论文检索

Jiwoo Ha, Jongwoo Baek, Jinhyun So

Recent Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across various multimodal tasks that require understanding both visual and linguistic inputs. However, object hallucination -- the generation of nonexistent objects in answers -- remains a persistent challenge. Although several approaches such as retraining and external grounding methods have been proposed to mitigate this issue, they still suffer from high data costs or structural complexity. Training-free methods such as Contrastive Decoding (CD) are more cost-effective, avoiding additional training or external models, but still suffer from long-term decay, where visual grounding weakens and language priors dominate as the generation progresses. In this paper, we propose First Logit Boosting (FLB), a simple yet effective training-free technique designed to alleviate long-term decay in LVLMs. FLB stores the logit of the first generated token and adds it to subsequent token predictions, effectively mitigating long-term decay of visual information. We observe that FLB (1) sustains the visual information embedded in the first token throughout generation, and (2) suppresses hallucinated words through the stabilizing effect of the "The" token. Experimental results show that FLB significantly reduces object hallucination across various tasks, benchmarks, and backbone models. Notably, it causes negligible inference overhead, making it highly applicable to real-time multimodal systems. Code is available at https://github. com/jiwooha20/FLB

Jingchen Sun, Shaobo Han, Deep Patel, Wataru Kohno, Can Jin, Changyou Chen

Knowledge distillation establishes a learning paradigm that leverages both data supervision and teacher guidance. However, determining the optimal balance between learning from data and learning from the teacher is challenging, as some samples may be noisy while others are subject to teacher uncertainty. This motivates the need for adaptively balancing data and teacher supervision. We propose Beta-weighted Knowledge Distillation (Beta-KD), an uncertainty-aware distillation framework that adaptively modulates how much the student relies on teacher guidance. Specifically, we formulate teacher-student learning from a unified Bayesian perspective and interpret teacher supervision as a Gibbs prior over student activations. This yields a closed-form, uncertainty-aware weighting mechanism and supports arbitrary distillation objectives and their combinations. Extensive experiments are conducted on multimodal VQA benchmarks by distilling a student Vision-Language Model from a large teacher VLM. The results demonstrate that Beta-KD consistently outperforms existing knowledge distillation methods. Code is available at https://github.com/Jingchensun/beta-kd.

Zeyuan Yang, Xueyang Yu, Delin Chen, Maohao Shen, Chuang Gan

Vision-language models (VLMs) excel at multimodal understanding, yet their text-only decoding forces them to verbalize visual reasoning, limiting performance on tasks that demand visual imagination. Recent attempts train VLMs to render explicit images, but the heavy image-generation pre-training often hinders the reasoning ability. Inspired by the way humans reason with mental imagery--the internal construction and manipulation of visual cues--we investigate whether VLMs can reason through interleaved multimodal trajectories without producing explicit images. To this end, we present a Machine Mental Imagery framework, dubbed as "Mirage", which augments VLM decoding with latent visual tokens alongside ordinary text. Concretely, whenever the model chooses to "think visually", it recasts its hidden states as next tokens, thereby continuing a multimodal trajectory without generating pixel-level images. Begin by supervising the latent tokens through distillation from ground-truth image embeddings, we then switch to text-only supervision to make the latent trajectory align tightly with the task objective. A subsequent reinforcement learning stage further enhances the multimodal reasoning capability. Experiments on diverse benchmarks demonstrate that \Model unlocks stronger multimodal reasoning without explicit image generation.

Shengyuan Ding, Xinyu Fang, Ziyu Liu, Yuhang Zang, Yuhang Cao, Xiangyu Zhao, Haodong Duan, Xiaoyi Dong, Jianze Liang, Bin Wang 等

Reward models are critical for aligning vision-language systems with human preferences, yet current approaches suffer from hallucination, weak visual grounding, and an inability to use tools for verification, limiting their reliability on complex multimodal reasoning tasks.We present **ARM-Thinker**, an **A**gentic multimodal **R**eward **M**odel that autonomously invokes external tools (e.g., image cropping, doc page retrieval) to ground judgments in verifiable evidence, replacing static, non-interactive reward scoring.This enables the model to verify fine-grained visual details, cross-reference multi-page evidence, and validate reasoning claims, which are capabilities absent in existing reward models.We train ARM-Thinker with multi-stage reinforcement learning, jointly optimizing tool-calling decisions and judgment accuracy.To evaluate agentic reward modeling, we introduce **ARMBench-VL**, comprising three benchmarks that assess fine-grained visual grounding (image-level tools), multi-page document understanding (retrieval tools), and instruction following (text-level verification).ARM-Thinker achieves +16.2% average improvement on reward modeling benchmarks, +9.6% on tool-use tasks, and outperforms baselines on multimodal math and logical reasoning benchmarks.Our results demonstrate that agentic capabilities significantly enhance both accuracy and interpretability of reward models.

Zelai Xu, Zhexuan Xu, Xiangmin Yi, Huining Yuan, Mo Guang, Kaiwen Long, Xinlei Chen, Yi Wu, Chao Yu, Yu Wang

Recent advancements in Vision Language Models (VLMs) have expanded their capabilities to interactive agent tasks, yet existing benchmarks remain limited to single-agent or text-only environments. In contrast, real-world scenarios often involve multiple agents interacting within rich visual and textual contexts, posing challenges with both multimodal observations and strategic interactions. To bridge this gap, we introduce Visual Strategic Bench (VS-Bench), a multimodal benchmark that evaluates VLMs for strategic abilities in multi-agent environments. VS-Bench comprises ten vision-grounded environments that cover cooperative, competitive, and mixed-motive interactions. The performance of VLM agents is evaluated across three dimensions: perception measured by element recognition accuracy; strategic reasoning measured by next-action prediction accuracy; and decision-making measured by normalized episode return. Extensive experiments on fifteen leading VLMs show that, although current models exhibit strong perception abilities, there remains a significant gap to optimal performance in reasoning and decision-making, with the best-performing model attaining 46.6% prediction accuracy and 31.4% normalized return. We further analyze the key factors influencing performance, conduct human studies, and examine failure modes to provide a deeper understanding of VLMs' strategic abilities. By standardizing the evaluation and highlighting the limitations of existing models, we envision VS-Bench as a foundation for future research on strategic multimodal agents.

Muhammad Kamran Janjua, Hugo Silva, Di Niu, Bahador Rashidi

Multimodal language models (MLLMs) are increasingly paired with vision tools (e.g., depth, flow, correspondence) to enhance visual reasoning. However, despite access to these tool-generated visual cues, MLLMs often fail to benefit fully from them. Existing approaches typically feed raw tool outputs into the model, but these dense, pixel-level representations are misaligned with the language-native reasoning strengths of LLMs, leading to weak perception due to reliance on language priors. We argue that, in problems where vision tools can provide the necessary visual cues, the bottleneck is not more tool calls or larger MLLMs, it is how tool outputs are represented. We introduce Perception Programs (P^2), a training-free, model-agnostic method that rewrites tool outputs into compact, structured, language-native summaries that MLLMs can directly parse and reason over. Across six perception-centric tasks in BLINK, P^2 consistently yields large improvements over base models and raw tool-augmented baselines. With GPT-5 Mini as the base model, P^2 raises its accuracy from 41.35% to 86.47% on multi-view reasoning, from 52.42 to 81.45 on relative depth, and achieves a 19.66 overall average gain across tasks, setting new state-of-the-art results. Even on smaller MLLMs, e.g., InternVL3.5-4B and Qwen3VL-4B, we observe 21-25 absolute gains from P^2 across tasks when comparing to image-based tool variants, surpassing prior agentic, supervised, and RL-based tool-use methods, without any training or model modifications.

Enhui Ma, Jiahuan Zhang, Guantian Zheng, Tao Tang, Shengbo Eben Li, Yuhang Lu, Xia Zhou, Xueyang Zhang, Yifei Zhan, Kun Zhan 等

Multimodal Large Language Models (MLLMs) are rapidly becoming the intelligence brain of end-to-end autonomous driving systems. A key challenge is to assess whether MLLMs can truly understand and follow complex real-world traffic rules. However, existing benchmarks mainly focus on single-rule scenarios like traffic sign recognition, neglecting the complexity of multi-rule concurrency and conflicts in real driving. Consequently, models perform well on simple tasks but often fail or violate rules in real world complex situations. To bridge this gap, we propose DriveCombo, a text and vision-based benchmark for compositional traffic rule reasoning. Inspired by human drivers' cognitive development, we propose a systematic Five-Level Cognitive Ladder that evaluates reasoning from single-rule understanding to multi-rule integration and conflict resolution, enabling quantitative assessment across cognitive stages. We further propose a Rule2Scene Agent that maps language-based traffic rules to dynamic driving scenes through rule crafting and scene generation, enabling scene-level traffic rule visual reasoning. Evaluations of 14 mainstream MLLMs reveal a pronounced decline in performance as task difficulty increases, especially in the rule conflict resolution task. After splitting the dataset and fine-tuning on the training set, we further observe substantial improvements in both traffic rule reasoning and downstream planning capabilities. These results highlight the effectiveness of DriveCombo in advancing compliant and intelligent autonomous driving systems.

Jing Bi, Filippos Bellos, Junjia Guo, Yayuan Li, Chao Huang, Yunlong Tang, Luchuan Song, Susan Liang, Zhongfei Zhang, Jason J. Corso 等

Test-time "thinking" (i.e., generating explicit intermediate reasoning chains) is known to boost performance in large language models and has recently shown strong gains for large vision-language models (LVLMs). However, despite these promising results, there is still no systematic analysis of how thinking actually affects visual reasoning. We provide the first such analysis with a large-scale, controlled comparison of thinking for LVLMs, evaluating 10 variants from the InternVL3.5 and Qwen3-VL families on MMMUval under generous token budgets and multi-pass decoding. We show that more thinking is not always better: long chains often yield long-wrong trajectories that ignore the image and underperform the same models run in standard instruct mode. A deeper analysis reveals that certain short "lookback" phrases, which explicitly refer back to the image, are strongly enriched in successful trajectories and correlate with better visual grounding. Building on this insight, we propose uncertainty-guided lookback, a training-free decoding strategy that combines an uncertainty signal with adaptive lookback prompts and breadth search. Our method improves overall MMMU performance, delivers the largest gains in categories where standard thinking is weak, and outperforms several strong decoding baselines, setting a new state of the art under fixed model families and token budgets. We further show that this decoding strategy generalizes, yielding consistent improvements on five additional benchmarks, including two broad multimodal suites and math-focused visual reasoning datasets.

Qixun Wang, Yang Shi, Yifei Wang, Yuanxing Zhang, Pengfei Wan, Kun Gai, Xianghua Ying, Yisen Wang

Thinking with images has emerged as an effective paradigm for advancing visual reasoning, extending beyond text-only chains of thought by injecting visual evidence into intermediate reasoning steps. However, existing methods fall short of human-like abstract visual thinking, as their flexibility is fundamentally limited by external tools. In this work, we introduce Monet, a training framework that enables multimodal large language models (MLLMs) to reason directly within the latent visual space by generating continuous embeddings that function as intermediate visual thoughts. We identify two core challenges in training MLLMs for latent visual reasoning--high computational cost in latent-vision alignment and insufficient supervision over latent embeddings, and address them with a three-stage distillation-based supervised fine-tuning (SFT) pipeline. We further reveal a limitation of applying GRPO to latent reasoning: it primarily enhances text-based reasoning rather than latent reasoning. To overcome this, we propose VLPO (Visual-latent Policy Optimization), a reinforcement learning method that explicitly incorporates latent embeddings into policy gradient updates.To support SFT, we construct Monet-SFT-125K, a high-quality text-image interleaved CoT dataset containing 125K real-world, chart, OCR, and geometry CoTs. Our model, Monet-7B, shows consistent gains across real-world perception and reasoning benchmarks and exhibits strong out-of-distribution generalization on challenging abstract visual reasoning tasks. We also empirically analyze the role of each training component and discuss our early unsuccessful attempts, providing insights for future developments in visual latent reasoning.

Jiangtong Tan, Lin Liu, Jie Huang, Xiaopeng Zhang, Qi Tian, Feng Zhao

Unified multimodal models significantly improve visual generation by combining vision-language models (VLMs) with diffusion models. However, existing methods struggle to fully balance sufficient interaction and flexible implementation due to vast representation difference. Considering abundant and hierarchical information in VLM's layers from low-level details to high-level semantics, we propose ParaUni. It extracts features from variants VLM's layers in a Parallel way for comprehensive information interaction and retains a flexible separation architecture to enhance generation in Unified multimodal model. Concretely, visual features from all VLM's layers are fed in parallel into a Layer Integration Module (LIM), which efficiently integrates fine-grained details and semantic abstractions and provides the fused representation as a condition to the diffusion model. To further enhance performance, we reveal that these hierarchical layers respond unequally to different rewards in Reinforcement Learning (RL). Crucially, we design a Layer-wise Dynamic Adjustment Mechanism (LDAM) to facilitate multiple reward improvements that aligns the hierarchical properties of these layers using RL. Extensive experiments show ParaUni leverages complementary multi-layer features to substantially improve generation quality and shows strong potential for multiple reward advances during RL stages.

Shreshth Saini, Bowen Chen, Yilin Wang, Neil Birkbeck, Balu Adsumilli, Alan C. Bovik

High Dynamic Range (HDR) user-generated (UGC) videos are rapidly proliferating across social platforms, yet most perceptual video quality assessment (VQA) systems remain tailored to Standard Dynamic Range (SDR). HDR's higher bit depth, wide color gamut, and elevated luminance range expose distortions such as near-black crushing, highlight clipping, banding, and exposure flicker that amplify UGC artifacts and challenge SDR models. To catalyze progress, we curate HDR-UGC-44K, a large-scale subjective dataset of ~44K videos from 6.5K sources with >1.5M crowd ratings, spanning diverse scenes, capture conditions, and compression settings. We further introduce HDR-Q, the first Multimodal Large Language Model (MLLM) for HDR-UGC VQA. We propose (i) a novel HDR-aware vision encoder to produce HDR-sensitive embeddings, and (ii) HDR-Aware Policy Optimization (HAPO), an RL finetuning framework that anchors reasoning to HDR cues. HAPO augments GRPO via an HDR-SDR contrastive KL that encourages token reliance on HDR inputs and a gaussian weighted regression reward for fine-grained MOS calibration. Across HDR-UGC-44K and public HDR-VQA benchmarks, HDR-Q delivers state-of-the-art performance. Project page: https://shreshthsaini.github.io/Beynod8Bits

Wei Xiang, Yexinrui Wu, Xinli Chen, Xinran Li, Shi Chen

User Interface (UI) display defect detection poses challenges far beyond UI understanding, requiring fine-grained element boundary understanding, missing-content detection, and reasoning about sequential interface semantic consistency. However, the capabilities of multimodal large language models (MLLMs) and vision-language models (VLMs) for detecting UI defects in realistic, complex interfaces have not been systematically validated. To fill this gap, we present UI-Lens, the first multi-dimensional UI display detection benchmark for Chinese-language UI scenarios. The dataset comprises 4,759 pages meticulously annotated by design experts, covering six core display defect categories. We conduct a systematic evaluation of 10 mainstream models (8 closed-source, 2 open-source). Results show clear shortcomings in current models: for tasks requiring fine-grained element boundary understanding, performance is near random, with task-average F1 scores of 20.36% and 31.21% on Text Overflow and Container Overlap, respectively; for sequential interface semantic consistency (e.g., Text Inconsistency), the task-average F1 score is only 10.61%, indicating severe underperformance. We release UI-Lens to catalyze research toward more robust UI display defect detection with fine-grained boundary awareness in realistic, complex interfaces.

Yuchen Feng, Zhenyu Zhang, Naibin Gu, Yilong Chen, Peng Fu, Zheng Lin, Shuohuan Wang, Yu Sun, Hua Wu, Weiping Wang 等

Multimodal large language models (MLLMs) have achieved remarkable progress on various vision-language tasks, yet their visual perception remains limited. Humans, in comparison, perceive complex scenes efficiently by dynamically scanning and focusing on salient regions in a sequential "blink-like" process. Motivated by this strategy, we first investigate whether MLLMs exhibit similar behavior. Our pilot analysis reveals that MLLMs naturally attend to different visual regions across layers and that selectively allocating more computation to salient tokens can enhance visual perception. Building on this insight, we propose Blink, a dynamic visual token resolution framework that emulates the human-inspired process within a single forward pass. Specifically, Blink includes two modules: saliency-guided scanning and dynamic token resolution. It first estimates the saliency of visual tokens in each layer based on the attention map, and extends important tokens through a plug-and-play token super-resolution (TokenSR) module. In the next layer, it drops the extended tokens when they lose focus. This dynamic mechanism balances broad exploration and fine-grained focus, thereby enhancing visual perception adaptively and efficiently. Extensive experiments validate Blink, demonstrating its effectiveness in enhancing visual perception and multimodal understanding.

Zirun Guo, Minjie Hong, Feng Zhang, Kai Jia, Tao Jin

Multimodal large language models (MLLMs) that "think with images" can interactively use tools to reason about visual inputs, but current approaches often rely on a narrow set of tools with limited real-world necessity and scalability. In this work, we first reveal a critical and previously overlooked weakness: even state-of-the-art MLLMs are surprisingly brittle, showing significant performance degradation on images with simple orientation changes, underscoring the need for more robust tool-based reasoning. To address this, we propose CodeVision, a flexible and scalable "code-as-tool" framework where the model generates code as a universal interface to invoke any image operation, moving beyond fixed tool registries. We train our model using a two-stage methodology, beginning with Supervised Fine-Tuning (SFT) on a high-quality dataset curated for complex, multi-turn tool composition and error recovery, followed by Reinforcement Learning (RL) with a dense process reward function to encourage strategic and efficient tool use. To facilitate this research, we construct new SFT and RL datasets and introduce a challenging new benchmark suite designed to rigorously evaluate robustness to orientation changes and multi-tool reasoning. Experiments on Qwen2.5-VL and Qwen3-VL series show that our approach significantly improves model performance and fosters emergent capabilities such as flexible tool composition, efficient chained execution, and robust error recovery from runtime feedback. Code is available at https://github.com/ByteDance-BandAI/CodeVision.

Yasmine Omri, Connor Ding, Tsachy Weissman, Thierry Tambe

Modern vision-language pipelines are driven by RGB vision encoders trained on massive image-text corpora. While these pipelines have enabled impressive zero-shot capabilities and strong transfer across tasks, they still inherit two structural inefficiencies from the pixel domain: (i) transmitting dense RGB images from edge devices to the cloud is energy-intensive and costly, and (ii) patch-based tokenization explodes sequence length, stressing attention budgets and context limits. We explore 2D Gaussian Splatting (2DGS) as an alternative visual substrate for alignment: a compact, spatially adaptive representation that parameterizes images by a set of colored anisotropic Gaussians. We develop a scalable 2DGS pipeline with structured initialization, luminance-aware pruning, and batched CUDA kernels, achieving over 90x faster fitting and 97% GPU utilization compared to prior implementations. We further adapt contrastive language-image pre-training (CLIP) to 2DGS by reusing a frozen RGB-based transformer backbone with a lightweight splat-aware input stem and a perceiver resampler, training only 9.7 - 13.8% of the total parameters. On a 12.8M dataset from DataComp, GS encoders yield competitive zero-shot performance on 38 datasets from the CLIP benchmark while compressing inputs 3-23.5x relative to pixels. When integrated into a full end-to-end LLaVA-style visual-language model (VLM), GS encoders consistently outperform the RGB VLM baseline across six VQA benchmarks, often with reduced token counts, demonstrating that a transmission-efficient, structured Gaussian representation can match or surpass pixel-based models in downstream multimodal reasoning. Our codebase can be found at https://github.com/ Tambe-Lab/GaussianVision.

Minh Dinh, SouYoung Jin

Large-scale image datasets frequently contain identifiable or sensitive content, raising privacy risks when training models that may memorize and leak such information. We present Unsafe2Safe, a fully automated pipeline that detects privacy-prone images and rewrites only their sensitive regions using multimodally guided diffusion editing. Unsafe2Safe operates in two stages. Stage 1 uses a vision-language model to (i) inspect images for privacy risks, (ii) generate paired private and public captions that respectively include and omit sensitive attributes, and (iii) prompt a large language model to produce structured, identity-neutral edit instructions conditioned on the public caption. Stage 2 employs instruction-driven diffusion editors to apply these dual textual prompts, producing privacy-safe images that preserve global structure and task-relevant semantics while neutralizing private content. To measure anonymization quality, we introduce a unified evaluation suite covering Quality, Cheating, Privacy, and Utility dimensions. Across MS-COCO, Caltech101 and MIT Indoor67, Unsafe2Safe reduces face similarity, text similarity, and demographic predictability by large margins, while maintaining downstream model accuracy comparable to training on raw data. Fine-tuning diffusion editors on our automatically generated triplets (private caption, public caption, edit instruction) further improves both privacy protection and semantic fidelity. Unsafe2Safe provides a scalable, principled solution for constructing large, privacy-safe datasets without sacrificing visual consistency or downstream utility.

Jiarui Wang, Huiyu Duan, Juntong Wang, Xiongkuo Min

As generative models rapidly evolve, the realism of AI-generated videos has reached new levels, posing significant challenges for detecting the authenticity of videos. Existing deepfake detection techniques generally rely on training datasets with limited generation methods and content diversity, which limits their generalization ability on more realistic content, particularly that produced by the latest generative models. Recently, large multimodal models (LMMs) have demonstrated remarkable zero-shot performance across a variety of vision tasks. Yet, their ability to discern deepfake videos remains largely untested. To this end, we propose **FVBench**, a comprehensive deep\underline f ake \underline v ideo \underline bench mark designed to advance video deepfake detection. It includes: (i) extensive content diversity, with over 120K videos covering real, AI-edited, and fully AI-generated categories, (ii) comprehensive model coverage, with fake videos generated and edited by 42 of the state-of-the-art video synthesis and editing models, and (iii) deepfake video detection benchmark for LMMs, which is a comprehensive benchmark for exploring the deepfake video detection capabilities of LMMs. The dataset and code are released at https://github.com/IntMeGroup/FVBench.

Ruosen Zhao, Zhikang Zhang, Jialei Xu, Jiahao Chang, Dong Chen, Lingyun Li, Weijian Sun, Zizhuang Wei

Large vision-language models (VLMs) show strong multimodal understanding but still struggle with 3D spatial reasoning, such as distance estimation, size comparison, and cross-view consistency. Existing 3D-aware methods either depend on auxiliary 3D information or enhance RGB-only VLMs with geometry encoders through shallow feature fusion. We propose SpaceMind, a multimodal large language model trained on SpaceMind-700K for spatial reasoning from RGB inputs. The model adopts a dual-encoder architecture integrating VGGT as a spatial understanding encoder and InternViT as a 2D visual encoder. The key idea is to treat the camera representation as an active guiding modality rather than passive metadata. Specifically, SpaceMind introduces a lightweight Camera-Guided Modality Fusion module before the language model to replace shallow fusion. It applies camera-conditioned biasing to spatial tokens, assigns query-independent weights reflecting their geometric importance, and uses the camera embedding to gate the fused representation. Empirically, SpaceMind establishes new state-of-the-art results on VSI-Bench, SQA3D and SPBench, surpassing both open and proprietary systems on VSI-Bench and SPBench by large margins and achieving state-of-the-art performance on SQA3D. These results demonstrate that camera-guided modality fusion is an effective and practical inductive bias for equipping VLMs with genuinely spatially grounded intelligence. We will release code and model checkpoints to support future research.

Rui Wu, Shuo Zhang, Xiaoxuan Tang, Ruirui Zhang, Yi Liu, Tao Jiang, Wenhao Xu, Yong Li

Multimodal Web Search Agents demonstrate a practically valuable capability by fusing information from diverse modalities (e.g., text and vision), retrieved iteratively from the internet, to address complex user queries. However, the visual modality is prone to information overload, and the noise contained within it--such as irrelevant background details or complex structures--can disrupt the model's attention, misdirecting its operational focus toward an erroneous path. To address the aforementioned challenge, we propose ReFAct (Reasoning, Focusing, and Acting), a novel framework that empowers the agent to actively manage its cross-modal context. This allows the agent to adjust its operational focus, thereby mitigating the impact of noise on multimodal Web Search Agents. Specifically, ReFAct employs a Grounding tool for active visual perception to dynamically filter information. We also design external memory-based Defocus/Refocus operations for selective information retention, further modulating information density within the multimodal context. Ultimately, this ensures the agent maintains focus during problem-solving. To evaluate and enhance agent capabilities in complex and noisy multimodal contexts, we first propose a pipeline for constructing datasets with flexible complexity. We introduce a new open-source benchmark: GroundedVQA. Finally, we experimentally demonstrate the effectiveness of our proposed method on GroundedVQA and other widely-used benchmarks.

Ruiying Peng, Xueyu Wu, Jing Lei, Lu Hou, Yuanzheng Ma, Xiao-Hui Li

Multimodal large language models (MLLMs) often suffer from perceptual impairments under extended reasoning modes, particularly in visual question answering (VQA) tasks. We identify attention dispersion as the underlying cause: during multi-step reasoning, model's visual attention becomes scattered and drifts away from question-relevant regions, effectively "losing focus" on the visual input. To better understand this phenomenon, we analyze the attention maps of MLLMs and observe that reasoning prompts significantly reduce attention to regions critical for answering the question. We further find a strong correlation between model's overall attention on image tokens and the spatial dispersiveness of model's attention within the image. Leveraging this insight, we propose a training-free Visual Region-Guided Attention (VRGA) framework that selects visual heads based on an entropy-focus criterion and reweights their attention, effectively guiding the model to focus on question-relevant regions during reasoning. Extensive experiments on vision-language benchmarks demonstrate that our method effectively alleviates perceptual degradation, leading to improvements in visual grounding and reasoning accuracy, while offering interpretable insights into how MLLMs process visual information.