execution. Instead, we posit that the most effective form of reasoning is one that deliberates directly in the action space. We introduce Action Chain-of-Thought (ACoT), a paradigm where the reasoning process itself is formulated as a structured sequence of coarse action intents that guide the final policy. In this paper, we propose ACoT-VLA, a novel architecture that materializes the ACoT paradigm. Specifically, we introduce two complementary components: an Explicit Action Reasoner (EAR) and Implicit Action Reasoner (IAR). The former proposes coarse reference trajectories as explicit action-level reasoning steps, while the latter extracts latent action priors from internal representations of multimodal input, co-forming an ACoT that conditions the downstream action head to enable grounded policy learning. Extensive experiments in real-world and simulation environments demonstrate the superiority of our proposed method. Code is available at: https://github.com/AgibotTech/ACoT-VLA.

Yun Zhu, Jianjun Qian, Jian Yang, Jin Xie, Na Zhao

Incremental 3D object perception is a critical step toward embodied intelligence in dynamic indoor environments. However, existing incremental 3D detection methods rely on extensive annotations of novel classes for satisfactory performance. To address this limitation, we propose FI3Det, a Few-shot Incremental 3D Detection framework that enables efficient 3D perception with only a few novel samples by leveraging vision-language models (VLMs) to learn knowledge of unseen categories. FI3Det introduces a VLM-guided unknown object learning module in the base stage to enhance perception of unseen categories. Specifically, it employs VLMs to mine unknown objects and extract comprehensive representations, including 2D semantic features and class-agnostic 3D bounding boxes. To mitigate noise in these representations, a weighting mechanism is further designed to re-weight the contributions of point- and box-level features based on their spatial locations and feature consistency within each box. Moreover, FI3Det proposes a gated multimodal prototype imprinting module, where category prototypes are constructed from aligned 2D semantic and 3D geometric features to compute classification scores, which are then fused via a multimodal gating mechanism for novel object detection. As the first framework for few-shot incremental 3D object detection, we establish both batch and sequential evaluation settings on two datasets, ScanNet V2 and SUN RGB-D, where FI3Det achieves strong and consistent improvements over baseline methods.

Jiangang Ding, Yiquan Du, Pengxiang Li, Lili Pei, Yuanlin Zhao, Wei Li

Bridging the modality gap between infrared and visible imagery is critical for cross-modal understanding and for enriching multimodal benchmarks. However, existing approaches remain confined to one-to-one mappings and are typically evaluated on unidirectional or closed-set scenarios. To address this challenge, we present SynthRGB-T, a unified framework for diverse and bidirectional image translation. Specifically, we formulate image translation as a vision-language guided denoising diffusion process, enabling flexible conditioning and open-world generalization. To enhance semantic alignment, a Visual Grounding Pipeline (VGP) is introduced to exploit the world knowledge of foundation models for fine-grained translation guidance. During the diffusion process, we propose to adopt a decoupling injection strategy to alleviate interference among multiple guidance. In addition, a Dual Conditional Cross-Attention (DCCA) module is designed to facilitate collaborative representation learning in latent space. SynthRGB-T is simple and versatile--capable of synthesizing diverse, high-fidelity data that substantially extends multimodal resources within the community. Comprehensive evaluations on multiple real-world benchmarks confirm that SynthRGB-T delivers superior performance and enhanced visual diversity over existing approaches.

Dachuan Zhao, Weiyue Li, Zhenda Shen, Yushu Qiu, Bowen Xu, Haoyu Chen, Yongchao Chen

Vision-Language Models (VLMs) have become indispensable for multimodal reasoning, yet their representations often encode and amplify demographic biases, resulting in biased associations and misaligned predictions in downstream tasks. Such behavior undermines fairness and distorts the intended alignment between vision and language. Recent post-hoc approaches attempt to mitigate bias by replacing the most attribute-correlated embedding coordinates with neutral values. However, our systematic analysis reveals three critical limitations of this coordinate-wise approach: feature entanglement, poor cross-dataset generalization, and incomplete bias removal. We find that bias is not localized to a few coordinates but is instead distributed across a few linear subspaces. To address these limitations, we propose Subspace Projection Debiasing (SPD), a geometrically principled framework that identifies and removes the entire subspace of linearly decodable bias while reinserting a neutral mean component to preserve semantic fidelity. Extensive experiments across zero-shot classification, text-to-image retrieval, and image generation validate the effectiveness of SPD: our method achieves more robust debiasing with an average improvement of 18.5% across four fairness metrics, while maintaining minimal loss in task performance compared to the best debiasing baseline.

Hyun Lee, Hyemin Jeong, Yejin Kim, Hyungwook Choi, Hyunsoo Cho, Soo Kyung Kim, Joonseok Lee

Modern multimodal large language models (MLLMs) typically keep the language model fixed and train a visual projector that maps the pixels into a sequence of tokens in its embedding space, so that images can be presented in essentially the same form as text. However, the language model has been optimized to operate on discrete, semantically meaningful tokens, while prevailing visual projectors transform an image into a long stream of continuous and highly correlated embeddings. This causes the visual tokens to behave differently from the word-like units that LLMs are originally trained to understand. We propose a novel Disentangled Visual Tokenization (DiVT) that clusters patch embeddings into coherent semantic units, so each token corresponds to a distinct visual concept instead of a rigid grid cell. DiVT further adapts its token budget to image complexity, providing an explicit accuracy compute trade-off modifying neither the vision encoder nor the language model. Across diverse multimodal benchmarks, DiVT matches or surpasses baselines with significantly fewer visual tokens, demonstrating robustness under limited token budgets, significantly reducing memory cost and latency while making visual inputs more compatible with LLMs. Our code is available at https://github.com/snuviplab/DiVT.