Large video models have exhibited impressive performance on a wide range of visual question answering tasks, owing to the rise of powerful, pretrained text and vision encoders. The usefulness of such models have also been demonstrated on a wide range of benchmarks, with an important caveat - the dominant approach in these benchmarks evaluates multiple choice reasoning via text options. This is a natural way to test text-based reasoning in these models, and has led to significant insights regarding model behavior in the community. In this work, we ask a different question - what happens when the evaluation modality is visual, rather than text? We introduce three new vision-centric evaluation benchmarks in temporal frame retrieval, video future prediction, and causal memory distortion, all designed around evaluating visual understanding capabilities in large video models. Our approach complements the existing approaches to evaluate video understanding in frontier models. We show that current frontier models exhibit significant weakness when attempting to reason through visual queries, rather than text. We conclude with an extended analysis section that provides pointers for future improvements in visual understanding for large video models.
Scientific documents contain complex multimodal structures, which makes evidence localization and scientific reasoning in Document Visual Question Answering particularly challenging. However, most existing benchmarks evaluate models only at the page level without explicitly annotating the evidence regions that support the answer, which limits both interpretability and the reliability of evaluation. To address this limitation, we introduce SciEGQA, a scientific document question answering and reasoning dataset with semantic evidence grounding, where supporting evidence is represented as semantically coherent document regions annotated with bounding boxes. SciEGQA consists of two components: a **human-annotated fine-grained benchmark** containing 1,623 high-quality question--answer pairs, and a **large-scale automatically constructed training set** with over 30K QA pairs generated through an automated data construction pipeline. Extensive experiments on a wide range of Vision-Language Models (VLMs) show that existing models still struggle with evidence localization and evidence-based question answering in scientific documents. Training on the proposed dataset significantly improves the scientific reasoning capabilities of VLMs. The project page is available at https://yuwenhan07.github.io/SciEGQA-project/.
Existing Multimodal Knowledge-Based Visual Question Answering (MKB-VQA) benchmarks suffer from "visual shortcuts", as the query image typically matches the primary subject entity of the target document. We demonstrate that models can exploit these shortcuts, achieving comparable results using visual cues alone. To address this, we introduce Relational Entity Text-Image kNowledge Augmented (RETINA) benchmark, automatically constructed using an LLM-driven pipeline, consisting of 120k training and 2k human-curated test set. RETINA contains queries referencing secondary subjects (i.e. related entities) and pairs them with images of these related entities, removing the visual shortcut. When evaluated on RETINA existing models show significantly degraded performance, confirming their reliance on the shortcut. Furthermore, we propose Multi-Image MultImodal Retriever (MIMIR), which enriches document embeddings by augmenting images of multiple related entities, effectively handling RETINA, unlike prior work that uses only a single image per document. Our experiments validate the limitations of existing benchmarks and demonstrate the effectiveness of RETINA and MIMIR. Our project is available at: Project Page.
Fine-grained audio-visual question answering (AVQA) requires models to resolve complex cross-modal correspondences and temporally grounding from sparsely distributed evidence. Omni-modal agents, which autonomously orchestrate perception tools to actively collect evidence, achieve strong performance, but it remains unclear whether active perception from both audio and visual modalities depends on broad specialized toolsets, especially audio or visual modality-specific tools. Such systems may also introduce long execution trajectories and substantial inference cost. We investigate whether comparable performance can be retained with a compact workflow exposing only essential perceptual capabilities. Starting from the specialized-tool OmniAgent, we conduct staged controlled experiments that prune the toolset and progressively consolidate modality-specific tools into omni-modal tools, motivating a two-tool agent comprising only Omni-modal captioning and Omni-modal QA. Across three widely used AVQA benchmarks, our compact agent achieves comparable or better accuracy than the specialized-tool agent while reducing inference cost by around 70%.
Document Visual Question Answering (DocVQA) often leverages Retrieval-Augmented Generation (RAG), where late-interaction encoders are commonly used to identify document pages relevant to a user query, before answer generation by a Large Vision-Language Model (LVLM). Existing approaches typically retrieve a fixed top-$k$ number of pages regardless of query complexity, which increases LVLM latency and may degrade answer accuracy when fewer pages would be enough. We introduce ViSAR (Visual Semantic Activation Retrieval), a training-free, budget-aware adaptive-$k$ retrieval method for late-interaction visual document retrieval. ViSAR operates directly in the embedding space to construct a query-conditioned page-level similarity matrix that highlights query-relevant semantics and dynamically determines the number of pages to retrieve, within a given page budget. We evaluate our method across multiple datasets, encoders, and LVLMs. ViSAR retrieves compact, query-adapted page sets that reduce end-to-end RAG latency by up to 58.9\%, with no statistically significant accuracy loss and significant gains in several settings, compared with fixed top-$k$ and adaptive retrieval heuristics. Furthermore, we show that the similarity matrix structure correlates with answer accuracy, suggesting future directions for retrieval quality-aware document understanding.
Establishing a clear link between model predictions and the visual evidence that supports them is critical for transparency and reliability in multimodal reasoning, yet current multimodal large language model (MLLM) evaluations do not explicitly enforce this alignment. Existing benchmarks assess either textual answer correctness or pixel-level localization in isolation, leaving the coupling of reasoning and grounding an open challenge. We introduce VISTAQA, a comprehensive benchmark for joint evaluation of free-form answer correctness and pixel-level evidence grounding in visual question answering. VISTAQA comprises 1,157 expert-curated samples spanning six task types and six visual domains, ranging from direct perception to compositional and relational reasoning. VISTAQA requires models to not only answer correctly, but to also provide precise segmentation masks that support their answers. It also includes hallucination-aware examples where no valid visual evidence exists. To support this enhanced evaluation, we introduce GROVE, a unified evaluation metric that enforces joint correctness by combining textual accuracy and grounding quality via a per-sample geometric mean, ensuring neither dimension can compensate for deficiencies in the other. Comprehensive experiments across grounding-aware models and hybrid pipelines with general-purpose MLLMs reveal that even the strongest systems achieve limited performance under GROVE, highlighting a substantial gap between answer accuracy and visual evidence alignment. Project page can be found at https://vistaqa.github.io.
Embodied Question Answering (EQA) couples active visual perception, vision-language reasoning, and sequential decision making, and has traditionally been evaluated in temporally stable environments where visual evidence can be accumulated reliably. In realistic multimedia settings such as dynamic human-populated scenes, however, moving people and transient occlusions introduce human-induced perceptual variations: task-relevant multimodal cues become transient, viewpoint-dependent, and incompletely captured in individual views. Conventional store-then-retrieve strategies generate redundant visual evidence and incur high inference cost, making them ill-suited to such variations. This exposes two core problems for embodied agents: resolving viewpoint-dependent ambiguity, and maintaining compact evidence for efficient inference. To address these challenges, we present DIVRR (Dynamic-Informed View Refinement with Relevance-guided Rotation and adaptive memory selection), a training-free active evidence-acquisition framework for EQA. DIVRR combines question-conditioned room guidance with relevance-triggered view augmentation to guide exploration and additional observation. Selective memory admission uses the same relevance signal as view augmentation to retain high-scoring evidence, reducing stored entries while keeping the retrieved visual context compact. To evaluate EQA under human-induced perceptual variations, we introduce DynHuman-EQA, a dynamic human-populated EQA dataset constructed in Habitat, comprising dynamic and static subsets under shared scene layouts. Experiments on DynHuman-EQA and HM-EQA show that DIVRR outperforms existing baselines in both accuracy and memory efficiency. The gains are most pronounced in dynamic human-populated scenes.
Vision-language models (VLMs) have achieved strong image and video understanding, yet their visual-spatial representations remain geometrically fragile, leading to failures in spatial reasoning needed for embodied AI, robotics, and autonomous driving. Prior approaches to geometry grounding either fine-tune VLMs on spatial question answering, which can perpetuate spurious visual representations, or fuse features from large geometry-grounded vision models, which substantially increases model size at inference. Knowledge distillation from geometry-grounded vision models offers an alternative, but directly matching multi-view teacher features can disrupt the pretrained alignment between visual and textual representations, degrading object- and language-semantic capabilities. We propose multi-view relational distillation (MVRD), which distills patch-wise cosine similarities across views instead of the teacher features themselves. These relations encode geometric correspondences adequate for spatial understanding, while leaving the student representation underdetermined, allowing it to remain close to its pretrained vision- language space. Across representative VLMs, MVRD improves visual-spatial reasoning, outperforming supervised fine-tuning and feature distillation while approaching feature fusion methods with considerably fewer added parameters and lower latency. We show that MVRD makes visual representations more geometric while retaining language alignment, and generalizes to 3D scene understanding tasks such as object grounding, dense captioning, and question answering. Code is available at https://github.com/kietngt00/MVRD
Vision-Language Models (VLMs) can produce Natural Language Explanations (NLEs) that sound plausible yet remain inconsistent with the visual evidence they cite. We present Explanation-Driven Counterfactual Testing (EDCT), an intervention-based protocol that extracts visual concepts cited in a model's explanation, applies verified minimal edits to them, and tests whether the resulting answer and explanation remain consistent with the edited image. Using this protocol, we create EDCT-Bench, a comprehensive benchmark spanning three complementary domains: knowledge-intensive visual question answering (OK-VQA), safety-critical driving (DriveLM), and 3D spatial reasoning (3DSRBench). Across the evaluated VLMs, EDCT reveals substantial faithfulness gaps, with models frequently producing responses inconsistent with verified visual changes. Finally, our fine-tuning study suggests that EDCT-generated counterfactuals provide high-impact training signals.
Marine video understanding is crucial for a wide range of real-world applications. However, existing marine video datasets suffer from limitations, including the limited number of data samples, fine-grained text descriptions, and detailed visual attributes. To address these limitations, we introduce a large-scale, video-text dataset featuring rich, in-the-wild visual attributes tailored for complex marine environments. Our dataset comprises 2,000 video-text pairs annotated with 13 attributes covering up to four objects per scene, object visual/geometric properties, behaviors, environmental context, and camera motion, spanning diverse lighting conditions, times of day, and 7 countries. From these annotations, we generate 16,080 video-question pairs (12,080 multiple-choice, 4,000 open-ended). Furthermore, we benchmark both proprietary and open-source video-LLM models on video question answering and video captioning tasks. We find that (i) proprietary models often achieves superior performances; (ii) VLMs handle object motion well but struggle with complex camera motion; (iii) closed-source models struggle to perceive camera motion; and (iv) VLMs struggle to reason in static cameras.
Robot tasks require grounding spatial language in physical scenes, but task success alone cannot establish whether spatial instructions are followed. We present BEHAVIOR-Spatial, a benchmark built on BEHAVIOR-1K that measures spatial question answering, point-based target grounding, and robot action to examine how VLM spatial understanding relates to VLA behavior. It comprises 3,256 spatial questions across 49 activities, 1,240 grounding queries across 14 manipulation tasks, and action evaluations covering those 14 tasks and 10 navigation tasks. Our manipulation protocol evaluates held-out configurations and instruction following when the requested target changes within a fixed training layout. The baseline achieves 41.6% on held-out configurations but only 16.1% under changed instructions. VLM adaptation and robot-data pretraining provide limited improvements, whereas oracle visual target cues raise these rates to 60.3% and 38.0%, respectively. Perception evaluations reveal policy-dependent correlations between VLM question-answering and grounding accuracy and VLA manipulation performance. Together, these findings motivate evaluating spatial instruction following alongside task completion and examining when spatial perception performance translates into robot behavior. Project website: https://arielchen07.github.io/behavior-spatial/.
Multimodal latent-space reasoning aims to replace explicit thinking with images by performing visual reasoning directly in a compact latent space. However, existing approaches largely rely on visual supervision and produce latent representations that lack sufficient semantic richness, limiting their ability to support diverse region-level reasoning tasks. In this work, we introduce Semantic-Enriched Latent Visual Reasoning (SLVR), a two-stage learning framework that enriches latent representations with attribute-level visual semantics and aligns them with diverse reasoning objectives. In the first stage, SLVR learns semantically enriched region-centric latents under fine-grained attribute supervision. In the second stage, we design Multi-query Group Relative Policy Optimization (M-GRPO) to align latent representations across multiple queries grounded in the same region. To support this framework, we construct SLV-Set, comprising approximately 400K region-level attribute annotations and 800K multi-query question answering samples, and introduce SV-QA, a benchmark that evaluates latent reasoning under semantic variation. Experiments demonstrate that SLVR improves the robustness and semantic consistency of latent visual reasoning compared to existing baselines.
In recent years, Multimodal Large Language Models (MLLMs) have made significant progress in visual question answering tasks. However, directly applying existing fine-tuning methods to remote sensing (RS) images often leads to issues such as overfitting on background noise or neglecting target details. This is primarily due to the large-scale variations, sparse target distributions, and complex regional semantic features inherent in RS images. These challenges limit the effectiveness of MLLMs in RS tasks. To address these challenges, we propose a parameter-efficient fine-tuning (PEFT) strategy called Guided Region-Aware Sparse Prompting (GRASP). GRASP introduces spatially structured soft prompts associated with spatial blocks extracted from a frozen visual token grid. Through a question-guided sparse fusion mechanism, GRASP dynamically aggregates task-specific context into a compact global prompt, enabling the model to focus on relevant regions while filtering out background noise. Extensive experiments on multiple RSVQA benchmarks show that GRASP achieves competitive performance compared to existing fine-tuning and prompt-based methods while maintaining high parameter efficiency.
Long-video question answering is challenging since the answer often hinges on a few decisive moments scattered throughout the video, while memory constraints force Multimodal Large Language Models (MLLMs) to sample frames at extremely low rates, resulting in sparsity that risks omitting key evidence. Though aggressively enlarging the context size is intuitive, we argue that performance remains fundamentally constrained by the bias, i.e., the misleading cues caused by spurious linguistic correlations and salient yet irrelevant visual observations. In this paper, we first introduce S-MME, a benchmark designed to systematically diagnose shortcut-biased behavior in MLLMs. To mitigate this failure, we further propose a training-free framework, Counterfactual Long-video Evidence-Aware Reasoning (CLEAR). By treating the full video-question pair as a complete view, CLEAR estimates bias effects via augmented views by only keeping the question or locally salient visual clips. Then, we intervene hidden states to mitigate the discrepancy at inference such that the correct answer can be predicted confidently. With experimental studies across five benchmarks, we show that our method is generalizable and can be integrated in different MLLMs, yielding consistent gains in prediction and robustness. We hope CLEAR can offer a complementary direction in test-time scaling for long-video understanding.
Reliable visual question answering requires correct answers when evidence is sufficient and abstention when it is not. We introduce a benchmark of 3,000 question groups: 1,000 each from PlotQA charts, CLEVR rendered scenes, and GQA photographs. Each group keeps the question fixed across original, answer-preserving, and evidence-removing images; charts and CLEVR also include answer-changing edits. Models receive each image independently. A group succeeds only if every supported answer and every required abstention is correct. For charts and CLEVR, executable witnesses establish ambiguity under the declared task semantics: two admissible complete charts or scenes give different answers but yield the same masked observation. GQA scoring compares source labels with a permissive rule that also credits masked-image answers compatible with the full-image reference. Across 72,000 responses from six model configurations, the best complete-question success is 57.0% on charts, 43.5% on CLEVR, and 33.7% on GQA under source labels. On charts, the strongest configuration makes correct answer-or-abstain decisions on 96.2% of individual images and on all five images of 83.5% of questions. Requiring correct answers lowers success to 57.0%: in 26.5% of all chart questions, every decision is correct but at least one answer is wrong. Adding answer correctness also reverses a chart ranking based on decision accuracy. Evaluating all response obligations together reveals failures that per-image decision accuracy conceals.
Multimodal large language models (MLLMs) have heterogeneous strengths across OCR, chart understanding, spatial reasoning, visual question answering, cost, and latency. Effective MLLM routing therefore requires more than estimating query difficulty: a router must match the multimodal requirements of the current image-question input with the capabilities of each candidate model. We propose LatentRouter, a router that formulates MLLM routing as counterfactual multimodal utility prediction. Given an image-question query, LatentRouter extracts learned multimodal routing capsules, represents each candidate MLLM with a model capability token, and performs latent communication between these states to estimate how each model would perform if selected. A distributional outcome head predicts model-specific counterfactual quality, while a bounded capsule correction refines close decisions without allowing residual signals to dominate the prediction. The resulting utility-based policy supports performance-oriented and performance-cost routing, and handles changing candidate pools through shared per-model scoring with availability masking. Experiments on MMR-Bench and VL-RouterBench show that LatentRouter outperforms fixed-model, feature-level, and learned-router baselines. Additional analyses show that the gains are strongest on multimodal task groups where model choice depends on visual, layout-sensitive, or reasoning-oriented requirements, and that latent communication is the main contributor to the improvement. The code is available at: https://github.com/LabRAI/LatentRouter.
Reasoning is essential for autonomous driving (AD) in long-tail scenarios, where vehicles must apply commonsense knowledge, understand spatial relations, infer agent interactions, and make safe decisions. However, existing AD datasets and benchmarks mainly target perception, prediction, or planning, and provide limited supervision for reasoning over realistic long-tail driving scenes. We introduce nuReasoning, a large-scale real-world dataset and benchmark for reasoning-centric AD. Following the lineage of nuScenes and nuPlan, nuReasoning advances real-world AD datasets and benchmarks toward reasoning in long-tail driving scenarios. The dataset contains 20,000 clips, each 20 seconds long, collected across multiple cities, with synchronized multi-camera images, LiDAR data, HD maps, object annotations, and human-verified reasoning annotations spanning Spatial Reasoning, Decision Reasoning, and Counterfactual Reasoning. Unlike prior datasets that focus primarily on visual question answering, nuReasoning supports both reasoning evaluation and planning evaluation, enabling a direct study of how reasoning supervision affects driving performance. Experiments show that fine-tuning VLMs on nuReasoning substantially improves driving-specific question answering, while incorporating reasoning supervision into VLA training improves planning performance even when textual reasoning outputs are disabled at inference time. These results establish nuReasoning as a foundation for evaluating and improving robust, interpretable, reasoning-driven AD systems in realistic long-tail settings.
Vision-Language Models (VLMs) have demonstrated remarkable capabilities in multi-modal understanding for tasks such as Visual Question Answering (VQA), which requires models to infer outputs based on visual and textual context simultaneously. However, models may memorize some private and/or sensitive information during training and regurgitate it in inference. Recently, machine unlearning has been leveraged to address such leakage of private data in LLMs. VLMs add a layer of complexity to this process, as the visual context in the query may additionally contain sensitive information. To address this issue, we explore unlearning for VLMs, specifically for the VQA task. We explore the role of visual tokens for output generation in VLMs using cross-modal attention and utilize it to formulate Cross-Modal Attention Guided Unlearning (CAGUL), a lightweight and efficient VLM unlearning framework. In contrast to computationally expensive finetuning methods, CAGUL utilizes external modules to encode unlearning information in visual tokens of low importance for relevant queries. Empirical results show that the transformed visual tokens prevent leakage and retain reference model behavior to achieve desirable trade-offs in forget quality and model utility without altering pretrained model parameters.
General-purpose vision-language models struggle with crisis imagery due to its inherent noise, ambiguity, and high variability across disaster events - yet timely, accurate scene understanding is critical for humanitarian response. To address this, we introduce DisasterVLG, a large-scale multi-task dataset containing 1.24 million disaster images from 142 countries paired with 33 million structured annotations spanning dense captioning with localization, incident assessment, relational scene descriptions, and crisis-grounded visual question answering (VQA) across diverse hazard types and scene contexts worldwide. Unlike prior disaster benchmarks that focus narrowly on classification, DisasterVLG unifies recognition, grounded reasoning, and uncertainty-aware interpretation. A distinguishing feature is its evidence-grounded annotation design: each VQA instance includes supporting evidence text that links visual observations to answers, alongside positive and negative question pairs, multiple answer formats, and difficulty labels. Visual grounding layer links caption phrases to image regions and relationship statements to subject-object boxes, with explicit labels for visual supports and clues. To support rigorous evaluation, we construct a human-verified test set of 38K images (1.15M annotations) spanning broad geographic regions and disaster types. Experiments with domain-adapted vision-language models demonstrate strong performance across disaster captioning, incident recognition, and grounded VQA, while also exposing the limitations of general-purpose models in crisis-specific reasoning. DisasterVLG establishes a scalable, reproducible benchmark for vision-language models that must reason reliably under real-world uncertainty. We will publicly release the dataset, trained models, and code.
Commonsense knowledge is widely used to support multimodal reasoning and generation, yet existing resources are predominantly text-only and lack grounding in visual context. This limitation often leads to the retrieval of semantically related but visually misaligned knowledge, and can degrade downstream performance. We introduce MMCOMET, a large-scale multimodal commonsense knowledge graph that integrates physical, social, and eventive knowledge with visual grounding. MMCOMET extends ATOMIC2020 by associating each commonsense triple with representative images through a hybrid alignment pipeline, resulting in over 900K multimodal tuples spanning 19 relation types. We further establish a unified evaluation framework across diverse vision-language tasks, including visual commonsense reasoning, visual question answering, visual storytelling, image editing, and image captioning. Our analysis shows that multimodal grounding provides more contextually aligned and reliable knowledge compared to both text-only knowledge graphs and LLM-based knowledge generation, particularly in reasoning-intensive scenarios. At the same time, we observe that the effectiveness of external knowledge is task-dependent, and naive knowledge integration can introduce noise in perception-dominated settings. To mitigate this, we propose an adaptive knowledge router that dynamically selects optimal paths or bypass knowledge in harmful settings. MMCOMET provides a scalable multimodal commonsense resource and a benchmark for studying the role of grounded knowledge in multimodal systems.