VLM Evaluation

Industrial anomaly detection has historically been a unimodal task. Recent multimodal vision-language models have produced systems that admit textual input alongside the image and are presented as enabling text-guided zero- and few-shot inspection. Yet these methods are evaluated with protocols inherited from unimodal benchmarks that hold the textual condition constant and therefore cannot measure whether language conditions the decision; whether reported gains reflect text guidance or strong pretrained visual features remains open. We introduce Text-Guided Anomaly Detection (TGAD), a structured benchmark that progressively increases the functional role of language across three scenarios: a controlled prompt-sensitivity setting on MVTec AD; a component-tagged extension of MVTec AD that requires the model to restrict its assessment to an instructed part; and the new Assembled Panel Dataset (APD), a realistic industrial setting that requires both defect-type and component-location knowledge. We evaluate one representative model per paradigm: generative large vision-language, training-free discriminative, and embedding-adaptive discriminative. In all three, the textual interface conditions the decision only superficially: prompt content is absorbed unless the object noun is removed (the generative model's I-AUROC drops from 97.4 to 82.6); component-level instructions do not constrain the decision once defects outside the instructed part are admitted as normal (from 90.3 to 66.3); and when both combine on APD, image-level discrimination collapses below the MVTec level, in one case below chance (71.2, 50.5, 31.5). These results suggest that standard benchmarks overstate the text-guided capabilities of current multimodal anomaly detection systems, and that a protocol of this kind is a prerequisite for models that can be reliably controlled through language for industrial deployment.

We introduce the Image Reconstruction Game, a fully automated benchmark in which a vision-language model issues corrective instructions to an image generator across multiple turns, making accumulated common ground directly observable as a rendered image. Benchmarking two Describer models crossed with two Generator models across seven image categories, we find that the describer is the dominant factor in reconstruction quality, while the generator determines whether iterative refinement helps or hurts. Mathematical and geometric images pose the greatest challenge. The describer's token budget strongly affects convergence: shorter budgets yield sparser first renderings with more room for visible improvement, while longer budgets raise absolute quality but leave less to fix. Stronger describers use a richer correction vocabulary spanning spatial, numeric, and structural categories, while weaker describers concentrate on surface properties and tend to stop after a few turns. Human validation shows that the best automated judge reaches only slight-to-fair agreement with human preferences, and automated scores require human recalibration to be used reliably.

Driving vision-language models (VLMs) must accurately understand scenes across diverse conditions defined by Operational Design Domains (ODDs), yet verification remains sparse: many slices are missing, making empirical failure rates unreliable. We propose SliceScorer, a deterministic scoring rule for missing-slice recommendation that combines (i) an exposure-based coverage prior to prioritize rare, under-tested regions, and (ii) a neighbor-failure prior that propagates risk from similar tested conditions. SliceScorer is deliberately simple - interpretable, auditable, and conservative - properties essential for safety-critical validation. For stress testing beyond the declared ODD, we embed SliceScorer within SliceNav, an LLM-orchestrated verification pipeline where the model interprets developer queries to select relevant operators (triage, scoring, acquisition, evaluation) and vocabulary extensions, composing verification workflows while keeping all scoring deterministic and auditable. Experiments on three driving VLMs (WiseAD, DriveMM, Cosmos-Reason2-2B) show that SliceNav surfaces high-risk coverage gaps more effectively than prior slice-discovery methods while maintaining diverse recommendations across the condition space. Ablations confirm both scoring components contribute, and qualitative analysis demonstrates end-to-end workflows from developer query to targeted evaluation.

Document parsing and recognition are fundamental capabilities for vision-language models (VLMs) and document processing systems. However, existing Optical Character Recognition (OCR) and document parsing benchmarks are increasingly limited in coverage and difficulty: many focus on common document genres or uniformly sampled pages where modern parsers already perform strongly, while offering limited annotation for expert-domain structures such as chemical formula, music notation, complex tables, and cross-page layouts. We introduce Dr. DocBench, a difficulty-aware benchmark for expert-level document parsing. Built from a large-scale multilingual book corpus, Dr. DocBench spans 52 BISAC subject domains and selects challenging documents through parser-failure-based sampling, targeting cases where multiple state-of-the-art systems struggle. It contains 4,514 annotated pages from long documents averaging around 100 pages, with 65k high-quality page- and block-level annotations for layout, reading order, hierarchical relations, and domain-specific visual contents. Evaluations of pipeline-based parsers and general-purpose VLMs show that strong performance on existing benchmarks does not transfer to our expert-level document parsing. Our analysis reveals substantial failures across subjects, content types, and structural attributes, highlighting Dr. DocBench as a comprehensive testbed for diagnosing and advancing document intelligence.

Charts are a primary medium for conveying quantitative and relational information, yet systematically evaluating chart parsing models remains difficult. Existing benchmarks focus on narrow chart types and leave diagrammatic structures such as flowcharts and mind maps largely unaddressed, while models produce outputs in incompatible formats, and datasets rarely include the printed or hand-drawn images encountered in practice. To address these issues, we introduce ChartArena, a comprehensive bilingual benchmark covering eight chart families spanning both numeric charts and diagrammatic structures, each evaluated across three visual scenarios: digital renderings, printed photos, and hand-drawn photos. The dataset is built via a human-agent collaborative annotation pipeline with multi-stage human verification to ensure annotation reliability. To enable fair cross-model comparison, we further design a format-agnostic evaluation protocol that maps heterogeneous outputs into two canonical semantic spaces, a normalized triple view and a directed graph view, and scores them with structure-aware metrics. Through extensive evaluation of 26 leading MLLMs, we observe three consistent findings: (i) frontier proprietary models such as Gemini 3.1 Pro lead overall, yet the strongest open-source systems are rapidly closing the gap; (ii) document parsing models handle numeric charts reasonably but fall sharply behind on diagrammatic structures; and (iii) expert chart parsers remain limited to narrow chart families. Across all models, radar charts and hand-drawn scenarios stay especially challenging. These findings show that ChartArena exposes clear capability gaps and provides a unified foundation for future progress. ChartArena is publicly available at https://github.com/pspdada/ChartArena.

Recent latent visual reasoning methods achieve substantial gains by inserting continuous latent tokens into multimodal language models. These gains are commonly attributed to the tokens encoding visual evidence; recent analyses, however, reveal a paradox: the tokens are loosely tied to the image and contribute little to the answer. Critically, these analyses treat latent tokens as a single unit, obscuring the true source of the gains. We therefore decompose latent tokens into three testable components: latent slots, boundary markers, and format, and develop a state-of-the-art method as a probe under favorable conditions. Across six method-stage settings and four perception-heavy benchmarks, latent slots fail every prediction of the visual-memory account. Strikingly, retaining only the boundary markers preserves 78 to 100% of the gain in several settings, while the model attends to the image more narrowly at latent positions than at answer positions. The gain therefore comes from boundary markers, format, and this attention pattern, not from latent slots. How each method engages this mechanism depends on its training supervision: at matched accuracy, mechanisms can still differ markedly. Latent visual reasoning thus needs evaluation not only by accuracy but by what the model actually relies on.

Despite tremendous recent progress, current text-guided image editing methods still struggle with many aspects of editing involving instruction following, minimally editing the source image, and ensuring high visual quality. These problems are especially apparent when the requested edit is challenging, such as those that involve position, motion, viewpoint, scale and creative edits. To systematically test generative image editors, we propose a novel image editing benchmark -- TECCI: Tricky Edits of Collected and Curated Images. TECCI consists of a completely new set of images we are releasing. The images in TECCI span 7 image categories. The images and these categories were curated intentionally to target weaknesses of existing methods. The edit instructions in TECCI are automatically generated by Gemini, covering 5 edit types per source image. We also curated a set of 530 images for which we created challenging manually written edit instructions. Overall, TECCI contains 7550 pairs of images and edit instructions. We conduct human evaluations of five leading image editing models on TECCI. Humans judge outputs along three dimensions: 1) instruction following, 2) minimality of the edits, and 3) visual quality. To scale-up the evaluation, we also build an auto-rater using Gemini that achieves 74.7% accuracy in matching human evaluations. Our evaluations reveal that: 1) none of the models exceed a 22% overall success rate, demonstrating the challenging nature of TECCI, 2) Nano Banana Pro is the best performing model overall, 3) models perform significantly better at instruction following compared to minimal edits and visual quality, 4) models struggle with editing architecture and nature images which require strong understanding of spatial layout and intricate visual details. 5) reasoning and creative edits are the most difficult, whereas color and appearance edits are the easiest.

Understanding chart and table images is essential for applying vision-language models (VLMs) to real-world document understanding. While English benchmarks have advanced rapidly, non-English counterparts remain scarce, leaving it unclear whether this progress generalizes across languages. A key obstacle is the difficulty of collecting realistic and diverse non-English chart and table images at scale. To address this, we leverage governmental white papers as a scalable source for benchmark construction beyond English, as they contain naturally occurring charts and tables across diverse formats and domains and are freely accessible in many countries. As a first instantiation, we introduce HakushoBench, a challenging Japanese chart and table VQA benchmark built from 33 governmental white papers. HakushoBench contains 2,053 images spanning over 10 image types, with manually annotated QA pairs, designed to assess deep and holistic understanding of charts and tables, rather than local visual cues alone. Experiments across a broad range of VLMs demonstrate that HakushoBench remains challenging for open-weight models: the best open-weight model achieves only 58.6% accuracy, and a 34.9-point gap between open-weight and proprietary models highlights substantial room for improvement in complex chart and table understanding. We release our dataset and code.

Procedural 3D modeling through code is emerging as a versatile paradigm, offering deterministic, engine-ready, and precisely editable assets that neural 3D generators inherently lack. Authoring such procedural content, however, demands deep expertise in 3D software APIs, parametric design, and code-level geometric reasoning. In this paper, we propose 3DCodeBench, a systematic benchmark for evaluating vision-language model (VLM) agents for procedural 3D generation in 3D modeling software. Specifically, 3DCodeBench evaluates how effectively 12 advanced VLMs can serve as procedural 3D modelers by translating text and image references into procedural code for 3D modeling software. Recognizing that automated metrics may not fully capture the perceptual quality of 3D shapes, we build 3DCodeArena, a ranking platform based on pairwise human preferences over generated 3D outputs. From extensive evaluations and results, we observe that: (1) Failures mostly arise from API mismatches, while successful renders still suffer from disconnected or floating 3D geometric components. (2) Test-time scaling, such as higher thinking budgets and multi-turn refinement, improves performance overall. Our findings highlight a critical need for high-quality procedural coding data to advance commercial VLMs. Furthermore, effective procedural 3D modeling requires a robust execution environment that provides high-fidelity feedback for iterative refinement. We release 3DCodeBench, including the curated large-scale dataset of multimodal (text/image) prompts, procedural code, 3D object triplets, evaluation protocol, and the public 3DCodeArena platform as a foundational toolkit for exploring VLM-based procedural 3D modelers.

Vision-language models (VLMs) are increasingly used to generate structured descriptions of street-level imagery for tasks such as streetscape auditing, mapping, and public consultation. These uses combine observable attributes with appraisal categories, and the human targets are often distributions of judgments with disagreement and explicit non-response. This paper argues that benchmarking VLMs for urban perception should treat disagreement and abstention as measurement outcomes, report inter-annotator reliability alongside model alignment, and treat the label space and scoring policy as negotiable artifacts when outputs are intended to inform urban governance. We ground the argument in a benchmark of 100 Montreal street scenes annotated along 30 dimensions by 12 participants from seven community organizations, and in a deterministic zero-shot evaluation of seven VLMs. Across dimensions, model agreement with human consensus co-varies with dimension-level human reliability, and for the appraisal dimension Overall Impression models and annotators exhibit distributional mismatch including different rates of Not applicable. We close with actions for benchmark creators, model developers, and institutions to make uncertainty and benchmark assumptions visible in evaluation reports.

Vision-Language Models (VLMs) have shown strong visual understanding and are increasingly deployed in embodied AI systems, where reliable perception under real conditions is essential. However, existing benchmarks assess VLMs using clean images or isolated perturbations rather than stresses caused by physical scene formation. This design has two limitations: it covers only a narrow subset of everyday visual stresses, and some perturbations rarely appear in realistic embodied scenes. This gap raises a fundamental question: how can we define visual stress in a principled way that captures the diverse factors encountered in physical environments? To address this question, we formulate visual perception from an inverse graphics perspective and introduce RoboStressBench, a benchmark for evaluating VLM robustness to physical visual stress in embodied scenes. Inspired by the physical rendering equation, RoboStressBench decomposes visual stress into four physically grounded dimensions: Material (M), Viewpoint (V), Lighting (L), and Geometry (G). This design enables RoboStressBench to cover a broad range of visual stresses in real-world environments, while allowing controlled analysis of their effects on VLM capabilities such as visual recognition, reasoning, and planning. Through comprehensive evaluations of state-of-the-art VLMs, we identify stress-specific failure modes and reveal that different physical factors degrade different embodied capabilities, which are often obscured by aggregate accuracy. We further introduce a stress-aware agentic solver that detects visual stressors and invokes visual-editing skills before reasoning, improving robustness in high-stress scenarios. Overall, RoboStressBench provides a principled evaluation framework for diagnosing and improving VLM perception under real-world physical stress, supporting the development of more reliable embodied AI systems.

A new framework for the estimation of the complexity posed by video-question pairs to video-LLMs, Video Attribute-Based Complexity (VideoABC), is proposed. Video complexity is defined as the probability of failure of a video-LLM for a given video-question pair. VideoABC is a non-parametric complexity measure, using a reference video dataset and a pre-defined vocabulary of video attributes informative of complexity, \eg the scene complexity or the speed of the video event informative of the question. In a training phase, reference videos are projected into the space of these attributes, which is then quantized. The expected ABC of each quantization cell is then computed. Given a new video and its projection into the attribute space, complexity is estimated by the expected ABC of the associated quantization cell. To enable the use of VideoABC with small reference video datasets, two quantizers are combined: a k-means quantizer that enables accurate complexity estimates for samples in the distribution of the reference dataset and a universal lattice quantizer that guarantees generalization to out-of-distribution samples. A synthetic video generation procedure, inspired by target-distractor manipulations of psychophysics studies, is proposed to populate the cells of the lattice quantizer during training, enabling the computation of their expected ABCs. Experimental results show that VideoABCis effective even with very low-dimensional attribute representations, substantially outperforming approaches like `video-LLM as judge' with much less complexity. Finally, the explainable nature of the VideoABC score, in terms of well-defined attributes, is shown to provide insights on how the attribute composition of benchmarks affects their complexity.

Effective visual communication stems from the harmony of multiple design principles, such as readability, contrast, alignment, overlap, and coherence, which collectively govern clarity and intent of the communicator. While human designers reason holistically over these principles, machine agents typically condense them into a single heuristic score, offering limited interpretability and diagnostic precision. To address this gap, we introduce PRISM (PRinciple-aware, Interpretable, and Structure-guided Design Modifications), a benchmark that systematically perturbs professional layouts from the Crello dataset along measurable design principles. The benchmark comprises 100K perturbed training samples and 10K perturbed validation designs, each isolating a specific principle violation for controlled analysis of multimodal reasoning about design quality. We show that models like Qwen-2.5-VL and GPT-4o-mini are largely insensitive to targeted principle degradations, whereas GPT-4o exhibits global awareness without fine-grained disentanglement. Building on these insights, we propose a multi-scale evaluation framework that integrates lightweight scorers for quantitative assessment, instruction-tuned vision-language models for localised feedback, and prompt-based methods for global reasoning. Our framework provides interpretable explanations of design failures. Using these localised insights, we show targeted refinements that improve layout quality. Together, PRISM and our framework lay the foundation for interpretable design-literate multimodal reasoning systems.

Measuring structured object understanding in vision foundation models remains challenging due to inconsistent evaluation protocols and limited part-level supervision. Semantic correspondence (SC) evaluates this capability by testing whether object parts can be matched across instances and categories under large variations in appearance, viewpoint, and geometry. To enable a systematic SC evaluation, we introduce SOCO, a new benchmark for Semantic Object Correspondence that introduces a taxonomy of correspondence types and provides consistent, functionally meaningful keypoint annotations across 100 categories and over 1M correspondence pairs. In addition, SOCO includes keypoint language descriptions, enabling the evaluation of large vision-language models (LVLMs) and their fine-grained part-level understanding. Comprehensive experiments reveal that (i) vision foundation backbones encode strong semantic structure but transfer correspondences poorly across related categories and only partially capture object-part position, (ii) LVLMs are stronger at text-prompted part localization than at visual-reference cross-image matching, exposing a gap between language-grounded localization and fine-grained visual correspondence, and (iii) correspondence performance predicts performance on dense downstream tasks, including segmentation, tracking, 3D pose estimation, and 3D detection, more strongly than ImageNet classification. Together, these findings position SOCO as a benchmark for structured, part-level representation quality in vision and multimodal foundation models.

Continuous episodic memory is a core capability for autonomous agents operating in dynamic, real-world environments, yet current streaming video benchmarks provide limited tools for diagnosing what models remember and for how long. We introduce Egostream, a diagnostic benchmark for streaming episodic memory evaluation in egocentric vision. \egostream organizes 2,250 curated questions along seven cognitive dimensions: detail, spatial, temporal, event, social, causal, and prospective memory. We introduce the Answer Validity Window (AVW), which specifies the temporal span an answer remains valid as the observed scene evolves. This allows us to expand the questions into 8,528 recall-conditioned evaluations, enabling controlled testing from instant to ultra-long-term recall while separating genuine model forgetting from natural world-state changes. We rigorously establish baseline performance through a unified streaming MLLM framework that compares several state-of-the-art memory-management mechanisms, covering sliding windows, attention sinks, KV-cache pruning, merging, and offloading. Experiments within a unified Qwen3-VL backbone reveal that comparable aggregate accuracies mask starkly different memory profiles. For instance, token pruning preserves fine-grained details and temporal structure significantly better than token merging, while quantized offloading rescues ultra-long-term recall. Ultimately, all mechanisms operate well below real-time (>1s per frame), and top performing methods ceil at about 45% accuracy, exposing critical gaps in current architectures. Egostream provides the diagnostic testbed needed to close these gaps. Project website, news and updates at: https://saroo25.github.io/Egostream/

Alignment teaches vision-language models (VLMs) to avoid expressing demographic biases, and when gender is clearly visible they largely succeed. Far less is known about ambiguous inputs (a worker in full gear, a figure seen from behind), cases common in practice yet rarely studied. We find that minimal prompting pressure exposes occupation-gender defaults when prompting ambiguous input images, with models collapsing to male even for strongly female-stereotyped occupations. But do these outputs reflect what models actually encode internally? We introduce LALS (Latent Association Leaning Score), a zero-shot metric that projects visual-token activations into the model's text-embedding space to measure concept associations per token and layer. Across 15 occupations, over 800 gender-ambiguous images, and four VLMs, internal representations and outputs often become systematically decoupled: models often encode a female association internally yet output male. Layer-wise analysis reveals an asymmetric filter: male signal amplifies end-to-end while female signal peaks mid-network and is suppressed before generation. A color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.

Visual evidence has long been treated as a reliable form of legal proof, but advances in artificial intelligence (AI) are undermining that assumption. This article asks how well humans and frontier multimodal large language models (MLLMs) can distinguish authentic evidentiary photographs from AI-generated counterparts in the object-centric scenarios typical of civil disputes. We built Synthetic Legal Evidence Detection (SLED-1400), a dataset of 200 authentic evidence images paired with 1,200 synthetic counterparts produced by six contemporary text-to-image generators across ten evidence categories. The same stimuli and response format were used in a controlled web experiment with 136 lay participants and in a standardized evaluation of four MLLMs (GPT-5.1, Gemini-3-Pro, Gemini-3-Flash, Qwen3-VL-235B). Human accuracy was 64.8% overall, and 48.5% and 51.0% on the two strongest generators (Gemini-3-Pro-Image and Flux-2-Max), indistinguishable from chance. MLLMs never misclassified an authentic image (100% specificity), but missed most synthetic outputs from the harder generators, with average MLLM detection at 5.9% on Gemini-3-Pro-Image outputs. Human and MLLM errors were largely uncorrelated, while the four MLLMs were strongly correlated with each other. Neither group is a reliable standalone authenticator. We argue that visual evidence in legal proceedings should be treated as inherently contestable, and that a workable procedural response must combine trained human review, MLLM screening, and provenance infrastructure such as C2PA Content Credentials.

Vision-Language Models (VLMs) largely follow the text-only LLM trajectory, excelling on English benchmarks but sharply degrading on low-resource languages, where neither large-scale image-text corpora nor culturally grounded evaluations exist. We present a systematic study of building a language-specific VLM for Romanian, covering the full pipeline from data construction to architectural choices. We translate established English VLM training and evaluation corpora into Romanian, applying machine translation to textual annotations and to in-image text, preserving visual grounding while adapting the textual content. Using this data, we train and ablate a series of VLMs to isolate the contribution of (i) vision backbones of varying scale and pretraining, (ii) language backbones from multilingual to Romanian-adapted LLMs, and (iii) OCR-style image-text data. We further curate HoraVQA, a culturally native evaluation set grounded in Romanian everyday scenes. Romanian-adapted VLMs consistently outperform their same-sized counterparts and, across all evaluated benchmarks, even surpass models from the next larger size category.

AI-based Visually Impaired Assistance (VIA) remains challenging, largely due to the high cost of human evaluation. The VLM-as-a-Judge paradigm may offer a promising alternative, although it has mostly been studied in general domains. We therefore ask whether such judges can be trusted for VIA tasks. To investigate this question, we introduce VIABLE (Visually Impaired Assistance Benchmark for VLM-as-a-Judge Evaluation), the first benchmark for VLM-as-a-Judge evaluation in VIA. VIABLE contains over 300K judgment samples across three scenarios and introduces an Effectiveness--Impartiality--Stability framework with a 12-mode failure taxonomy. Based on VIABLE, our systematic study of seven judges across different model scales shows that existing models are largely unreliable across all evaluation axes. The strongest judge, GPT-5.4, achieves only 52.6% single-failure diagnostic accuracy, yet exhibits the highest self-preference rate at 94.2%; while open-source judges are strongly biased and adversarially fragile. To address these issues, we propose VIA-Judge-Agent, a model-agnostic inference-time harness that augments judges with visual evidence extraction and a taxonomy-guided workflow. It enables positive improvements in diagnostic accuracy and downstream VIA responses more preferred by BLV users. Data and code are available at: https://github.com/YiyiyiZhao/VIABLE

Hateful meme detection remains a formidable challenge for vision-language models, as existing benchmarks are structurally observational - confounding rhetorical hate mechanisms with target community features and preventing causal evaluation of model vulnerabilities. To address this, we introduce FBHM, a systematically curated benchmark of Functionality Based Hateful Memes constructed along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities (5,000 memes total). Benchmarking state-of-the-art VLMs reveals a severe generalization gap: models highly accurate on standard datasets catastrophically drop to near-random performance on FBHM, proving they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, we propose LSV (learnable steering vectors), an ultra-low data regime strategy that applies a causal intervention objective on as few as 500 steering samples (50 unique base memes), boosting FBHM performance by ~30 Macro-F1 points while outperforming in-context learning and PEFT without degrading source-domain performance.

Recent aerial vision-language-action (VLA) models show promising single-UAV capabilities, such as tracking moving objects and navigating to language-specified landmarks. However, it remains unclear whether these capabilities can transfer to air-ground cooperation, where a UAV and a UGV must act jointly in a shared, closed-loop physical world. We study this question with CARLA-Air, a single-process air-ground evaluation environment that unifies CARLA and AirSim inside one Unreal Engine runtime. By sharing the same world state, physics tick, and sensing pipeline, CARLA-Air enables physically consistent UAV--UGV interaction and precise measurement of simulation-timestamp alignment and effective coordination latency. Using CARLA-Air, we evaluate representative aerial VLA and planning baselines on two complementary diagnostic tasks: moving-platform landing and occlusion-recovery escort. The results show that current aerial VLA models can often track or follow a ground partner, but struggle to convert this single-agent competence into stable cooperative behavior. State prompting provides limited benefit, and naive bidirectional interaction fails to consistently improve performance and can amplify errors for most baselines. These findings suggest that, under the tested text-based cue interfaces, zero-shot cooperative air-ground VLA requires three components beyond the current paradigm: explicit partner-state grounding, low-latency action coordination, and team-level objective alignment. Our code is available at https://github.com/louiszengCN/CarlaAir.

Vision-Language-Action (VLA) models have demonstrated promising capability in autonomous driving, highlighting the potential of unified multimodal architectures for jointly modeling perception and planning. However, how current VLA-based driving behavior is grounded in visual information remains poorly understood. Existing evaluation protocols mainly focus on aggregate performance metrics, lacking structured and practical diagnostics to quantify visual-behavior dependency. In this work, we introduce a structured multi-level visual perturbation framework to analyze visual-behavior dependency in VLA-based driving models systematically. The framework organizes controlled visual perturbations along three complementary dimensions: channellevel degradation, information-level disruption, and structurelevel modification. We apply it to VLA-based driving systems and evaluate behavioral responses under both open-loop trajectory prediction and interactive closed-loop safety evaluation. Experimental results reveal evaluation-dependent dependency patterns and uneven visual grounding across abstraction levels. These findings call for more structured analyses and principled design of VLA driving models to better understand how visual information shapes behavior and develop safer, more robust systems.

Hallucination remains one of the key challenges undermining the reliability of Large Vision-Language Models (LVLMs). But what makes an LVLM hallucinate less? Many existing efforts focus on improving internal components of the model. We argue that hallucination fundamentally stems from how the model architecture is designed. To investigate this, we factor the architecture design into three dimensions: Linguistic Foundation (LF), Visual Representation (VR), and Semantic Alignment (SA), and categorize hallucinations into Co-occurrence, Similarity, and previously overlooked Uncertainty types. Building on this formulation, we propose CoSimUE, a benchmark that creates fine-grained hallucination scenarios through controlled textual perturbations and random perturbations, enabling mapping between design choices and hallucination behaviors. Experiments across 7 design aspects show that: 1) the widely emphasized scaling of model parameters has only limited impact on reducing all three types of hallucinations; 2) larger and better-trained language foundations can reduce co-occurrence hallucinations; 3) stronger visual encoders and higher resolutions mitigate similarity errors; 4) effective alignment strategies alleviate uncertainty hallucinations. 5) Furthermore, cross-dimensional analysis reveals that jointly enhancing visual fidelity and alignment quality yields the most comprehensive improvements. This study provides the first systematic exploration linking architecture-level design to hallucination robustness, offering practical guidance for developing reliable and efficient LVLMs.

Multimodal large language models (MLLMs) often know the rule but pick the wrong answer: on abstract visual reasoning (AVR) tasks, a model can describe what it sees and name the underlying pattern, yet still fail to choose the matching candidate. Existing AVR benchmarks cannot detect this because they collapse perception, rule induction, and answer selection into a single right-or-wrong signal. We introduce StemBind, a shared-stem diagnostic benchmark that probes the same visual stem with three aligned questions: Perception (what is in the image), Rule (what pattern governs it), and Full (which option completes it), so a final-answer error can be attributed to a specific sub-step on the same evidence. StemBind contains 2,298 curated knowledge-light stems across nine auditable visual operations, totaling 19,533 P/R/F tasks, with each full item annotated by Sternberg's four reasoning stages (S1 Encode, S2 Infer, S3 Map, S4 Apply). Evaluating 24 frontier MLLM configurations yields four findings. (i) The R-F chasm: rule accuracy exceeds full-item accuracy on 22 of 24 models, so most failures happen after the rule is identified. (ii) A persistent binding gap: even when P and R are both correct on the same stem, models still answer F incorrectly 51.2% of the time. (iii) The bottleneck is S3: process diagnostics and Stage-wise Stimulus Augmentation localize the dominant failure to rule-to-instance mapping. (iv) Scaling and thinking do not help: neither larger models nor explicit thinking mode reliably closes the gap, and thinking even lowers rule and full-item accuracy. StemBind reframes AVR evaluation from final-answer ranking to locating where abstract visual reasoning breaks down, identifying rule-to-instance binding as a concrete next target for vision-grounded reasoning.

Test-time compute (TTC) strategies have emerged as a lightweight approach to boost reasoning in large language models (LLMs). However, their application and benefits for vision-language models (VLMs) remain underexplored. We present a systematic study of TTC across seven VLMs and six benchmarks, specifically analyzing feature-based scoring and majority voting methods. We find that feature heuristics fail and voting yields only modest gains in single-model settings. We theoretically show that this limitation stems from a lack of prediction diversity: when outputs are highly correlated, voting provides little benefit. In contrast, multi-model ensembles offer richer diversity, yet standard majority voting fails to account for varying model capabilities. To address this, we propose Entropy-based TTC (ETTC), which selects the most confident prediction based on predictive entropy. Our method reduces to majority voting in the single-model case, but in model ensembles, it leverages confidence disparities to prioritize stronger models. We prove that ETTC outperforms majority voting under mild assumptions and empirically demonstrate that it consistently surpasses both voting and the best individual model. Crucially, our results show that smaller models can synergistically enhance larger ones, unlocking ensembling gains not achievable with standard strategies.

Spatial reasoning benchmarks typically evaluate whether vision-language models can derive the correct answer from a visual observation. Yet in real 3D environments, the observation itself may be unreliable: occlusion can remove task-relevant evidence, while perspective can make visible geometry misleading. Reliable spatial reasoning therefore requires more than answering a question correctly. A model must also assess whether its current observation provides sufficient and trustworthy evidence for that answer. We introduce SPATIALUNCERTAIN, a controlled evaluation framework for studying viewpoint-dependent observational uncertainty. We study two complementary failure modes: missing evidence caused by occlusion and misleading evidence caused by perspective. We further evaluate whether models can recognize when the current view is unreliable and identify a more informative observation. Across eight open- and closed-source vision-language models, we find that model behavior does not track the reliability of visual evidence. Models do not reliably become more cautious as evidence disappears, and under perspective conflict, their judgments increasingly follow projected appearance rather than the unchanged physical 3D relation. Internal analysis suggests a corresponding representational asymmetry: projected 2D relations are readily available, whereas the underlying physical 3D relation is barely decodable. Moreover, models that can identify an informative viewpoint when explicitly asked often fail to recognize when such an additional view is needed. These failures are not fully resolved by prompting or fine-tuning, and providing a better viewpoint is substantially more effective than adding depth information to the same misleading observation. Our results identify assessing the reliability of visual observations as a distinct and missing component of current spatial reasoning evaluation.

Vision-language models (VLMs) achieve strong performance on spatial reasoning benchmarks, yet it remains unclear whether this reflects structured 3D understanding or reliance on statistical shortcuts in natural images. We introduce a representation-level analysis framework that constructs minimal contrastive pairs to measure how spatial axes are organized and disentangled within VLM embeddings. Our analysis across multiple model families reveals a consistent vertical-distance entanglement: models conflate vertical image position with distance, mirroring the perspective bias of natural photographs. This bias produces a significant accuracy gap between perspective-consistent and counter-heuristic examples, and intensifies under data scaling even as overall benchmark accuracy improves. We further show that models with similar benchmark scores can exhibit different internal representations, and that these differences predict accuracy and robustness across diverse spatial reasoning benchmarks. To isolate this bias from evaluation-set skew, we introduce SpatialTunnel, a synthetic benchmark designed to expose spatial shortcut biases by removing common correlations present in natural images. Experiments confirm that the entanglement is model-intrinsic, and that models with well-separated spatial axes exhibit greater robustness, suggesting that well-structured spatial representations lead to more reliable spatial reasoning across diverse benchmarks. Code and benchmark are available on the project page: https://cheolhong0916.github.io/whyfarlooksup.github.io/.

Vision-language models (VLMs) have made strong progress on high-level image-text alignment, yet their ability to perceive subtle visual differences remains limited. We study this problem in rendered web interfaces, where localized visual changes are both a diagnostic test of fine-grained perception and a practical requirement for GUI agents and design tools. We introduce \textbf{DiffSpot}, a code-driven benchmark for open-ended spot-the-difference on web interfaces. DiffSpot constructs controlled image pairs by mutating a single CSS property of a target element in self-contained HTML, re-rendering the page, and recording the changed property, element, and mutation magnitude. A grounding gate retains only pairs whose rendered pixel difference is confined to the target element. The benchmark contains 4{,}400 pairs, including 3{,}900 has-diff pairs balanced across 13 CSS-property operators and three difficulty tiers, plus 500 no-diff pairs for hallucination control. Evaluating 13 frontier VLMs zero-shot, we find that even the best model identifies only $40.7\%$ of true changes, with Hard-tier Recall below $23\%$ for every model. DiffSpot further shows that difficulty is strongly property-dependent: across CSS operators, neither pixel magnitude nor CLIP distance reliably predicts Recall.

Vision-language models (VLMs) are increasingly used to answer questions about physical scenes, yet most evaluations reduce performance to a final answer. This hides whether the model perceived the right objects, represented the right physical state, predicted a plausible transition, or merely selected the right option for the wrong reasons. We introduce \wmw, an evaluation framework for auditing the \emph{language-expressed physical commitments} of VLMs. Instead of scoring only $I,q\mapsto a$, we ask models to produce a typed trace $I,q\mapsto(s_0,Δs,s_1,a)$: an initial state, a state transition, a resulting state, and an answer. A hybrid verifier then checks schema validity, state grounding, transition consistency, and answer-trace compatibility, yielding typed error labels such as object, relation, force, transition, temporal, unit/scale, and faithfulness errors. We release \tracebank, a controlled trace resource with \nSeed schema- and recomputation-validated synthetic scenarios across \nFamilies physics families, \nPairs minimally perturbed contrastive preference pairs, verifier code, audit guidelines, and model outputs. We evaluate \nModels VLMs on both controlled and external physical-reasoning examples. \wmw reveals failures that answer-only evaluation misses: 35% of correct answers from mid-tier models are backed by physically invalid traces. Verifier-guided reranking recovers up to 7 percentage points of trace validity without sacrificing answer accuracy, and trace-level preference tuning reduces hidden inconsistency by 41% relative. The contribution is not another final-answer physics benchmark, but a reusable protocol for measuring whether a VLM's stated physical world can be true at the same time as its answer.

This work presents a comparative evaluation of machine translation systems applied to images containing textual information, a task that lies at the intersection of computer vision and natural language processing. The study compares three main paradigms: modular pipelines that separate text detection, recognition, and translation; multi-modal large language models (MLLMs) capable of processing both image and text jointly; and an end-to-end model, Translatotron-V, which directly generates translated images. The modular systems employ state-of-the-art OCR (docTR) combined with multilingual LLMs such as Llama and EuroLLM, while the evaluated MLLMs include different configurations of Gemini 2.5. Experiments were conducted on parallel multilingual datasets covering multiple language pairs, with evaluation based on BLEU, chrF, and TER metrics. The results show that modular pipelines outperform the end-to-end approach, while MLLMs achieve the best overall performance, demonstrating superior flexibility and contextual understanding. These findings underscore the effectiveness of multi-modal reasoning for image-to-text translation and provide a solid foundation for future research on integrating visual understanding and language generation in multilingual settings.