Vision-Language

Automated evaluation is essential for scaling generative 3D systems, where exhaustive human review is costly and slow. However, the reliability of an automated judge depends on the entire evaluation pipeline, not only the underlying vision-language model (VLM), but also how assets are rendered, what visual evidence is provided, how the task is specified, and how human reference labels are constructed. We introduce 3D-DefectBench, a benchmark and framework for systematic analysis of VLM-based 3D defect detection pipelines. It complements holistic ratings and pairwise preferences with nine fine-grained binary defects spanning geometry, texture, and prompt adherence, providing actionable diagnostics for generator development and judge evaluation. Using a balanced factorial design, we vary four pipeline factors, VLM, camera protocol, visual input, and prompt schema, across 84 inference designs and approximately 3.2 million scored defect decisions, followed by staged validation on a broader set of frontier models. Model choice is the largest determinant of agreement with human labels, but the remaining factors also affect performance, interact with model selection, and can change the best configuration. Within the evaluated design space, a compact six-view RGB protocol performs comparably to denser multi-view settings and inputs augmented with depth or surface normals, making it a strong cost-effective default. Under this standardized pipeline, the best of 12 VLM judges still lag behind trained human labelers, while texture agreement drops sharply when expert-consensus labels are replaced by noisier silver labels. These findings show that automated judges should be evaluated as complete pipelines and calibrated across human reference regimes, rather than benchmarked only as standalone models. We release labels, prompts, predictions, and Croissant metadata on Hugging Face.

Vision language models (VLMs) have made remarkable progress in visual reasoning during the last decade. Most evaluations have used simple scenes (MS-COCO) that do not showcase complex human interactions or behaviors, only a handful of non-curated human descriptions as a benchmark, and have not focused on understanding the model's error types. Here, we introduce the Complex Social Behavior (CSB) dataset, containing 100 images depicting complex social interactions/behaviors. We analyze the progression of scene descriptions over a decade (2017-2025) of VLMs (four pre-Multimodal Large Language Models, MLLMs, and five MLLMs). We evaluate the accuracy of the models and 20 human descriptions relative to a gold standard on the CSB dataset and on a sample from MS-COCO. We analyzed five visual-cognitive error types: object detection, recognition, hallucination, scene understanding, and spatial dependence. The CSB dataset showed a more pronounced improvement than MS-COCO in scene description accuracy, with pre-MLLMs achieving much lower accuracy than the bottom-ranked human descriptions and MLLMs attaining accuracies similar to the top-ranked human descriptions. We show that MLLMs have eliminated the gap in scene description accuracy between simpler MS-COCO scenes and scenes depicting complex behaviors (CSB). MLLMs have almost eliminated all error types in our tested datasets, except for occasionally relying on different image regions for scene descriptions than humans do (spatial dependence error). We also show that detection, recognition, and hallucination errors have the highest impact on scene description accuracy. Together, our findings provide a more thorough evaluation of how visual language models have advanced over the last decade.

Foundation models are reshaping computational pathology, yet their capabilities remain shaped by pretraining objectives, data sources, and spatial scales, fragmenting complementary expertise across separate backbones. Here we present ALICE, a unified foundation model trained through multi-stage agglomerative distillation that sequentially distills eight vision-only, vision-language, and slide-level teacher models into dedicated modules of a single backbone. ALICE is pretrained on 24,985,184 tile-level pathology images and 155,604 high-resolution images, and evaluated across 21 task scenarios, 96 downstream tasks, and 48 data sources, spanning region-of-interest tissue analysis, vision-language multimodal evaluation, and whole-slide clinical assessment. In all three evaluation settings, ALICE achieved the best average rank among task-matched pathology foundation models. These results demonstrate that agglomerative distillation can consolidate complementary capabilities from specialized models into a unified backbone for broad computational pathology applications. The model is available at https://github.com/WonderLandxD/ALICE.

Pre-trained Vision-Language Models (VLMs) such as CLIP achieve strong zero-shot generalization, but their performance degrades sharply under adversarial perturbations. Existing test-time adaptation methods typically rely on sample-level confidence heuristics, overlooking the intrinsic distributional structure of the data. This sample-centric approach limits robustness, as it fails to distinguish confident adversarial mispredictions from true semantic consistency. In this work, we observe that adversarial distortion is structurally brittle: while holistic representations are corrupted, semantic integrity is often preserved in the distribution of augmented views. Motivated by this insight, we propose RITA, a Robust test-tIme prompt-TAdaptation framework that shifts from sample-level estimates to distribution-level alignment. Specifically, RITA employs optimal transport to align the distribution of augmented visual features with textual prototypes, mitigating adversarial outliers and rectifying cross-modal semantic misalignment. Furthermore, we introduce a dynamic cache to progressively accumulate reliable cues from the test stream for online refinement. Extensive experiments demonstrate that RITA significantly improves adversarial robustness without compromising clean accuracy.

Large Language Models are reshaping how software is developed and maintained. They are typically deployed in production using inference engines such as vLLM, which can efficiently serve pre-trained, highly configurable models. While prior work has focused on model architectures and hardware acceleration, the impact of inference engine configuration on energy consumption, performance, and output quality remains poorly understood. In this paper, we present a large-scale controlled study of three selected vLLM configuration options: attention kernel type, prefix caching, and chunked prefill. We evaluate all combinations of these configurations across 5 open-weight LLMs and 5 diverse inference tasks, totaling $9,000$ runs and $93,600$ measures. We analyze energy consumption, latency, and accuracy, and examine both main effects and interaction effects between configuration options and tasks. Our results show that the studied configuration options significantly impact energy and performance, mainly driven by attention type and prefix caching, while chunked prefill has a limited effect under the default vLLM serving configuration and evaluated workloads. These effects are highly model- and workload-dependent, and no configuration is universally optimal. We further show that model choice dominates global trade-offs, while configuration tuning provides local improvements along the Pareto frontier. Unexpectedly, inference options can also affect model accuracy.

Vision-language alignment, the stage that bridges pretrained vision encoders and large language models, is widely treated as a form of pretraining requiring full-parameter updates. We challenge this view and investigate what happens when low-rank adaptation is applied to the LLM during this stage instead. We find that low-rank alignment not only reduces computational costs but also outperforms full-parameter alignment on most benchmarks. To understand this phenomenon, we systematically characterize the implicit biases introduced by low-rank adaptation during alignment. Empirically, we find that low-rank alignment shifts model behavior from hallucinatory to conservative and preserves per-token linear separability of visual features that full-parameter alignment disrupts, a phenomenon we term LS-curse. Geometrically, low rank aligned models exhibit more homogeneous and structurally stable visual representations, maintaining modality-specific knowledge rather than prematurely fusing entity-level semantics. Theoretically, we establish two theorems showing that low-rank alignment induces preferences for parameter subspaces with flat gradients and feature subspaces robust to perturbations, providing a principled explanation for the observed structure-preserving behavior. Extensive experiments cover ablation over 100 alignment configurations, three families of low-rank operators, and various rank, encoder, and other settings.

Long-horizon robot planning requires jointly reasoning over semantic task structure and geometric feasibility. To successfully execute a task, a robot must decompose goals, select task-relevant objects, and sequence actions, while ensuring that plans satisfy spatial constraints such as limited free space and object collisions. In this work, we propose APIVOT, a VLM-based planner that adaptively interleaves language and visual thoughts for long-horizon planning. APIVOT learns to leverage language for semantic reasoning, while using visual thoughts as imagined future states for internal verification of geometric feasibility. On long-horizon kitchen tasks, APIVOT outperforms general-purpose VLMs and prior planning frameworks, achieving the largest gains in spatially constrained settings. We find that APIVOT learns meaningful modality selection behavior, demonstrating that adaptive interleaving of vision-language thoughts improves both planning success and reasoning efficiency.

Traditionally, continual learning has assumed access to labeled data, yet many real-world applications -- such as lifelong robotics -- require models to adapt continuously from unlabeled streams. This has led to the development of continual self-supervised learning (CSSL), a rapidly growing area that lacks a dedicated, systematic review. In this work, we present a comprehensive survey of CSSL for vision, with connections to emerging vision-language settings. First, we analyze existing evaluation protocols and highlight inconsistencies that hinder fair comparison. We then examine why self-supervised objectives exhibit improved robustness to catastrophic forgetting, relating this to task-agnostic representations and smoother loss landscapes. Next, we organize existing methods into a unified taxonomy based on their forgetting-mitigation strategies, including distillation, replay, regularization, architectural approaches, model merging, and objective-level adaptation. Finally, we identify open challenges such as scalability and the need for fast adaptability. We argue that advancing CSSL requires moving beyond small-scale benchmarks towards continual pre-training paradigms for large-scale systems.

In this report, we present Pelican-VLA 0.5, a unified VLA model that integrates vision-language understanding, future-frame generation, and action prediction within a single architecture. Pelican-VLA 0.5 achieves attention-level generalization: without object annotations, segmentation masks, attention supervision, or task-specific fine-tuning, its action pathway already focuses on the manipulation-relevant object and contact region. This behavior persists across unseen scenes and unseen robot embodiments, and is substantially stronger than in other open-source VLA baselines. We verify that this ability originates from the learnable Bottleneck Token inserted between perception and action: by routing task-relevant visual information through a compact bottleneck, the tokens interface induces manipulation-centric attention during pre-training and remains effective across different policy structures, including a MoT-style architecture.

3D dense captioning, an emerging vision-language task, aims to generate descriptive sentences for each object in the 3D scene. Despite the impressive results achieved by previous methods, they suffer from two limitations. First, current research often employs global rigid transformations, such as rotation, to augment scenes without changing their spatial layouts. However, diverse spatial layouts are crucial for training a 3D dense captioning model to describe spatial relations between objects. Second, previous works mainly focus on the design of the caption generation pipeline while utilizing a simple network architecture for other components, i.e., backbone and detection head, which is crucial for extracting rich semantic information for captioning. In this paper, we propose PVCap to alleviate the aforementioned problems. Our PVCap consists of PseudoCap and VoxelCapNet. Specifically, PseudoCap employs a random mixing technique on instances within the dataset, generating numerous pseudo frames with diverse spatial layouts at the instance level. By utilizing a teacher-student framework, PseudoCap obtains pseudo caption labels for these pseudo frames. This data augmentation approach significantly increases the number of training samples and enhances the model's ability to describe the environment effectively. Regarding VoxelCapNet, we introduce a robust caption network that utilizes voxel features and adapts the caption head to the voxel-based network architecture. Our VoxelCapNet can serve as a competitive baseline for future research on 3D dense captioning. Extensive experiments are conducted on two prevalent benchmarks, i.e., ScanRefer and Nr3D. Notably, our method surpasses current state-of-the-art by 11.41% and 13.99% in CIDEr@0.5IoU, respectively. Codes will be made publicly available.

Open-vocabulary 3D scene understanding aims to segment 3D scenes beyond predefined categories by transferring semantic knowledge from vision-language models. Existing methods have advanced this task by lifting language-aligned 2D features into 3D, yet they often rely on context-independent semantic representations, leaving object relationships underexplored for contextual refinement. We propose RelGraphOV, a relationship-aware framework that uses 3D scene graphs to enhance open-vocabulary 3D understanding. Our method constructs relational scene graphs from multi-view observations by leveraging vision-language reasoning to infer object relationships and prune geometrically implausible connections, without manual relationship annotations. To aggregate relational context while avoiding feature interference, we introduce an Adaptive Gated Dual-Stream Contextual GAT that separates dense geometric features and semantic CLIP embeddings, performs edge-guided message passing, and adaptively fuses complementary semantics. A hierarchical contrastive objective further promotes instance-level consistency and category-level discrimination. Experiments on ScanNetV2, ScanNet200, ScanNet$++$, and Replica demonstrate strong performance and generalization ability. Project Page: https://cxavireh.github.io/relgraphov-projectpage

Adverse driving conditions, such as bad weather, remain a principal barrier to autonomous driving because they degrade two things at once: what the vehicle can perceive and what it can physically do. Human drivers cope by anticipation, reasoning about the scene and re-budgeting speed, following distance, and steering before grip or sight is lost, whereas current autonomous driving systems at best react after the fact. This paper proposes VLM-CASE, a framework that gives an autonomous vehicle this anticipatory capacity while keeping its motion bounded by a formal safety model at all times. A vision-language model (VLM), fine-tuned with low-rank adaptation (LoRA), reasons about the scene from the front-camera image and reports the road surface and visibility conditions. This output parametrizes a context-adaptive safety envelope (CASE), derived from physical limits and the guarantees of responsibility-sensitive safety, that couples braking and steering through a shared friction budget. A model predictive controller then drives freely within the envelope, while the VLM runs asynchronously so it never blocks the real-time control loop. We validate the framework in closed-loop CARLA simulation on tasks that demand both lateral and longitudinal control, across a range of weather, road-surface, and lighting conditions. The resulting controller, VLM-CASE-MPC, completes all trials, outperforming a conventional MPC baseline and a state-of-the-art VLM-integrated controller. Ablations confirm that the gains come from context adaptation, with the friction and visibility adaptations proving complementary. Furthermore, the framework is controller-agnostic and pairs with almost any low-level controller, offering a promising direction for safe autonomous driving. The dataset and supplementary materials for VLM-CASE are available at https://github.com/ytj254/VLM-CASE.

Reasoning segmentation aims to predict pixel-wise masks for targets given complex language queries. Existing approaches leverage Multimodal Large Language Models (MLLMs) for vision-language reasoning and generate intermediate target cues (e.g., points or boxes) to guide a segmentation model. However, compressing rich reasoning into sparse cues often introduces ambiguity and noise, preventing these cues from accurately preserving the reasoning intent. While multiple complementary cues can enrich target information, existing methods typically feed them jointly into a single segmentation process, allowing ambiguous or erroneous cues to affect the entire prediction. Therefore, we propose DGSeg, a reasoning segmentation framework that learns to fuse predictions guided by semantic and spatial cues. Specifically, the MLLM jointly reasons about both target identity and spatial location, producing complementary semantic and spatial cues that are fed into separate segmentation branches. Their predictions are adaptively integrated by a lightweight dynamic gating module trained with relative branch-quality supervision to suppress noisy or conflicting regions. Extensive experiments demonstrate that DGSeg consistently outperforms strong baselines on multiple benchmarks and achieves 69.6% and 67.3% gIoU on the challenging ReasonSeg validation and test splits. Code is available at https://github.com/RZZeng/DGSeg.

Open-vocabulary object detection aims to localize and classify objects beyond the fixed set of categories seen dur ing training. Recent open-vocabulary object detection methods improve localization and classification for unseen categories by leveraging a frozen VLM as a detector backbone. However, VLM classification score lacks recognizing position and scale of the object in an image. We observe that pretrained VLMs en able to classify foreground and background regions. According to this observation, we propose a simple inference-time Pro posal Calibration (ProCal) that improves localization quality of the classification score. ProCal computes a proposal prior by combining two scores: localization-aware foreground score and background-aware suppression score. Localization-aware foreground score captures whether a proposal contains an object area. Background-aware suppression score measures the extent to which the proposal resembles background. We analyze that ProCal suppresses false novel activation on background proposals and consistently ranks true novel proposals above background and partial novel proposals. Applied to CLIPSelf ViT-L/14, ProCal improves APr +2.5 on OV-LVIS. The analyses show that proposal-level localization-aware reranking effects to mitigate ranking miscalibration for novel objects.

Video world models are emerging as a scalable alternative for evaluating generalist robot policies, bypassing the physical constraints and engineering burdens of real-world deployment. However, evaluating policies with video world models remains challenging, as world-model errors can make generated rollouts unreliable and slow inference limits large-scale throughput. We introduce RoboWorld, an automated evaluation pipeline that pairs a fast autoregressive video world model with a task-progress-aware vision-language model scoring. To enable reliable long-horizon autoregressive world-model rollouts, we propose Step Forcing, which combines anchored and one-step self-forwarded contexts to reduce train-test mismatch while preserving action-observation dynamics. Together, these components enable RoboWorld to align strongly with real-world robot evaluation across tasks and environments, achieving Pearson's r = 0.989 and Spearman's $ρ$ = 0.970.

We present DroneIQA-VLE, our solution to the ICME 2026 Drone-IQA Grand Challenge on Target-aware Image Quality Assessment for Low-altitude UAV Images. The framework jointly predicts global, target, and background quality scores by ensembling two complementary pipelines: (1) SigLIP2 vision encoders with multi-task regression heads, and (2) a LoRA-adapted Qwen3.5-9B multimodal large language model for quality score regression. The final global quality prediction is obtained by arithmetically averaging the outputs of both pipelines. Our method achieves 2nd place in the challenge, demonstrating its effectiveness. The code is available at https://github.com/sunwei925/DroneIQA-VLE.

In collaborative dialogue, shared perception does not guarantee shared interpretation. Mutual understanding must be established through interaction. We investigate whether vision-language models (VLMs) can distinguish what could be shared from what has been shared between dialogue participants through grounding. We formulate this as an interpretation-matching task on 13,077 annotated reference expressions from HCRC MapTask dialogues, and evaluate VLMs under systematically controlled manipulations of dialogue context and map-information access. Our results show that providing authentic map images improves overall performance but shifts models toward over-predicting alignment. Textual descriptions of the same map content reproduce this bias, while non-informative images suppress alignment predictions entirely, indicating that the bias is driven by task-relevant map content, not the visual channel. This improvement comes at the cost of degraded accuracy on non-aligned cases. Calibration analysis and reference-chain tracking further suggest that models rely on static referential cues on the maps rather than tracking how grounding unfolds through dialogue history. We observe these patterns most clearly in Qwen3-VL-8B-Instruct and, to varying degrees, in four additional models from two architecture families. In models that exhibit the bias, map content, whether presented visually or textually, is treated as evidence of mutual understanding, conflating potential with established common ground.

Generalizable robotic grasping in cluttered environments is essential for deploying manipulators in unstructured human spaces, yet existing VLM-based methods rely on visual similarity for object matching, neglecting physical affordances such as handle graspability and material fragility, and operate open-loop without spatial reasoning or failure recovery, limiting their effectiveness when objects are densely packed or physically diverse. We present Agentic RAG-VLM, a unified framework that bridges VLM-based semantic understanding and physically grounded grasp execution by integrating retrieval-augmented generation (RAG) with vision-language models (VLMs) and agentic self-reflective planning. Agentic RAG-VLM introduces three tightly coupled components: (1) a Hierarchical Affordance-Aware RAG (HAA-RAG) that encodes four-dimensional affordance descriptors, including type, material, fragility, and graspable region, and retrieves strategies by functional affordance compatibility rather than visual appearance; (2) a Scene Graph Constraint Reasoner that constructs spatial relationship graphs from VLM perception and translates proximity, occlusion, and support constraints into concrete grasp parameter adjustments; and (3) an Agentic Self-Reflective Pipeline with a 14-type failure taxonomy and three-level adaptive retry for closed-loop grasp refinement. Evaluated on a 12-task benchmark spanning single-grasp, interactive, and long-horizon scenarios with 360 trials per configuration, Agentic RAG-VLM achieves 78.3 percent overall success, a 53.3 percentage-point absolute gain over VLM-only baselines, demonstrating that affordance-aware retrieval, scene graph reasoning, and agentic recovery are jointly essential for robust manipulation.

Vision-Language Models (VLMs) provide powerful semantic understanding and commonsense reasoning for End-to-End Autonomous Driving (E2E-AD) planning. However, trajectories directly generated by VLMs often encode only coarse driving intentions and remain insufficient for geometrically accurate, future-aware, and multi-view-grounded planning. To address these limitations, we develop the Layer-Wise World-Model-Guided Driving framework (LWDrive). LWDrive is a VLM planning framework that refines coarse trajectories through layer-wise world-model guidance. Instead of treating the VLM output as the final trajectory, LWDrive uses it as an intent-aware coarse plan, expands a diverse candidate space around it, and progressively refines the candidates through a Foresight Cascade Planner (FCP). Specifically, we introduce future-frame generation supervision to encourage the VLM to learn forward-looking scene representations, thereby injecting planning-relevant predictive dynamics into its internal hidden states. Built upon these world-model-supervised representations, FCP exploits VLM features across multiple layers and integrates historical temporal states, Action-Query representations, and current-frame multi-view Bird's-Eye-View (BEV) features to refine candidate trajectories in a coarse-to-fine manner. This design enables progressive correction of spatial positions and motion trends while grounding trajectory refinement with multi-view scene cues and preserving the high-level driving intention produced by the large model. Finally, a score head evaluates the refined candidates and selects the best trajectory as the final planning output. Experiments show that LWDrive achieves a score of 92.0 on the NAVSIM benchmark and 89.6 on NAVSIM-v2. Code and models will be made publicly available.

Large Vision-Language Models (VLMs) are increasingly deployed as content moderation tools, yet they remain vulnerable to jailbreak attacks in which harmful text is visually encoded as ASCII art. This can allow inappropriate or harmful content to bypass moderation systems. To address this vulnerability, this paper investigates how image resolution affects VLM detection of harmful ASCII art across eight character construction modes (L1-L8), ranging from dense block characters to word-embedded designs. We evaluate eight state-of-the-art VLMs on English and Chinese corpora using a pipeline that generates ASCII art images at ten resolution scales, probing whether a consistent detection-failure threshold exists across models, modes, and languages. Results indicate that detection rates decline sharply above certain resolution thresholds, and that word-based modes are the most resistant to detection across the full resolution range. These findings reveal a systematic vulnerability in VLM-based content moderation systems and motivate resolution-aware evaluation standards.

Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies. We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training. As part of DCVLM, we collect 160 datasets spanning four data types -- image-caption pairs, multimodal interleaved documents, text-only, and instruction-tuning data -- into a corpus of 6T multimodal tokens. DCVLM allows participants to test curation strategies (filtering, mixing, formatting, sampling) across 1B-8B models and 6.25B-200B token budgets. Models are then evaluated on a carefully selected suite of up to 52 downstream benchmarks across 9 domains. We conduct extensive experiments on DCVLM and find that data mixing, not filtering, is key to a high-quality training dataset: instruction-heavy mixtures scale better than caption-heavy ones, with gains widening at larger scales. The resulting dataset, DCVLM-Baseline, enables training an 8B VLM to 63.6% accuracy on our 33-task core suite with 200B training tokens. Compared to FineVision, the state-of-the-art open VLM training dataset, this represents an improvement of +5.4pp. DCVLM and all accompanying artifacts will be made publicly available at https://www.datacomp.ai/dcvlm/.

Concept Bottleneck Models (CBMs) promise transparent reasoning by predicting through human-interpretable concepts, yet their effectiveness fundamentally depends on how well visual and textual representations are aligned or matched. Existing vision-language CBMs often rely on pre-aligned encoders or global cosine similarity, which obscures fine-grained concept localization and fails to reflect true semantic geometry. In this work, we rethink concept alignment as a dynamic cross-modal transport process instead of static projection and propose the Optimal Transport Flow Concept Bottleneck Model (OTF-CBM). It first learns a data-driven semantic cost via Inverse Optimal Transport to measure cross-modal distances, and then performs unbalanced optimal-transport-based flow matching to model semantic transitions between visual patches and textual concepts. With velocity-based concept activation, OTF-CBM captures interpretable geometric relations without ODE integration. Experiments further show that OTF-CBM achieves superior classification accuracy and concept faithfulness, offering a new geometric and dynamical perspective for interpretable cross-modal reasoning.

Multimodal Large Language Models (MLLMs) have recently made remarkable progress in unifying vision-language understanding and reasoning, especially following the introduction of models such as OpenAI's O-series and DeepSeek's R-series, which have driven a paradigm shift toward perception-centric intelligence. However, there remains a lack of systematic surveys that examine perception from a truly unified vision-language perspective -- one that treats vision and language as an inseparable modality. Existing reviews are often fragmented, focusing separately on either vision or language, and thus rarely capture the cross-modal evolution of perception as an integrated capability. To bridge this gap, we present the first systematic survey of unified vision-language perception in MLLMs. Specifically, we (1) formalize MLLM perception as an intrinsic, unified vision-language capability analogous to human innate perception, (2) introduce a five-stage taxonomy tracing the paradigm evolution of MLLM perception and survey representative methods and milestones at each phase, and (3) identify open challenges and outline promising research directions toward truly general, unified multimodal intelligence. We hope our study will provide both a foundational understanding and an actionable roadmap to foster further innovation on the path toward artificial general intelligence (AGI).

Production vision pipelines silently degrade on blurry input, wasting compute on downstream OCR, retrieval, and vision-language model (VLM) calls that cannot recover a usable output. We present MagikaDocumentFromPixel, a lightweight, CPU-friendly image quality gate that classifies a single image as sharp, blurred, or uncertain in roughly 7 ms on a single CPU core. The contributions are (i) a recipe selected from a 46-configuration, 8-sweep empirical search that isolates input resolution as the dominant lever and shows architecture capacity only pays off at >= 384 px; (ii) a confidence-aware routing formalism grounded in classical selective prediction; (iii) the Edge Prior Module (EPM), a Laplacian-magnitude auxiliary input channel that gives the network direct access to the spectral evidence that classical blur heuristics rely on and that lifts test F1 by +1.3 points in a matched-env comparison; and (iv) an observation that the gate is one instance of a recurring design pattern that appears independently in Magika content-type detection, risk-controlled OCR with VLMs, and DocVLM. The final recipe MobileNetV3-Large with the EPM trained at 384x384 on paired GoPro Large frames, evaluated with 5-scale test-time augmentation reaches F1 = 0.9803 (AUC 0.9989) with a 17 MB ONNX artifact, improving over our fixed-scale baseline on the same hardware (F1 = 0.9672) by +1.31 points. We are explicit about limitations: results are on a single motion-blur distribution, numbers are from a single seed, and calibration is qualitative rather than measured.

Vision-language pre-training (VLP) holds great promise for general-purpose medical AI by leveraging radiology reports as rich textual supervision, yet existing methods struggle with 3D CT imaging due to inefficient visual backbones and coarse semantic alignment. To address these issues, we propose a tailored VLP framework featuring three key components: (1) a CNN-ViT hybrid encoder that replaces ViT's patch embedding with a 3D CNN backbone to efficiently capture local anatomical details while preserving global attention and compatibility with pre-trained cross-modal priors; (2) a disease-level contrastive learning mechanism using learnable query tokens to dynamically extract disease-specific semantics from full reports and align them with corresponding visual features, thereby disentangling distinct diseases within the same anatomical region; and (3) a diagnosis-aware prompt strategy that employs real clinical phrases and aggregated disease prototypes to bridge the pre-training-inference gap and enhance zero-shot diagnostic reliability. Our model achieves state-of-the-art performance on CT-RATE (84.4% AUC, +5.1%) and Rad-ChestCT (75.4% AUC, +5.4%), with even larger gains (+9.8% AUC) on a challenging 60-disease benchmark, and demonstrates strong transferability to radiology report generation, underscoring the generality and clinical utility of our approach.

Vision Language Models (VLMs) have shown promising capabilities in medical image analysis by jointly understanding visual and textual information for tasks such as Visual Question Answering. However, existing hematology vision-language resources remain predominantly English centric, limiting their applicability in multilingual healthcare environments. This challenge is releveant generally to South Asia and specifically to Pakistan, where Urdu is widely used despite healthcare information and digital medical systems being largely dependent on English. To investigate this gap, we conducted a survey among healthcare professionals, which revealed substantial language mismatches between clinical documentation and patient communication, emphasizing the need for multilingual healthcare technologies. To address this limitation, we introduce WBCMor VQA, a clinically validated bilingual English, Urdu morphology aware VQA benchmark for leukemia and normal white blood cell analysis. The benchmark is constructed using morphology-aware annotations from LeukemiaAttri and WBCAtt datasets and supported by a domain specific Urdu hematology dictionary to ensure linguistic consistency and clinical correctness. The final benchmark contains 110K bilingual question answer pairs serving as VQA annotations for 20K leukemic and normal single-cell images. Furthermore, we establish baseline performance by evaluating multiple open-source VLMs on the proposed benchmark. The proposed resource aims to facilitate the development of accessible and clinically relevant AI systems for multilingual healthcare environments.

Vision-Language Large Models (VLLMs) trained on massive crawled corpora raise pressing copyright and data-provenance concerns. These concerns are particularly acute in healthcare, where patient medical images paired with clinical reports demand rigorous privacy safeguards. However, existing training data detection methods either fail in cross-modal scenarios or rely on superficial output signals with insufficient discriminative power. We introduce GradAudit, a gradient-based auditing framework that examines internal optimization dynamics rather than treating VLLMs as black boxes. Our approach builds on a key observation: model parameters converge to regions where gradients on training samples become stable and well-aligned, whereas gradients on non-training samples remain noisy and inconsistent. By analyzing these gradient signatures, GradAudit achieves strong separability and detects genuine image-text associations learned during training, not merely individual modality membership. Empirically, across both medical and general-domain datasets, GradAudit substantially outperforms state-of-the-art baselines in both pretraining and fine-tuning VLLMs. In a case study employing copyrighted content, we show that existing training data detection methods not only underestimate the extent of unauthorized data usage, but that this underestimation becomes more pronounced as models become more recent and more advanced.

Vision-language contrastive pretraining has become the dominant recipe for 3D medical foundation models, leveraging the large volumes of paired scans and reports produced in clinical practice. However, medical images usually span dozens of organs, and radiological reports are much longer than typical natural image captions and are composed of multiple structured sections. CLIP-style pretraining compresses this structure by encoding each modality into a single global token, at the risk of losing important details. We introduce ConQuer (Concept Queries), an image-text pretraining method that augments CLIP's global alignment with a set of localized alignments, one per concept. ConQuer splits the report into concept-specific sections and learns cross-attention queries that pool the matching image features without using any segmentation mask or spatial supervision. Contrastive learning is then applied independently for each concept. Concepts can be any unit of semantic localization; here, they are anatomical regions, one query per organ or gross body region. As a byproduct, each query learns attention maps focused on its concept, providing built-in spatial interpretability. We use ConQuer to train Jolia, a 3D CT foundation model on chest and abdominal CT. Jolia consistently outperforms a CLIP baseline on findings classification, report generation, and cross-center transfer, and sets a new state of the art across multiple public benchmarks. Jolia's weights are available at https://huggingface.co/raidium/Jolia

Vision-language-action (VLA) models have the potential for open-world generalization by leveraging pretrained vision-language representations, yet downstream finetuning on limited robot data often degrades these representations, leading to brittle policies that ignore language instructions in favor of visual shortcuts, a failure mode we term instruction blindness. We hypothesize that standard finetuning with limited data applies gradients to a sparse set of points, which manifests as a sharp loss landscape with high-curvature minima. We propose to address this directly through flatness-preserving optimization while finetuning on the exact same data, where learning a flatter landscape results in a model more robust to perturbations in the weight space. Specifically, we demonstrate that simply applying sharpness-aware minimization during VLA finetuning significantly improves instruction following by over 60% across multiple simulation and real-world benchmarks without additional data, architectural modification, or retraining. We further analyze the effect of selective sharpness, quantify its effects, and show that our approach is complementary to existing guidance techniques. Project page can be found at https://haochenz11.github.io/papers/flatness-vla/.

The availability of large amounts of clean data is paramount to training neural networks. However, at large scales, manual oversight is impractical, resulting in sizeable datasets that can be very noisy. Attempts to mitigate this obstacle to producing performant vision-language models have so far involved heuristics, curated reference datasets, and using pre-trained models. Here we propose a novel, bootstrapped method in which a CLIP model is trained on an evolving, self-selected dataset. This evolving dataset constitutes a balance of filtered, highly probable clean samples as well as diverse samples from the entire distribution. Our proposed Self-Filtering method iterates between training the model and selecting a subsequently improved data mixture. Training on vision-language datasets filtered by the proposed approach improves downstream performance without the need for additional data or pre-trained models.

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for generalist robot manipulation, yet they lack geometric inductive biases: policies trained at specific orientations require substantially more data to generalize across rotational configurations. We present \textsc{EquiVLA}, the first general framework for end-to-end $\mathrm{SO}(2)$-equivariant VLA models, applicable to any architecture coupling a frozen vision-language backbone with a flow-matching Diffusion Transformer action head. \textsc{EquiVLA} introduces \textsc{EquiPerceptor}, which produces approximately $\mathrm{SO}(2)$-equivariant visual representations from frozen ViT features; and \textsc{EquiActor}, an exactly $\mathrm{SO}(2)$-equivariant flow-matching Diffusion Transformer action head. Together, they establish an approximate $\mathrm{SO}(2)$ equivariance chain from camera observations to predicted action sequences. Instantiated on GR00T~N1.5 and evaluated across four LIBERO suites, CALVIN ABCD$\to$D, and five real-robot tasks on Mobile ALOHA, \textsc{EquiVLA} achieves $92.6\%$ average success on LIBERO (vs. $78.1\%$ baseline), an average sequence length of $4.03$ on CALVIN (vs. $3.45$), and improves real-robot success from $54\%$ to $72\%$.

Embodied Vision-Language-Action (VLA) models are typically obtained by fine-tuning powerful pretrained VLMs on robotics data, yet it is unclear how much commonsense and factual knowledge they retain after adaptation. Failures on knowledge-sensitive tasks are ambiguous, conflating missing knowledge with poor generalization of low-level control. We introduce Act2Answer, a lightweight protocol that adapts VLM knowledge benchmarks to VLA evaluation by requiring agents to answer through action. Each question becomes a short tabletop episode where the agent performs a single object-placement action to select among candidate answers, yielding an action-grounded success rate with reduced control confounds. We curate a test suite of such environments across diverse commonsense and world-knowledge categories and introduce layerwise intent probing to localize answer-relevant information across the VLM backbone and action head. In a large-scale study of 7 VLA models and 9 VLM baselines, we systematically rank models across categories, finding that VLAs show solid performance on simple concepts while exhibiting larger gaps on richer semantic categories relative to their source VLMs, that VQA co-training is associated with better knowledge retention, and that answer-relevant signals peak in middle VLA layers but attenuate in upper layers. Act2Answer is available at https://tttonyalpha.github.io/act2answer/.

Open Vocabulary Action Recognition (OVAR) enables the recognition of novel actions by leveraging vision-language representations, overcoming the limitations of traditional closed-set approaches. However, achieving robust performance in real-world scenarios typically requires domain-specific fine-tuning, which is often costly and raises privacy and regulatory concerns. In this work, we propose an alternative paradigm that bypasses target-domain training and recombines knowledge from existing datasets and models. Leveraging model merging and task arithmetic, we extract and combine task vectors from models fine-tuned on diverse public OVAR datasets. We show that, in out-of-distribution settings, the resulting merged model achieves superior zero-shot generalization to the pre-trained base model. Code is available at https://github.com/omaymaMoussadek/robust-ovar

Vision-language models (VLMs) are now widely used in downstream tasks. However, real-world applications often expose VLMs to distribution shifts induced by semantic variation (e.g., shape, size, and style). Robustness certification determines if a model's prediction changes when transformations are applied to its input. While most certification frameworks study geometric or pixel-level transformations over inputs, this work proposes a novel framework that enables certifying VLM robustness under semantic-level transformations. Leveraging the open-vocabulary capability of VLMs, we use text prompts as semantic proxies to construct transformations parameterized by an extent that controls the degree of semantic variation. By characterizing the VLM decision boundary in closed form, our framework quantitatively certifies extent intervals for which the predicted class remains unchanged under the semantic transformation. Our framework is the first to certify VLM robustness under semantic-level variations without requiring additional data for each variation, making it practical to apply. Experiments on both synthetic and real-world data show that our framework enables certifying robustness under diverse semantic variations across scenarios.

To locate a target object while exploring the unknown environment is a fundamental capability for autonomous agents, with applications ranging from search-and-rescue to field robots. A simplified version of such task is Object Goal Navigation (ObjNav). In ObjNav, successful arrival at the target object provides a basic measure of performance; however, the efficiency of the navigation trajectory is equally important, as it indicates how intelligently the agent explores and how much time remains for subsequent tasks. In unknown environments, the key to efficient navigation lies in deciding where to explore next. While many prior works aim to address this core challenge and achieved promising performance in certain settings, recent training-based models and non-training frameworks still suffer from generalization and efficiency issues respectively, which in the worst cases can lead to excessive exploration of already-visited areas or redundant back-and-forth motion. We evaluate EffiNav on two widely used simulation benchmarks Habitat Matterport 3D (HM3D) and Open-Vocabulary Object goal Navigation (OVON), and further validate its effectiveness on physical robots in real-world settings. We conduct failure analysis on massive simulation episodes. With minimal modification, we also extend EffiNav to a memory-augmented ObjNav task on the GOAT-BENCH dataset, demonstrating its adaptability beyond standard ObjNav settings. Across two standard metrics--Success Rate (SR) and Success weighted by Path Length (SPL), EffiNav matches or outperforms recent baselines, reflecting its efficiency, robustness, and practical applicability. Recognizing the different emphases of the two datasets, the performances reveals this framework is more balanced and generalizable for efficient ObjNav.

Visual information helps resolve ambiguity in coreference resolution, leading to notable performance gains. However, existing Multi-modal Coreference Resolution (MCR) methods require training with (partially) annotated data from the target dataset before they can be applied, preventing their direct usability and raising concerns about generalization. While Vision-Language Large Models (VLLMs) with billions of parameters offer promising zero-shot capabilities, they remain largely inaccessible. Their massive size limits deployability, and many are only accessible through paid APIs. In this paper, we propose a plug-and-adapt method that strategically adapts a carefully pre-trained \emph{alignment model} for immediate use in MCR tasks, designed to eliminate the need for training on scarce benchmark datasets or relying on resource-intensive VLLMs. Specifically, we first pre-train a fine-grained alignment model between textual and visual contextual information using vision-language alignment datasets. We then repurpose the alignment model to MCR through similarity aggregation by fusing visual and categorical cues with evidence theory, thereby enhancing effectiveness. Experiments on the Coreference Image Narratives (CIN) benchmark dataset demonstrate the effectiveness of our method, achieving a 5.31% and 2.12% improvement in CoNLL F1 over SOTA dedicated methods and popular VLLMs, respectively. We further evaluate our method on a masked CIN dataset for robustness testing and on a specially constructed VCR-MCR dataset for generalization assessment, with results confirming both capabilities.

Text-guided open-vocabulary object counting (TOOC) aims to count objects belonging to the categories specified by natural language descriptions. Although vision-language pre-trained models have been successful applied to TOOC tasks, they still struggle with fine-grained spatial understanding and real-time inference requirements in counting scenarios. To address these limitations, this paper proposes a real-time TOOC framework, called the Real-Time Counter (RT-Counter), that achieves not only good counting accuracy but also high computational efficiency. RT-Counter designs a novel Visual Prototype Textualization (VPT) module that can project learned visual features into a text feature space and then generate features containing the abstract information that is hard to capture with visual prototypes and the detailed prototype information that is difficult to describe in text, enhancing the object-level visual-language model's counting capabilities. Additionally, RT-Counter incorporates our Weaving Transformer (Weaformer) layers, maintaining high descriptive power at a fraction of the computational cost. The Weaformer layer adopts a novel hybrid attention mechanism that can efficiently weave together local and global visual features. Extensive experiments on three public datasets show that RT-Counter successfully breaks the accuracy-speed trade-off in TOOC. While achieving a competitive MAE of 13.30 on FSC147, RT-Counter operates at 112.48 FPS, making it 7.4x faster and over 4$\times$ more parameter-efficient than the existing leading methods in TOOC. Our work aims at balancing high accuracy and real-time performance in TOOC. Code is available at: https://github.com/Jason-Mar1/RT-Counter.

Multimodal Foundation Models are increasingly used as reasoning agents, making reliability, knowing when a model may hallucinate, critical. A common intuition, which we call the Attention-Confidence Assumption, holds that reliability follows from "structural" visual perception: tight attention on relevant regions should signal a trustworthy answer, while scattered attention signals confusion. We challenge this through the VLM Reliability Probe (VRP), a systematic cross-family study of reliability signals in contemporary Vision-Language Models (VLMs). We introduce structural-attention metrics, cluster counts (C_k) and spatial entropy (H_s), to quantify the visual encoder's gaze, and track its evolution (Delta H_s) across layers. This reveals a "Symbolic Detachment": models often "Early Lock" visual features only to diffuse attention later, severing early perception from final generation. Contrary to the grounding hypothesis, we find a "Cluster Failure": spatial attention has near-zero correlation (R approx 0.001) with accuracy. Instead, reliability is a phenomenon of generation dynamics and internal-state distributions. Self-Consistency, the agreement rate across sampled reasoning paths, is the dominant predictor of truth (R = 0.429). Scaling causal interventions exposes a sharp architectural divergence: LLaVA locks its prediction in a fragile late-stage bottleneck, whereas PaliGemma and Qwen2-VL distribute reliability globally, staying resilient even when ~50% or more of their most predictive layer is destroyed. For current VLMs, reliability signals are detached from visual grounding maps and are best inferred from generation-time dynamics and hidden-state probes.

Current multilingual evaluations for Vision-Language Models (VLMs) assume a one-to-one mapping between language and orthography, overlooking billions of users of multi-script languages. We introduce PuMVR (Punjabi Multimodal Visual Reasoning), a benchmark of 1,000 strictly parallel image-text instances across Punjabi's three active scripts: Gurmukhi, Shahmukhi, and Roman. Evaluating 10 state-of-the-art VLMs, we expose a substantial and systematic Script Gap. Models frequently solve visual tasks in one script while failing identical tasks in another, with accuracy deltas reaching 16%. Crucially, visual input boosts absolute performance uniformly yet does not close the orthographic gap. Furthermore, cross-script in-context transfer is highly brittle, exposing script-locked knowledge representation. Supported by McNemar tests across all script pairs, our findings demonstrate that current "multilingual" VLMs are not truly multi-script. We propose the Script Consistency Rate (SCR), which falls as low as 24.8% on our benchmark, as a mandatory metric for script-agnostic evaluation to ensure equitable AI access. Data and code are available at: https://github.com/prabhjotschugh/Not-Truly-Multilingual-PuMVR.

Vision-language models (VLMs) have sparked growing interest in zero-shot Earth Observation (EO) downstream tasks, with further gains enabled by remote-sensing-adapted models. We examine this setting across 17 VLM variants and 12 remote sensing (RS) datasets under Meta-Prompting for Visual Recognition (MPVR), and show that zero-shot performance remains highly sensitive to textual design choices, from the meta-prompts used to guide the LLM in generating class descriptions to the descriptions themselves. We explore why semantically rich LLM-generated class descriptions do not translate into consistent gains over simple domain-adapted CLIP-style descriptions. While LLM descriptions are more semantically expressive, they can also introduce noise in the text embedding space, reducing robustness in downstream tasks. We support this observation through a text log-likelihood analysis in the whitened CLIP feature space, comparing LLM-generated and template-based descriptions. Building on this finding, we study query embedding calibration and show that lightweight calibration of the query space consistently yields strong improvements in zero-shot classification and retrieval. Overall, our results provide practical insight into the trade-off between semantic richness and robustness, and identify embedding calibration as a simple and effective tool for improving zero-shot remote sensing performance.

Vision Language Model (VLM) has great potential to enhance the quality of pseudo labels in semi-supervised spine segmentation by leveraging textual class prompts to generate segmentation map, but no one has studied it yet. Although promising, it lacks explicit constraints to ensure consistency between spine class prompts and spine unit region, resulting in unsatisfactory performance in multi-class segmentation map generation. In this paper, we propose CPS4, the first text-guided semi-supervised spine segmentation network using class prompts to enhance the quality of spine pseudo labels. Specifically, CPS4 is implemented through two training stages. (i) Class-specific consistency constrained VLM pretraining stage: we propose token- and pixel-level attention loss to optimize the consistency between class prompts and spine units, forcing the textual class prompt to be closely coupled with the target spine unit in the semantic space. (ii) Class Prompt driven semi-supervised spine segmentation stage: using the pretrained vision-text encoder, we derive each class-specific binary segmentation map for the unlabeled spine image and integrate them into an unified multi-class segmentation map, improving the quality of the spine pseudo label generated by the semi-supervised spine segmentation network. Experimental results show that our CPS4 achieves superior spine segmentation performance with Dice of 80.44%, only using 5% labeled data on the public spine segmentation dataset, surpassing popular semi-supervised learning and VLM methods. Our code will be available.

Continual vision-language models are commonly addressed through sequential fine-tuning; however, although this paradigm enables adaptation to new environments (tasks), it inherently emphasizes the contribution of previously learned environments (tasks) at the expense of the stability required to preserve previously acquired knowledge. While existing approaches have adequately studied continual learning and catastrophic forgetting in vision-language models (VLMs), the theoretical understanding of modality-specific contributions across a sequence of environments remains largely unexplored. In this paper, we present a new theoretical perspective to understand the cross-modal (vision-language) contributions to consecutive environments. We empirically evaluate our theoretical findings on large VLMs and demonstrate their effectiveness in capturing environment-level cross-modal contributions. Our analysis provides deeper insights into continual VLMs, highlighting their contribution robustness to varying task orders and inter-task similarities, and their improved generalization performance.

Current representation learning paradigms force a fundamental compromise: self-supervised methods scale to massive datasets but yield opaque features, whereas interpretable models remain bottlenecked by the need for dense human annotation. We introduce Self-Supervised Concept discOvery via Preference lEarning (\model), a label-free framework that resolves this dilemma. Instead of treating Vision-Large-Language Models (VLLMs) as static feature extractors, \model leverages them as active participants in a self-supervised preference optimization loop. By autonomously hypothesizing, validating, and reinforcing candidate visual attributes directly from raw imagery, our framework discovers novel, structured concepts without a single label. Extensive experiments across natural, medical, and physics domains demonstrate that \model successfully extracts domain-specific concepts where standard VLLMs often fail to generate. By amortizing concept discovery directly into the VLLM backbone through our self-supervised preference objective -- rather than relying on static generation and disjoint filtering -- we achieve up to a 24-point absolute improvement in downstream top-1 classification accuracy on unseen data. Our work suggest that interpretability can emerge through a model's autonomous interaction with incidental visual structures, without any human supervision.

Autonomous UAV navigation is conventionally solved by pipelines that separate perception, mapping, and planning into distinct stages, which propagates errors, accumulates latency, and requires environment-specific retuning. End-to-end generative models remove these interfaces by mapping raw observations directly to trajectories, but inherit a subtle failure mode: trained on clean data, they cannot recognise when an observation is unreliable, and treat degraded regions such as glass, mirrors, and overexposed surfaces as valid evidence for planning. We present a reliability-aware diffusion planner for 3D UAV navigation. It conditions trajectory generation on the observation together with a scene-level reliability heatmap that marks where perception cannot be trusted, produced by a lightweight network that distils the open-vocabulary reasoning of a vision-language model within the real-time planning budget. To generalise to unseen environments without retraining, we steer the denoising process with a differentiable two-stage ESDF cost that treats physical obstacles from depth and virtual obstacles from highly unreliable regions on equal footing. In simulation and on a real quadrotor, our planner produces markedly safer trajectories than a state-of-the-art diffusion baseline, reducing the obstacle-violation rate from 40.3% to 9.6% and raising the mean reliability of traversed regions from 0.588 to 0.925. Ablating the reliability term alone drops mean reliability from 0.898 to 0.783, confirming it as the decisive component, while distillation runs the framework up to 2 times faster than the full vision-language model.

End-to-end navigation policies trained on large simulation corpora degrade sharply when transferred to out-of-distribution scenes, categories, or goal modalities. Modular pipelines such as Modular GOAT are bottlenecked by closed-set object detection recall, while 3D snapshot-memory systems (e.g. 3D-Mem) accumulate dense, view-dependent representations that are heavy to maintain. We present AnyGoal, a training-free multi-robot architecture that places a Vision-Language Model (VLM) at the core of frontier-based exploration and coordinates agents through a shared 2D Gaussian Bayesian Value Map (BVM). The BVM maintains a per-pixel (mu, sigma^2) posterior over goal relevance, updated via precision-weighted fusion of VLM scores through a depth-cone mask, and is never reset between subtasks, yielding lifelong evidence accumulation. Frontiers are ranked by a convex blend of a VLM-as-judge softmax and a Bayesian UCB term on the BVM. A greedy allocator with spatial-separation penalty and commitment hysteresis distributes frontiers across agents without a centralized controller. On the full GOAT-Bench val unseen split (360 episodes, 2,669 subtasks), our dual-agent system achieves 52.4% Subtask SR at 12.7% SPL--state of the art under the strict physical regime (discrete 0.25 m steps, no teleportation, 42 deg HFOV) and a +27.5 pp improvement over Modular GOAT (24.9%). Single-agent AnyGoal achieves 41.9% Subtask SR, showing gains arise from the decision architecture. A four-way perception ablation shows that open-vocabulary detectors shift the dominant failure mode from exploration to goal verification.

On-policy distillation (\textsc{OPD}) has recently become a prominent post-training recipe by combining two desirable ingredients: on-policy student trajectories and dense teacher supervision. However, how this hybrid changes a model's parameters remains unclear. Across several language and vision-language model pairs and \textsc{OPD} use cases, our analysis yields two main findings. On sparsity, \textsc{OPD} updates are small and coordinate-sparse. They are distributed across layers, with the largest relative movement usually appearing in FFN modules. This sparse structure is operationally useful: training only the discovered subnetwork nearly recovers full-training performance. The sparse support does not remove the need for adaptive optimization: SGD, previously reported to be competitive in \textsc{RLVR}, underperforms AdamW in our \textsc{OPD} optimizer ablation, suggesting that dense teacher supervision preserves useful momentum structure and heterogeneous second-moment scales. On geometry, the updates are numerically full-rank but spectrally concentrated; they lie mostly away from the principal singular subspaces of the source weights and fall disproportionately on coordinates where the source weights are close to zero. These findings suggest that dense teacher supervision does not turn \textsc{OPD} into ordinary dense parameter rewriting; instead, \textsc{OPD} retains important geometric signatures of on-policy post-training.

Vision-language models (VLMs) are increasingly used to detect whether AI-generated images contain visible artifacts, yet their ability to analyze such artifacts remains poorly understood. A correct image-level decision can still hide important failures: a model may correctly flag an artifact while relying on the wrong visual cue, selecting the wrong region, or describing a defect that the image does not support. To evaluate these behaviors directly, we introduce SalArt-VQA, a diagnostic benchmark for fine-grained SALient ARTifact understanding in AI-generated images. SalArt-VQA contains 950 images and 3,681 human-authored multiple-choice questions spanning artifact images, matched real reference images, and paired generated reference images. Four aligned question types evaluate presence detection, semantic localization, spatial grounding, and evidence-grounded defect identification, while the reference splits test calibration and abstention when the annotated defect is absent. Across 20 VLMs, SalArt-VQA reveals failures that image-level detection accuracy hides: the strongest model reaches 99.37% detection recall on artifact images but answers all four artifact-side questions correctly on only 53.26% of images. Comparing artifact images with artifact-free references reveals a sensitivity-calibration tradeoff: sensitive models often make unsupported artifact claims, while conservative models avoid false alarms largely by missing real artifacts. These results show that high artifact detection accuracy alone does not imply grounded artifact understanding. SalArt-VQA exposes these hidden failure modes and provides a fine-grained evaluation of whether VLM artifact claims are supported by local visual evidence.

Vision-Language-Action (VLA) models that couple pretrained Vision-Language Models (VLMs) with continuous action experts have achieved strong manipulation performance, yet generalization to out-of-distribution (OOD) language instructions remains poor. A known challenge is the structural imbalance in VLA data, where language is far less diverse than visual and action content, making policies prone to visual shortcuts. While discrete-action methods mitigate this through vision-language co-training, continuous action experts lack such protection: they start from random initialization and learn entirely from imbalanced data, producing noisy gradients that corrupt the VLM and fail to exploit its language capability. We address this from a Bayesian perspective, factorizing the policy into a language-agnostic Vision-Action (VA) prior and a language-conditioned VLA likelihood, and propose APT, a two-stage training method emphasizing Action expert PreTraining. In Stage 1, the action expert is pretrained as a VA prior on vision-action pairs from a frozen VLM, bypassing the language imbalance. In Stage 2, language tokens are injected through a gated fusion mechanism that integrates VLM features while preserving the learned visuomotor prior. APT applies to mainstream VLA architectures, including the $π$ and GR00T-style architectures. Comprehensive experiments validate that APT achieves consistent gains on unseen instructions and compositional tasks. Project Page: https://xukechun.github.io/papers/APT/

Vision-language-action (VLA) models inherit a shared synchronous clock from vision-language pretraining, processing every input at one rate. This is misaligned with physical interaction, where a high-frequency modality changes at hundreds of hertz, vision evolves more slowly, and language stays constant across an episode. A synchronous VLA oversamples slow modalities, undersamples fast ones, and caps action generation at the lowest effective frequency. We hypothesize that decoupling temporal processing per modality, letting each update and retain information at its own sensor rate, yields stronger representations and more robust control. We present DAM-VLA, which maintains per-modality latent buffers refreshed at sensor rates and read continuously by the action head, integrating new high-frequency modalities through gated cross-attention that leaves the pretrained backbone intact. Across seven contact-rich real-world manipulation tasks, DAM-VLA more than doubles the average success rate of the strongest synchronous baseline (95.2% vs.\ 40.95%) while sustaining smooth, reactive 100,Hz control. Project website: \href{https://intuitive-robots.github.io/DAM-VLA/}{intuitive-robots.github.io/DAM-VLA/}

Sign language translation (SLT) converts sign language video into spoken language text and holds significant promise for improving accessibility and enabling communication between signing and non-signing communities. While large weakly-aligned datasets have enabled pre-training at scale and gloss-free methods have reduced reliance on expert annotation, high-quality parallel sign video-text pairs for fine-tuning remain scarce, limiting generalisation on long-tail vocabulary and unseen constructions. We propose a corpus augmentation approach that requires no additional human annotation, external sign-language video corpora, or generative video models, relying only on the existing gloss-annotated training corpus and an LLM for sentence generation: per-gloss clips are extracted from training videos via CTC forced-alignment, novel gloss-sentence pairs are generated by a corpus-anchored LLM, and synthetic sequences are assembled through random sentence sampling and clip assignment. The resulting synthetic RGB video-text pairs are architecture-agnostic at the downstream training stage and can be consumed directly by RGB-based SLT models, or converted into pose or feature representations by pipelines that derive such inputs from video. Sincan et al. re-evaluated five recent gloss-free methods under strictly identical conditions; the largest verified gain over the GFSLT-VLP baseline was only 0.98 BLEU-4. Our augmentation, applied within the same framework, achieves +2.92 BLEU-4 without any change to architecture or training protocol. We further identify that synthetic data harms vision-language pretraining despite improving its objectives, and that optimising clip transitions for visual smoothness is counter-productive under L2-based criteria; we propose that abrupt boundaries may act as a form of implicit regularisation. Code is available at https://github.com/robizso/slt-datagen.

Understanding multi-label images remains a challenging task in computer vision. With the rapid progress of vision-language multimodal learning, vision-language models (VLMs) enable zero-shot recognition without labeled data. However, due to their intrinsic design, these models often prioritize the most iconic object and omit other contextual positives. This intrinsic bias conflicts with the nature of multi-label learning, thereby limiting their applicability. In this work, we propose an unsupervised framework that adapts VLMs from iconic recognition toward inclusive understanding, enabling label-free multi-label image recognition. Our approach consists of two key stages, cutting'' and sewing'': In the cutting stage, we present the multi-sampling response estimator to prevent the model from concentrating only on one single object. In the second sewing stage, the multi-object blend adaptation is introduced to adjust the labels to better conform to the multi-label distribution while preserving the intrinsic characteristics of the original model within only one epoch. Extensive experiments show that our framework significantly outperforms existing unsupervised approaches on four public datasets, even surpassing several representative weakly supervised baselines. These results demonstrate the potential of adapting pre-trained VLMs for more comprehensive visual understanding without manual annotations. Our code is publicly available at https://github.com/iCVTEAM/TailorCLIP.

Vision-language models like CLIP can provide rich semantic priors for open-vocabulary object detection. However, jointly integrating both textual and visual knowledge into detection architectures remains challenging. In this paper, we propose VL-DINO, an open-vocabulary detector that enhances DINO through more effective exploitation of CLIP's vision-language knowledge. Specifically, a Query-guided Positive Sample Construction (QPSC) module is first developed to construct additional high-quality positive samples, enabling the vanilla DINO framework to better accommodate mixed training across heterogeneous data sources while providing more vision-language alignment signals, thereby incorporating richer textual knowledge during training. A Visual Semantic Encoder (VSE) module is then introduced to distill CLIP visual knowledge into backbone-extracted features, producing fused features for subsequent encoder refinement. Based on the fused features, an Object-Region Semantic Alignment (ORSA) module extracts object-centric region features and aligns them with the corresponding textual embeddings, further incorporating textual cues. In the zero-shot setting, VL-DINO-T and VL-DINO-L achieve 36.3 and 38.1 AP on the LVIS benchmark, respectively, consistently outperforming prior advanced approaches. Extensive experiments demonstrate the effectiveness and competitive performance of the proposed design.

Vision-Language Models (VLMs) demonstrate strong performance on general multimodal reasoning benchmarks, yet their ability to perform engineering reasoning remains largely unexplored. Unlike general visual question answering, engineering problem solving requires interpreting technical diagrams, selecting governing physical principles, and maintaining physically consistent multi-step reasoning. These capabilities are increasingly important for AI systems used in engineering education, scientific assistance, and technical decision-making, where reasoning failures may produce physically invalid yet superficially plausible solutions. Existing benchmarks primarily evaluate final answers and provide limited assessment of intermediate reasoning processes. We introduce EngVQA, a multimodal benchmark for evaluating engineering reasoning across 5 engineering subjects containing 696 problems. We introduce an 8-stage automatic evaluation framework for assessing VLM-generated solutions. The framework independently evaluates each stage of the solution, enabling fine-grained analysis of reasoning failures. We benchmark multiple state-of-the-art open and closed source VLMs on our evaluation framework and demonstrate substantial limitations in current engineering reasoning capabilities. Human evaluation shows strong agreement with our automated framework, achieving a Pearson correlation of 0.975 and a mean absolute error of 0.67 on a 10-point grading scale. Our results highlight the importance of process-oriented evaluation for reliable assessment of multimodal engineering reasoning systems.

3D vision-language segmentation aims to segment target objects in 3D scenarios according to the linguistic instructions and visual observations. Prior art heavily relies on the coarse superpoint representation to reduce the computation complexity, which suffers from poor segmentation quality and messy object boundaries. In this paper, we propose the SEGment-And-select (SEGA3D) paradigm for 3D visionlanguage segmentation that directly operates on the fine-grained visual information and is free from the superpoint dependency. Specifically, we first leverage a mask candidate generator to provide fine-grained categorical mask candidates, substantially improving the quality of candidate masks over the superpoint counterparts. Then, a Large Language Model (LLM) is utilized to generate the semantic and spatial information based on the linguistic description and visual features. The LLM output and visual features are fed to the Semantic-Spatial Selector (SSS) to produce the top-ranking mask candidates. Eventually, the Loopback Verification Module (LVM) is designed to yield the segmentation mask from the selected candidate masks. Our SEGA3D attains competitive performance on ScanRefer, ScanNet and Matterport3D benchmarks. Notably, our SEGA3D surpasses the top-performing counterpart by 8.3 mIoU and 5.3 mIoU on ScanNet and Matterport3D, respectively. Codes will be available upon publication.

Adversarial examples reveal vulnerabilities in Vision-Language Pre-training (VLP) models and provide insights for improving robustness. A key property is cross-model transferability, which enables transfer-based black-box attacks. However, existing attacks often rely heavily on the surrogate model, causing cross-model performance drops. One reason is that adversarial optimization may follow surrogate model responses more than input semantics, making the update direction effective on the surrogate but less transferable to unseen targets. We refer to this dependency as surrogate-specific bias. Motivated by this observation, DeBias-Attack improves transferability by correcting surrogate-specific bias in adversarial optimization directions. It maintains two perturbation branches. The main branch optimizes a perturbation on the original image and obtains the adversarial gradient used to disrupt image-text alignment. The reference branch optimizes a perturbation on a weak-semantic image constructed from the dataset mean image with small Gaussian noise resampled at each iteration. Since this weak-semantic image contains little clear visual content, its optimization reflects surrogate responses more than image semantics, and its reference gradient estimates surrogate-specific bias. DeBias-Attack removes the aligned projection of the main gradient on the reference gradient before updating the adversarial image, then performs context-aware text substitution using the updated adversarial image. DeBias-Attack is the first transfer-based VLP attack that corrects surrogate-specific bias through gradient correction. Experiments show strong performance across VLP models, downstream tasks, and open-source and closed-source multimodal large language models.

The cost signal that constrained-RL algorithms optimize against is almost always reactive: the simulator emits a non-zero cost only after a collision has begun, and the Lagrange multiplier of PPO-Lagrangian grows only after the episode budget has been exceeded. At race speeds, where collisions are instantaneous and irreversible, any safety mechanism that waits for cost to accumulate is structurally too late. We present VLM-Safe-RL, a framework that integrates a frozen vision-language model into the CMDP Lagrangian update as an anticipatory cost term. The framework comprises four contributions: (i) Decoupled Dual-Path CLIP, independent reward/cost paths that respect the CMDP's factorization; (ii) VLM-Lagrange, an augmented multiplier update that incorporates a per-step VLM cost as an anticipatory term; (iii) Confidence Gating, a Bayes-optimal weight derived from a logistic noise model on the CLIP margin; and (iv) VLMPPOLag, the composed algorithm. On Safety-Gymnasium FormulaOne L2, our principal evaluation ($n{=}5$ seeds, $10^{6}$ steps, budget $d_{\text{lim}}{=}25$) VLMPPOLag$+$Conf is the only configuration in our default budget comparison that simultaneously retains substantive return ($J_r{\approx}40$) and holds cost within budget on a majority of seeds; the five constraint-aware baselines (PPOLag, CPO, CPPOPID, CPO-CLG, PPOLag-RND) each fail at least one requirement. The mechanism generalizes to held-out MetaDrive Medium (catastrophe rate $41\%{\to}26\%$, 95% bootstrap CI $[-26,-5]$,pp) and shows directionally consistent transfer to Bullet Safety-Gym; we report honestly where it does not (MetaDrive Easy/Hard, Qwen2-VL backbone) and trace the Hard failure to a Lagrangian-regulation pathology rather than the VLM signal itself. To our knowledge, this is the first work to use frozen VLM signals as an anticipatory cost term inside the CMDP Lagrangian update.

Two-stage post-training -- a Stage-1 warm-start (supervised fine-tuning, SFT, or on-policy distillation, OPD) followed by Stage-2 reinforcement learning (RL) -- is increasingly used for vision-language models (VLMs). We ask what Stage-1 actually controls in a small-data study using Qwen2.5-VL-7B with a same-modality 72B VLM teacher for OPD. First, the three warm-starts reach a narrow $53$--$54\%$ band on Geometry3K internal validation, consistent with the narrow range reported by recent specialized methods; this setup provides little evidence that Stage-1 changes the in-domain endpoint. Second, a matched-recipe, early-stopped SFT improves out-of-domain MathVista by $+2.1$ points, reversing the $-9.5$-point drop of an over-trained variant. The clearest difference is the \emph{entropy regime}: OPD enters RL with substantially higher policy entropy than either SFT initialization, and the separation remains visible through the available trajectories. At the in-domain initialization, OPD also has higher answer diversity and pass@16 ($+2.0$ to $+5.2$ points over SFT), although problem-level bootstrap intervals show that the smaller contrast is uncertain. The advantage is absent after RL (endpoint pass@16 values within $1.1$ points) and on MathVista (six models within $1.2$ points). Our contribution is therefore a bounded empirical characterization: Stage-1 is strongly associated with the entropy regime in this setup, but the downstream payoff is small, localized, and not evidence that OPD is a better RL warm-start.

Vision-language models (VLMs) with varying performance and resource requirements are widely deployed, making it difficult for users to select the most appropriate one among numerous VLM candidates. Existing work reveals the performance paradox phenomenon in language models and focuses on routing methods to solve it. However, developing a router for VLM selection is still a critical yet challenging problem, which primarily faces: 1) lack of specialized data, 2) ineffective feature representation, and 3) rigid model space and costly adaptation. In this paper, we construct a multimodal dataset for VLM selection, containing the outputs of seven mainstream VLMs on 32,626 unique image-text queries. We then propose ARMS, a router for VLM selection. ARMS enhances input signals with VLM profiles, employs a simple but effective architecture to improve representations of queries and VLM capabilities. To improve ARMS' adaptation to new VLMs, we propose two extension training strategies: incremental training and independent training. Experimental results on both in-distribution and out-of-distribution test sets demonstrate the effectiveness of ARMS. In particular, using our training strategy, ARMs (only 800M in size) can adapt to a broader VLM space and defeat commercial models like GPT-4o that are hundreds of times larger in scale. Our code, models, and datasets are available in the anonymous repository.

Semiconductor lithography inspection requires reliable detection of small pattern defects such as bridge, burr, pinch, and contamination. In this study, we propose a two-stage vision-language framework that combines initial defect detection with prediction refinement. In the first stage, Qwen3-VL is fine-tuned with LoRA as a vision-language adapter to predict defect counts, defect categories, and normalized bounding boxes from lithography images. However, direct fine-tuning may still produce common test-time errors, including false positives, missed defects, and incorrect defect types. To address this limitation, the second stage trains a refinement module using first-stage prediction failures and their corrected labels, allowing the model to review and revise initial outputs. By learning from cases where the initial adapter fails, the refinement process improves defect inference beyond single-stage fine-tuning.

Reasoning Vision-Language Models (VLMs) achieve strong performance on complex multimodal tasks, but reliable real-world application requires handling visual inputs that are messier than clean, curated benchmarks. Existing works mainly evaluate such reliability of VLMs through input corruptions, such as noise, blur and weather effects, which make visual evidence harder to perceive. This leaves a critical reliability failure mode underexplored: a model may perceive the evidence correctly, yet reason from plausible but irrelevant and distracting evidence and propagate this mistake to its final answer. To address this gap, we introduce \textbf{Distract-Bench}, a benchmark for evaluating VLM robustness to \textbf{semantic visual distractions}, defined as meaningful but task-irrelevant visual cues added to inputs while preserving the ground-truth answer. We comprehensively evaluate eight leading open-source and two closed-source VLMs across conventional vision corruptions and Distract-Bench. Our results show that Distract-Bench exposes a robustness failure distinct from vision corruptions: reasoning VLMs largely track their non-reasoning base models under perceptual degradation, but show consistently lower robustness to semantic distractions. Further analysis shows that these distractions often enter the reasoning process of VLMs, are treated as evidence, and lead to incorrect answers. Together, these findings reframe robustness evaluation for reasoning VLMs, shifting the focus from degraded perception to distractions for reliable real-world visual reasoning. Our data and code are available at https://github.com/Yizheng-Sun/Distract-Bench.