Visual Question Answering

Vision-language models (VLMs) remain unreliable when predictions require fine-grained visual evidence. We identify a previously overlooked cause: spectral response rigidity. Despite substantial frequency variation across images and tasks, pretrained vision encoders exhibit persistent, encoder-specific layerwise spectral profiles that change only marginally under downstream fine-tuning. Since pretrained vision encoders only receive images, they cannot adapt spectral extraction to the evidence required by the current query. We therefore propose HAFI-VLM, which introduces a task-conditioned frequency pathway while preserving the pretrained semantic representation. Hierarchical Adaptive Frequency Injection (HAFI) retrieves complementary low-, mid-, and high-frequency evidence at multiple encoder depths using text-modulated, spatially aligned cross-attention. A Visual Enrichment Layer Adapter further recalibrates shallow LLM attention to effectively utilize the enriched visual tokens. Experiments on LLaVA-1.5 and Qwen2.5-VL demonstrate consistent improvements in general VQA, text-rich understanding, and hallucination robustness, outperforming representation-level enhancement methods and most resolution- or cropping-based approaches without additional high-resolution encoding. Mechanistic analyses show that HAFI restores task-dependent spectral allocation while retaining semantic attention, establishing frequency enrichment as a distinct and effective route for improving VLM perception.

We present our ImageCLEF 2026 Multimodal Reasoning system for the Visual Multiple Choice Question Answering (Visual MCQ) and Visual Open Question Answering (Visual OpenQA) subtasks. The challenge requires reliable reasoning over multilingual educational and scientific images with dense text, diagrams, charts, tables, formulas, and units, while enforcing strict answer formats. Our central finding is that robust output control is as important as model choice. For Visual MCQ, we replace fragile free-form generation with direct candidate label scoring from vision-language model logits, then combine complementary runs through score fusion and voting. For Visual OpenQA, we use image enhancement, concise final answer prompting, deterministic decoding, and targeted post-processing to remove reasoning traces and formatting artifacts. Without task-specific model training, our official submissions achieved third place in Visual MCQ with 0.7108 accuracy and first place in Visual OpenQA with 0.6488 COMET, 0.1391 BLEU, 0.2762 ROUGE L, and 0.2383 METEOR. The results highlight the practical value of inference engineering: careful scoring, ensembling, prompting, and cleanup can turn strong VLMs into reliable competition systems.

Multimodal large language models have advanced visual understanding, yet perception-intensive reasoning remains challenging. Recent latent visual reasoning methods introduce hidden-space computation before answering, but they often rely on costly intermediate supervision, such as bounding boxes, sketches, or interleaved rationales. These strategies focus on how latent states should be shaped, but do not explicitly assess whether the latent is useful for the final answer. We propose LUT, a latent reasoning framework trained with only standard VQA pairs. LUT centers training on Latent Utility at two levels. At the trajectory level, we propose Utility-Aware Latent Distillation SFT, which explores answer-relevant latent trajectories, selects qualified trajectories by their information gain, and distills more reliable and learnable supervision through curriculum learning. At the step level, we propose Latent Attribution Policy Optimization, which uses answer-to-latent attribution to differentially optimize latent steps during reinforcement learning. Experiments on perception-intensive visual reasoning benchmarks show that LUT outperforms previous latent reasoning methods and remains competitive with latent-text interleaved methods with lower annotation cost.

Multi-page document visual question answering requires locating sparse evidence at both the page and region levels. Existing approaches typically emphasize one level over the other: page-centric methods focus on page acquisition, with region operations serving mainly as navigation aids, whereas region-centric methods assume that the relevant pages have already been supplied. Consequently, page and region selection remain disconnected rather than forming successive evidence decisions. We propose HierDoc, a hierarchical evidence-routing framework that formulates long-document evidence acquisition as two-stage set prediction from pages to regions. A page policy selects evidence pages from the full document; these pages are then parsed for semantic elements, after which a region policy selects the elements passed to a downstream answer model. Both answer-agnostic policies are optimized with stage-wise GRPO using granularity-specific structured-set rewards. The answer model receives selected full pages together with selected region crops and OCR or table text, preserving global context while emphasizing fine-grained evidence. Across the evaluated benchmarks, HierDoc achieves state-of-the-art or competitive performance among open-weight systems, improving LongDocURL by 16.87% relative to the strongest reported open-weight baseline. Controlled ablations further show that selected regional evidence improves the page-only system in accuracy and F1 by 5.51% and 4.82%, respectively. These results demonstrate the benefit of organizing coarse page routing and fine-grained region routing as successive, separately optimized stages of a unified evidence-acquisition process.

High-resolution visual question answering (HR-VQA) is often treated as a problem of insufficient evidence acquisition, where failing multimodal large language models must inspect images again through cropping, re-encoding, or multi-round search. We show that this view is incomplete: in many cases, fine-grained evidence has already survived visual encoding and become identifiable and influential within an intermediate-layer routing window, but is later diluted before answer generation. We propose Thinking-Once, a training-free, single-visual-pass evidence-routing method that reconstructs question-conditioned attention at this window, preserves core entity tokens and compact background context, and routes this evidence to later layers without extra visual encoding. Across five base models, Thinking-Once consistently improves or matches the corresponding base setting, increasing the average scores on V$^*$Bench, HRBench-4K, and HRBench-8K by +3.1, +3.0, and +2.7 points while reducing the average peak memory by about 4,GB. On Qwen2.5-VL-7B, it improves the three benchmarks by +9.9, +4.6, and +5.5 points, raising the cross-benchmark mean from 72.5 to 79.1. With the ZwZ-8B base model, Thinking-Once reaches a mean score of 82.7. Against 11 open-source HR-VQA baselines, it obtains the best or tied-best score on all three benchmark averages and the best overall mean; for example, compared with DeepScan, it reduces V$^*$Bench inference time by 97.2% while improving the cross-benchmark mean from 77.8 to 79.1. These results show that HR-VQA can be improved by routing already encoded evidence rather than repeatedly acquiring new visual inputs. Code is available in the appendix.

Scientific figure comprehension and reasoning using multimodal AI requires integrating visual perception with domain-specific reasoning to extract meaningful knowledge, often not presented in the text of a research publication. The Sci-ImageMiner benchmark dataset, accompanied by a community-driven competition, raises the bar over prior scientific competitions by curating a comprehensive, expert-annotated dataset across four end-to-end complementary tasks. The competition attracted 68 active participants and 1,263 public/private submissions from 9th January 2026 to 8th April 2026. Our results show that state-of-the-art multimodal models perform well on classification and summarization tasks but struggle with data extraction and scientific reasoning, particularly in visual question-answering. These findings reveal key limitations and highlight challenges and opportunities for improving domain-aware multimodal AI systems. Overall, the Sci-ImageMiner benchmark and competition establish a rigorous platform for advancing research in scientific figure comprehension and reasoning and demonstrate the potential of state-of-the-art approaches for a challenging and complex research area.

Stateful multimodal assistants encode an image once but may answer questions about it many turns later. Attention-guided visual-KV eviction assumes that evidence irrelevant now will remain dispensable, although future questions are unknown. We ask when a visual fact is actually safe to forget and introduce the Causal Visual Memory Audit (CVMA), a paired single-prefill framework that tests what later answers lose when a visual region, the whole image, or prior assistant text becomes unavailable. On VisDial and ConvBench, current attention can rank future-useful regions worse than random even though a diagnostic marginal-utility control shows substantial selection headroom. Aggregate scores hide this failure when later turns do not need vision; controlled and stock-generated histories reveal a second escape route, in which assistant-text KV replaces image KV for facts already stated but not reliably for unstated facts. In the tested stacks, safe forgetting is supported by low future visual dependence or fact-specific verbalization---not by low current attention.

Automatic prompt optimization (APO) has been widely adopted to adapt vision-language models (VLMs) to downstream tasks without weight updates, yielding promising results. However, on multimodal tasks, the effectiveness of APO is fundamentally bottlenecked by a blind feedback channel: the optimizer reads the question, the prediction, and the gold answer, but never the input image on which the model failed, and therefore cannot diagnose visually grounded errors. As a remedy, we introduce Cross-Modal Visual Feedback (CMVF). CMVF incorporates (1) a failure-conditioned visual diagnosis stage, in which a stronger optimizer VLM inspects each failed image without access to predictions or labels, and (2) an error-aware aggregation stage that compresses these observations into reusable, task-level visual blind-spot patterns that drive the prompt rewrite. Crucially, the image is consumed only during optimization; the deployed artifact is an ordinary text prompt that runs at the same inference cost as any text-only baseline. Extensive results across 12 VQA datasets and 4 target VLMs demonstrate that CMVF consistently ranks first, improving over the strongest baseline on every target by 2.4 points on average, with gains of up to 6.5 points on individual benchmarks. Moreover, the optimizer self-organizes into expert-style visual checklists that transfer across models without re-optimization.

Recent Vision-Language Models (VLMs) have achieved remarkable success in visual understanding, driven by the growing availability of high-quality image-text pairs. However, the performance of VLMs often degrades in the video domain due to the essential need for temporal understanding and the scarcity of large-scale annotated video data. In this work, we focus on marine video understanding, which brings further challenges: first, it requires substantial domain expertise; and video VLMs usually struggle with localizing and interpreting critical information from marine videos, as the informative events are typically sparse, unpredictable, and unevenly distributed. To address these challenges, we carefully curate the first event-centric marine video understanding dataset called MarineEVT, which features 20K multi-task, video-level visual question-answering pairs spanning multiple dimensions of marine understanding and analysis. Meanwhile, based on MarineEVT, we decompose marine video understanding as an Event-centric Visual Tool-integrated Reasoning process EVT-R1 for short, where we leverage powerful visual tools to drive the model to localize and interpret critical information aligned with visual questions and human intent. To demonstrate its effectiveness, we compare EVT-R1 against 11 SOTA VLMs in different settings. EVT-R1 outperforms the top open-source and top commercial models by 5.22 and 11.09, respectively. MarineEVT and EVT-R1 lay the foundation for ecological discovery and marine education, fostering the development of VLMs capable of interpreting marine dynamics, reasoning about ecological interactions, and supporting sustainable ocean video understanding and analysis.

Answer grounding in visual question answering aims to locate the region from a given natural language question associated with the visual content of an image, which has garnered significant attention due to its practical applications. In this paper, we introduce the Dynamic Dual-level Vision Transformer Fusion Network (DDVT) for answer grounding in visual question answering. Specifically, we propose a question-guided dynamic regional-level module (QGDR) that combines complementary image context through ROI Align and text content, enabling precise localization of text-related visual content. Moreover, we present a cross-modal multi-scale aggregation module (CMA) that enhances feature fusion between pixel-level and region-level features, facilitating the effective localization of visual content associated with grounded answers. Furthermore, we fuse the located visual content with text features to locate the region and provide answers to questions posed about the image. Experimental results demonstrate that our DDVT outperforms state-of-the-art methods on several widely-used benchmarks.

This technical report presents our approach to Challenge Track~3: SeePhys Pro at the 3rd AI for Math Workshop, where the task is to answer college-level physics questions whose statement and figure may be given partly or entirely as an image. Visual physics problems become substantially harder for large language models when the decisive information resides in a figure rather than in the text, and this modality gap widens as more of the problem migrates into the image. We address the task with a two-stage framework: a visual information extraction stage that re-expresses figure content as solver-readable text to close the modality gap, and a reasoning stage that orchestrates three heterogeneous solvers through multi-agent debate. Our analysis yields two findings: the gain from orchestration comes from reliable answer selection rather than from additional debate, and the value of a figure aid scales with how much of the problem is locked inside the image. The resulting pipeline improves overall accuracy over a single-agent baseline from 0.643 to 0.802 on the public split, and won 1st place on both the public and the private leaderboard (private overall 0.743).

Autoregressive multimodal large language models (MLLMs) suffer from error snowballing: a single incorrect inference early in a chainof-thought (CoT) trace corrupts all downstream reasoning. We find that in state-of-the-art open-source MLLMs, once the first error occurs, the reasoning cascades into failure across all remaining steps in 65% of such cases (a metric we term the snowball rate). Existing mitigations-sampling multiple chains, post-hoc self-verification, or full program synthesis-either lack symbolic grounding, catch errors too late, or sacrifice the flexibility of natural language reasoning. We propose Constraint-Anchored Reasoning Traces (CART), a neuro-symbolic framework that trains MLLMs to interleave natural language reasoning steps with symbolic constraint assertions: lightweight, machine-checkable statements about visual content (e.g., count(red_objects) = 3). A dual-pronged Constraint Propagation Module-combining a learned neural grounding head with Boolean Constraint Propagation-continuously verifies these anchors against extracted visual features and checks their mutual logical consistency. When a contradiction is detected, a backtrack controller halts generation and reverts to the last consistent checkpoint, preventing error propagation. A variable-frequency emission mechanism allows the model to adaptively control anchor density, avoiding trace bloat. We construct 218K training instances by augmenting GQA, CLEVR-CoGenT, and VCR with ground-truth constraint annotations derived from scene graphs, and fine-tune open-source MLLMs (LLaVA-NeXT, Qwen2-VL) via LoRA. On five benchmarks, CART reduces the snowball rate from 0.65 to 0.14, improves GQA accuracy by +4.6 percentage points over trainingonly baselines, and achieves 89.1 F1 on POPE-all with at most 18% inference overhead.

Compositional visual question answering requires Vision-Language Models (VLMs) to execute multiple reasoning operations like object selection, spatial relation resolution, and attribute verification. Despite strong aggregate performance, the mechanistic basis of VLM failures on this task remains underexplored. To address this gap, we analyze vision-operation misalignment in VLMs by examining how failures relate to specific reasoning operations and the internal computational pathways through which they arise and propagate. We introduce an Operation-centric mechanistic framework that decomposes VLM failures by both the reasoning operation where they originate and the internal computational pathway through which they propagate. Our analysis reveals four mechanistically distinct failure modes: grounding failure, reasoning failure, attribute extraction failure, and language prior dominance failure. Each characterized by a unique relationship between visual grounding strength and answer correctness. Through three complementary causal interventions applied across all transformer layers, we further demonstrate a pathway dissociation: grounding failures route exclusively through the feedforward network, reasoning failures route through late-layer attention, and attribute extraction failures localize to the answer-position feedforward computation. This dissociation demonstrates that different failure types require fundamentally different corrective strategies, providing a principled foundation for targeted improvements to VLM reliability in multimedia reasoning.

Where should the question go in a vision-language model (VLM) prompt: before the image or after it? Intuition says before: knowing what is asked should tell the model where to look. Yet across visual question answering benchmarks, question-first prompting consistently underperforms the image-first ordering recommended for frontier VLMs, a phenomenon we term the question-first paradox. We trace this paradox to a conflict between two stages of VLM computation. Logit-lens and attention probes show that question-first prompting steers perception, shifting image patch representations toward question-relevant concepts. But downstream, stranded behind hundreds of image tokens, the question is barely attended by the answer token, which instead commits to image-driven, often wrong answers. Causal attention knockout confirms that the answer reads the question only when it follows the image. This diagnosis yields a training-free fix: question echoing, restating the question on both sides of the image so one copy steers perception while the other is available at answer time. A similar division of labor appears in a fifty-year-old finding on human 'adjunct questions', where repeating a question before and after a passage improves comprehension. Echoing the image as well brings further gains by restoring the whole-image view otherwise lost by a causal decoder. The paradox holds across five open VLMs, costing up to 17.5 group-accuracy points. Echoed prompts recover most of the gap and, on NaturalBench and Winoground, surpass the best single-pass ordering by up to 19 group-accuracy points on Winoground, with no training, fine-tuning, or architecture change. The paradox reveals a tension between steering what a model sees and preserving access to what it was asked; echoing resolves this through prompt design. Project Page: https://rakshanda-cmu.github.io/ask-twice-look-twice/

Egocentric Visual Question Answering (VQA) has attracted widespread attention as an important task for enabling Multimodal Large Language Models (MLLMs) to interact with the real world. However, existing MLLMs struggle to perform effective spatial reasoning in complex egocentric scenes due to their limited spatial perception capabilities. To this end, we introduce Ego Scene Augmentation (ESA), an egocentric spatial perception framework, which actively enhances the spatial perception capabilities from the egocentric perspective, powered by the proposed Ego-element Graph. Our core insight is leveraging the Ego-element Graph as an intermediary representation to augment the egocentric spatial perception of MLLMs via visual foundational models. Specifically, we 1) construct the Ego-element Graph, which encapsulates and integrates egocentric spatial features enabled by visual foundational models; 2) enhance the spatial perception capabilities of MLLMs via the Ego-element Graph for ego-perspective scenes. Our proposed ESA framework presents significant performance improvement on the EgoTextVQA benchmark. We achieve an 8.14% gain on the indoor setting and an 8.72% gain on the outdoor setting. Furthermore, our ESA shows the most impressive performance improvement in the shopping subset of the indoor setting. The project code is publicly available.

As vision-language models (VLMs) are increasingly deployed in geospatial question answering and visual scene understanding, improving their spatial cognition capability on street view imagery for complex logical reasoning has emerged as a key research priority. However, existing VLMs frequently suffer from "spatial semantic hallucinations" when perceiving object locations, distances, and directions in real-world street view scenes. Furthermore, such errors are often recalcitrant to tracing and calibration, posing a critical bottleneck for their practical deployment in geospatial tasks. To address this pressing challenge, this study proposes DM-KG (Direction-Metric Knowledge Graph), a structurally grounded spatial representation framework for street view imagery. By explicitly extracting directional and metric relationships between entities from a single 2D image, this framework enhances the spatial reasoning accuracy of VLMs through a structured knowledge graph. Specifically, we integrate panoptic segmentation with metric depth estimation to robustly compute entity-level 3D spatial coordinates. Subsequently, we encode the clock azimuths and Euclidean distances of entity pairs into a JSON-formatted knowledge graph, which is injected into the VLM as an explicit geometric prior to guide spatial reasoning. Experimental results on public spatial question-answering (QA) benchmarks demonstrate that DM-KG reduces the mean absolute error (MAE) in distance estimation by 31.1% and the mean angular error in direction judgment by 65.8%, while simultaneously maintaining a high QA success rate. By establishing a complete, augmented reasoning pipeline, this research significantly improves the spatial cognitive capabilities of VLMs in street view scenarios, thereby providing a flexible, generalized, and interpretable framework for geographic visual question answering (GeoVQA) in open environments.

Efficient waste segregation is critical for sustainable urban management and environmental governance. Existing automated systems are limited by single-modality visual processing, insufficient contextual understanding, and weak regulatory alignment. To address these issues, we propose a language-guided vision-AI framework that integrates vision-language models and multimodal large language models for joint visual-linguistic reasoning. This framework implements a visual question answering paradigm aligned with India's Solid Waste Management Rules 2016. We construct a new WasteVQA dataset with 13,500 question-answer pairs across 21 waste categories. Experiments show that the BLIP-based model achieves a BLEU score of 0.8291 and a BERTScore of 0.9273, outperforming traditional CNN-based methods. This work improves source-level segregation accuracy, ensures regulatory compliance, and supports scalable deployment for municipal and citizen-facing waste management, promoting multimodal AI in sustainable urban infrastructure. The source code and dataset are available at: https://github.com/Khushkataruka/WasteAssistant

Short-answer VQA benchmarks conflate two distinct quantities: whether a model's answer is semantically correct, and whether that answer matches the surface form expected by the automatic evaluator. We study this conflation across six vision--language models and six benchmarks, using a human-validated semantic judge (97.6% precision) to audit over 37k official errors. A second text-only judge reproduces the same benchmark-level false-negative pattern, showing that the effect is not an artifact of a single audit model. On text-rich benchmarks, up to half of these errors are semantically acceptable answers penalized purely for surface-form mismatch. This instability is structured by answer type: extractive and multi-span answers are far more evaluator-sensitive than scalar answers. Benign prompt and context rewrites further destabilize official outcomes, flipping item-level correctness at substantial rates without changing the underlying task. A deterministic CPU-only contract repair confirms that the undercount is partially recoverable. These findings imply that official short-answer VQA scores should be accompanied by semantic audits and answer-type diagnostics to remain interpretable.

Test-time scaling (TTS) reliably improves reasoning in large language models, but whether it transfers to small open vision-language models remains unclear. We examine this on EXAMS-V, a multilingual visual multiple-choice benchmark, comparing self-consistency, describe-then-reason with PRM-guided beam search, and two post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B. What matters is the conditions under which TTS runs, not the search or verification machinery. The largest factor is parseability: an early prompt format left many chains reasoning correctly yet never committing to an answer letter, which a standard answer cue and a guided repair step largely remove. A larger decoding budget removes the rest: raising the per-chain token limit from 1k to 2k recovers 3.7 pp, whereas sampling more chains (8 to 16) adds only 0.15 pp. Once chains have room to finish, elaborate methods contribute little: PRM-guided beam search trails plain self-consistency by 0.39 pp at over eight times the cost, and neither a training-free generative critic nor a trained multimodal PRM beats majority vote across both policies. The largest gain comes instead from the policy model itself (+11.4 pp). Our best configuration reaches 84.1% on the held-out ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard.

Recent advancements in LVLMs necessitate robust benchmarks for complex, visually grounded reasoning. A critical limitation is identified in many document understanding benchmarks: visual content is often reducible to text, enabling high performance without genuine visual grounding. To address this limitation, OmniMapBench is introduced to foster visual-centric reasoning for map documents. The benchmark comprises 2,096 manually annotated question-answer pairs across 1,603 map documents from nine categories. It is designed to probe a hierarchy of skills, ranging from perception to multi-step visual reasoning. To quantify benchmark properties, a simple yet effective benchmark-level metric is proposed: the Visual Dependency Index (VDI), defined as the accuracy drop when images are replaced with question-agnostic descriptions. OmniMapBench exhibits higher VDI than established benchmarks, which quantitatively validates its focus on irreducible visual reasoning. Comprehensive evaluations of 25 leading LVLMs are conducted on OmniMapBench. A significant performance gap is observed, with the top-performing model achieving only 75.03% accuracy. This result underscores the challenges posed by OmniMapBench to current LVLMs. This work aims to catalyze progress in visual-centric reasoning for document understanding of LVLMs. The dataset and code are publicly available at https://github.com/SIGMME/OmniMapBench.

Document Visual Question Answering (DocVQA) presents a complex multimodal challenge, requiring models to exploit visual, textual, and layout information from documents. Although Vision-Language Models (VLMs) have shown remarkable performance in text-vision tasks, their robustness and transferability to different document domains remains underexplored. In this study, we present a comprehensive evaluation of 8 open-source pretrained VLMs on DocVQA in three different document domains: industrial documents of varying type, infographics, and presentation slides. We systematically assess model performance under zero-shot evaluations, fully supervised finetuning with inter- and intra-dataset evaluations, and few-shot learning evaluations of knowledge transfer between domains. Our findings demonstrate that while large pretrained VLMs possess strong zero-shot baselines for structured layouts, their performance strongly decreases on visually complex layouts of infographics and slides. Although parameter scaling is a dominant factor on performance, supervised finetuning yields higher relative gains in smaller architectures. Furthermore, our cross-domain and few-shot experiments show that visual understanding is the main bottleneck for DocVQA, not a lack of knowledge from the VLMs. Using 50 target domain samples, the models finetuned in DocVQA with datasets of different domains rapidly adapt to the target domain documents, even surpassing their fully supervised counterparts in some cases.

While Large Vision-Language Models (VLMs) demonstrate remarkable generic capabilities, their clinical reasoning in specialized domains like ocular surface diseases (OSDs) is severely hindered by a paucity of high-fidelity, multimodal instruction-tuning data. To dismantle this data bottleneck, we introduce IRIS, an Intelligent Recognition and Interaction System tailored for fine-grained OSD understanding via external eye photography. First, we curate IRIS-120K, the largest and most comprehensive OSD visual question-answering (VQA) dataset to date. Crucially, to overcome the semantic shallowness of conventional image-caption pairs, we propose a synergistic data generation paradigm to explicitly inject clinical priors. Our data engine operates via a dual-branch framework: 1) a Topic Finding Tree (TFT) that hierarchically anchors visual features to precise anatomical and pathological concepts, enforcing rigorous medical deduction logic; and 2) a Scene-driven strategy that synthesizes role-adaptive clinical dialogues to ensure pragmatic generalization. By explicitly aligning a compact 4B-parameter VLM on this structurally enriched corpus, IRIS achieves state-of-the-art performance, comprehensively outperforming both generalist and specialized medical VLMs with up to 34B parameters. Our findings underscore that structured knowledge injection profoundly prevails over sheer parameter scaling, unlocking the potential for resource-efficient, expert-level AI deployment on mobile edge devices for scalable OSD screening. Code, datasets, and model weights will be publicly released by this repo.

In safety-critical traffic scenarios, answering complex questions relies on minute, localized visual cues. However, standard Multimodal Large Language Models (MLLMs) tend to over-attend to backgrounds, overwhelming crucial small objects during visual-language alignment, a failure mode we term 'critical evidence dilution.' Furthermore, existing visual question answering (VQA) datasets rarely expose this flaw, as they lack large-scale, distractor-heavy evaluations that require pinpointing local evidence. To bridge this evaluation and architecture gap, we introduce the Fine-Grained Traffic Reasoning Benchmark (FGTR-Bench) and the Text-Guided Small-Object Reasoning MLLM (TSR-MLLM). FGTR-Bench comprises 40,236 single-image Multiple-Choice Questions (MCQs) created via multi-agent generation, consistency checks, and expert audits, alongside a disjoint 4,947-sample blind test split. To resolve evidence dilution, TSR-MLLM, built on Qwen3-VL-4B, uses a query-conditioned Text-Guided Small-Object Focus (TG-SOF) map. Applied once at the decoder boundary, the map adds sparse Top-K gated residuals to the most question-relevant vision slots while leaving text tokens unchanged. Together with lightweight decoder adaptation, TSR-MLLM preserves single-pass inference without external detectors or image re-encoding. Under matched settings, TSR-MLLM outperforms the strongest 4B baseline by 2.1 points on FGTR-Bench (74.1% overall), with larger gains on evidence-local tracks. Furthermore, it remains competitive on DriveQA-V (CARLA Signs) under greedy decoding without task-specific fine-tuning.

Knowledge-Based Visual Question Answering (KB-VQA) aims to evaluate whether Visual Language Models (VLMs) can retrieve, ground, and reason over external structured knowledge beyond visual evidence. In practice, answer accuracy is widely adopted as the primary evaluation metric, implicitly treating correctness as a proxy for knowledge-grounded reasoning. However, for existing KB-VQA benchmarks, this proxy relies on critical assumptions that are often overlooked and rendered unreliable by benchmark issues: annotated answer must be derivable from the associated knowledge base, question must be well-posed with sufficient constraints, and visual setting must meaningfully require grounded disambiguation. In this work, we show that these assumptions are systematically violated in existing KB-VQA benchmarks. Our audit reveals substantial instances with missing or contradicted answers and underspecified questions that render accuracy a misleading metric. Furthermore, we find that existing datasets rely on visually trivial, single-entity scenes that bypass the need for sophisticated visual-to-knowledge mapping. We demonstrate that even with controlled architectures, these flaws lead to distorted model rankings and overestimations of reasoning capabilities. To address this, we introduce (1) a principled audit-and-repair protocol that restores answer derivability and question clarity, and (2) a controlled multi-entity augmentation protocol that introduces visual ambiguity to challenge initial retrieval and grounded reasoning. Re-evaluation under corrected and augmented settings yields markedly different performance trends. Our findings call for rethinking evaluation protocols and designing more interaction-aware KB-VQA benchmarks that prioritize verifiable reasoning over simple matching.

Vision-language models can produce confident answers on visually ambiguous inputs, resulting in biased predictions. Common entropy-based methods, such as Semantic Entropy (SE), rely on output diversity. Yet our analysis shows that overconfident visual embeddings suppress output diversity under stochastic decoding, causing SE to underestimate uncertainty in such cases. Recent methods instead probe output diversity through input perturbations, including textual paraphrasing or joint text-image perturbations, and show improved performance. We study these approaches and reveals that the resulting variability is often dominated by textual changes rather than visual evidence, causing uncertainty estimates to reflect prompt sensitivity rather than visual ambiguity. We therefore propose Visual Semantic Entropy (VSE), which perturbs only the image to probe nearby visual variations while keeping the text query fixed. VSE measures uncertainty by clustering generated answers into semantic prototypes and computing the mass-weighted dispersion among them. Extensive evaluation across five modern vision-language models and five diverse VQA benchmarks demonstrates that VSE effectively captures visual ambiguity, establishing a new state-of-the-art for VLM uncertainty estimation.

Multimodal large language models (MLLMs) often fail in fine-grained visual reasoning, as question-relevant visual cues are diluted by dense and redundant image tokens. Recent multimodal reasoning methods usually extend chain-of-thought from language models into visual or latent spaces, seeking to add intermediate reasoning states while overlooking the negative impact of redundant visual tokens. We propose LatEnt Noise maSk (Lens), a question-conditioned visual evidence purification framework that empowers MLLMs to reason with cleaner visual cues in latent space. Lens introduces a lightweight Lens Evidence Token (LET) to score which visual tokens support the current question and preserve them during decoding. Guided by the LET scores, it injects adaptive latent noise into low-relevance tokens, softly suppressing distractors without changing the model backbone or token sequence. With only one temporary learnable control token and a lightweight noise generator, Lens adds minimal overhead while improving the base MLLM by 2.4-6.4 points on most VQA datasets and by 4.1-6.4 points on grounding tasks. These results show that multimodal reasoning can benefit more directly from cleaner question-relevant visual evidence than from simply extending the reasoning trace.

Multimodal large language models (MLLMs) are widely applied to visual document understanding. However, comprehending long documents remains an issue by the limited context window. Though recent multimodal retrieval-augmented generation (MMRAG) can address this challenge by retrieving relevant pages. It still struggles with the visual question answering (VQA) requiring holistic comprehension of a document. To cope with this, knowledge graph (KG) that summarizes global knowledge of a document can provide an effective solution. However, most existing LLM-based KG construction methods handle only the language modality, leaving the automatic creation of multimodal KGs (MMKGs) for visually rich documents largely unexplored. In this paper, we introduce a multimodal graph-based RAG approach to tackle this problem. Existing LLM-based KG methods evaluate the QA performance relying on indirect evidence such as comprehensiveness, diversity, empowerment, and so on. The lack of annotated datasets for comprehensive document-level VQA poses a significant challenge to effective model evaluation. To overcome this limitation, we also introduce a new benchmark, DLVQA (document-level VQA), which provides reference summaries and corresponding supporting facts for global document-level questions. Experimental results show that our approach outperforms existing MMRAG or KG-based approaches on multi-hop QA/VQA benchmarks and DLVQA.

Wildfire monitoring from UAVs requires reliable reasoning over complex aerial scenes, where smoke, scale variation, and occlusions often limit RGB-only interpretation. We introduce FlameVQA, a multiple-choice visual question answering benchmark for UAV-based wildfire intelligence built on FLAME 3, leveraging paired RGB imagery and radiometric thermal TIFFs for temperature-grounded, safety-critical reasoning. FlameVQA includes 34 multiple-choice questions per image spanning six operational capability groups, covering tasks such as detection, localization, distribution/coverage estimation, cross-modal reasoning, and flight planning. To ensure label reliability, we combine MLLM-assisted annotation with deterministic thermal rules and cross-question consistency checks, followed by human auditing. We also evaluate representative MLLMs on FlameVQA to provide baselines for future work. Results show strong performance when explicit cross-modal cues are available, but notable failures on presence detection under heavy smoke and on coverage estimation. These findings suggest that current MLLMs require domain-specific adaptation to better support disaster and wildfire monitoring. The dataset and benchmark code are open-source at github.com/mobiiin/WildFire_VQA

Multimodal Large Language Models (MLLMs) have shown remarkable progress in single-image perception, yet their ability to reason about complex cross-view human-centric scenes remains largely unverified. Current multi-view benchmarks evaluate models using a fixed "bag of frames" and thus conflate a model's robustness to visual distraction with its genuine ability to fuse fragmented cross-view evidence. To address this issue, we introduce SSMNBench, a diagnostic benchmark comprising 3,300 curated QA pairs for cross-view human and human-object understanding. SSMNBench uniquely categorizes tasks into Single-View Sufficiency (SVS) and Multi-View Necessity (MVN). By systematically perturbing view availability across 17 state-of-the-art MLLMs, critical limitations are revealed: models suffer from severe "distraction degradation" when presented with redundant views (SVS), and fail to integrate fragmented geometric evidence across cameras (MVN). Our evaluations demonstrate that modern MLLMs rely on multiple single-image semantic averaging and view preference rather than genuine cross-view synthesis. By exposing these fundamental vulnerabilities, SSMNBench provides a rigorous diagnostic framework to drive the advancement of future cross-view-aware multimodal architectures. The code is available at: $[SSMNBench](https://github.com/gtc-gh/SSMNBench)$

Vision Language Models have achieved near-human performance on single-document Visual Question Answering, yet their effectiveness degrades significantly when retrieving information from large collections of visually homogeneous documents. Existing multi-document benchmarks aggregate diverse document types, creating artificial separation in embedding space that does not reflect enterprise document repositories where thousands of records share identical visual templates. We identify this as embedding collapse and introduce Invoice Haystack, a benchmark with 1,500 anonymized invoice images paired with 200 discriminative question-answer pairs, specifically designed to stress-test retrieval under strong visual homogeneity. Invoice Haystack exhibits a mean pairwise cosine similarity of 0.73, compared to 0.38 (DocHaystack) and 0.31 (InfoHaystack) in existing benchmarks, posing a fundamentally more challenging retrieval problem. Addressing the identified challenge, we propose VL-RAG, a hybrid retrieval-augmented generation framework that jointly leverages text and visual embeddings to harness the complementary strengths of both modalities, followed by a VLM-based verification filter for precise document identification. VL-RAG achieves 60.0% Recall@1 on Invoice Haystack-500, outperforming existing state-of-the-art method by up to an absolute 13.5 percentage points. It further improves retrieval considerably on DocHaystack-1000 (77.1% vs.\ 75.2%) and InfoHaystack-1000 (84.5% vs.\ 80.0%), establishing the proposed dual-stream fusion as a consistently superior retrieval strategy across both homogeneous and heterogeneous document collections.