Vision-Language

In Vision-Language Models (VLMs), processing a massive number of visual tokens incurs prohibitive computational overhead. While recent training-aware pruning methods attempt to selectively discard redundant tokens, they largely rely on continuous-gradient relaxations. However, visual token pruning is inherently a discrete, non-convex combinatorial problem; consequently, these continuous approximations frequently trap the optimization in sub-optimal local minima, especially under aggressive compression budgets. To overcome this fundamental bottleneck, we propose GRIP-VLM, a Group-Relative Importance Pruning framework driven by Reinforcement Learning. Rather than relying on smooth-gradient assumptions, GRIP-VLM formulates pruning as a Markov Decision Process, employing a Group Relative Policy Optimization (GRPO) paradigm anchored by supervised warm-up to directly explore the discrete selection space. Integrated with a budget-aware scorer, our lightweight agent dynamically evaluates per-token importance and adapts to arbitrary compression ratios without retraining. Extensive experiments across diverse multimodal benchmarks demonstrate that GRIP-VLM consistently outperforms heuristic and supervised-learning baselines, achieving a superior Pareto frontier and delivering up to a 15% inference speedup at equal accuracy.

Standard dual-encoder vision-language models that map images and text to deterministic points on a shared unit hypersphere through $\ell_2$ normalization typically expose neither \emph{aleatoric} uncertainty (cross-modal ambiguity) nor \emph{epistemic} uncertainty (lack of training-distribution support). Existing post-hoc methods either recover at most one of the two uncertainty components, or ignore the hyperspherical geometry of these models' embeddings. We propose \textbf{GeoFlowVLM} as a post-hoc adapter that learns the joint distribution of paired $\ell_2$-normalised dual-encoder VLM embeddings on the product hypersphere $\mathbb{S}^{d-1} \times \mathbb{S}^{d-1}$ via Riemannian flow matching with a single masked velocity field. A consistency result shows that, in the population limit, the trained network exposes the joint flow and both cross-modal conditional flows as valid Riemannian flow-matching velocity fields on their respective domains. We derive two quantities from this single model: a conditional retrieval entropy that quantifies aleatoric ambiguity with a decision-theoretic interpretation via a Fano-type bound, and a marginal-typicality epistemic score justified by an exact chain-rule decomposition of the joint NLL. This decomposition isolates a cross-modal pointwise-mutual-information term that is structurally discriminative rather than epistemic, and is empirically the only consistently uninformative standalone component. Empirically, the entropy tracks Recall@1 with near-ideal monotonic calibration across three retrieval benchmarks in both directions, and the marginal-typicality sum yields consistently calibrated selective accuracy across four zero-shot classification benchmarks.

Vision-language models typically reason over post-ISP RGB images, although RGB rendering can clip, suppress, or quantize sensor evidence before inference. We study whether grounding improves when the visual interface is moved closer to the underlying camera measurement. We formulate measurement-grounded vision-language learning and instantiate it as PRISM-VL, which combines RAW-derived Meas.-XYZ inputs, camera-conditioned grounding, and Exposure-Bracketed Supervision Aggregation for transferring supervision from RGB proxies to measurement-domain observations. Using a quality-controlled 150K instruction-tuning set and a held-out benchmark targeting low-light, HDR, visibility-sensitive, and hallucination-sensitive cases, PRISM-VL-8B reaches 0.6120 BLEU, 0.4571 ROUGE-L, and 82.66% LLM-Judge accuracy, improving over the RGB Qwen3-VL-8B baseline by +0.1074 BLEU, +0.1071 ROUGE-L, and +4.46 percentage points. These results suggest that part of VLM grounding error arises from information lost during RGB rendering, and that preserving measurement-domain evidence can improve multimodal reasoning.

Testing robots requires assessing whether they perform their intended tasks correctly, dependably, and with high quality, a challenge known as the test oracle problem in software testing. Traditionally, this assessment relies on task-specific symbolic oracles for task correctness and on human manual evaluation of robot behavior, which is time-consuming, subjective, and error-prone. To address this, we propose VISOR, a Vision-Language Model (VLM)-based approach for automated test oracle assessment that eliminates the need of expensive human evaluations. VISOR performs automated evaluation of task correctness and quality, addressing the limitations of existing symbolic test oracles, which are task-specific and provide pass/fail judgments without explicitly quantifying task quality. Given the inherent uncertainty in VLMs, VISOR also explicitly quantifies its own uncertainty during test assessments. We evaluated VISOR using two VLMs, i.e., GPT and Gemini, across four robotic tasks on over 1,000 videos. Results show that Gemini achieves higher recall while GPT achieves higher precision. However, both models show low correlation between uncertainty and correctness, which prevents using uncertainty as a correctness predictor.

Long-horizon reasoning requires deciding not only what actions to take, but how deeply to commit before the next observation. We formalize this as \emph{commitment depth}: the number of primitive actions executed open-loop between replans. Commitment depth induces a trade-off between replanning cost and compounding execution error, yet most existing long-horizon systems fix it as a hand-designed scalar. In this work, we instead treat commitment depth as a learnable, state-conditioned variable of the policy itself. We instantiate this within a model-native vision--language policy that jointly predicts both what to execute and for how long. Across Sliding Puzzle and Sokoban, the resulting adaptive policy Pareto-dominates every non-degenerate fixed-depth baseline, achieving up to 12.5 percentage points higher solve rate while using approximately 25% fewer primitive actions per episode. Despite using a 7B backbone, our method outperforms GPT-5.5 and Claude Sonnet on both tasks, while every tested open-weight vision--language model achieves 0% zero-shot success. We further present a theoretical analysis showing that, under the standard commitment-depth surrogate, state-conditioned commitment strictly dominates any fixed depth whenever the locally optimal depth varies across states.

Adaptive prompting mechanisms have been proposed to enhance vision-language models by dynamically tailoring prompts to inputs. However, in frozen few-shot prompt learning with CLIP-style backbones, we systematically observe that adaptive gates and prompt-selection modules often collapse: they produce nearly constant outputs, contribute negligible gradient signals, and frequently fail to outperform fixed prompts. To further explore this issue, we present a systematic diagnostic study to uncover the underlying causes and conditions of adaptation failure. Through controlled experiments across datasets and multiple prompt learning architectures, we identify two recurring failure modes: gradient magnitude imbalance and gate degradation. Our findings invite a re-examination of indiscriminately adding architectural complexity in parameter-efficient learning and clarify when prompt-level adaptive gating is, and is not, effective in this regime.

Are low-attention visual tokens truly redundant in vision-language reasoning? Existing pruning methods often assume so, ranking visual tokens by shallow text-to-image attention and discarding low-scoring patches to accelerate LVLM inference. We show that this scalar criterion is unreliable for compositional reasoning: tokens ignored in early layers can later become essential for resolving secondary objects, spatial relations, and contextual cues. Premature pruning can therefore induce Visual Aphasia, a failure mode in which the model loses visual grounding and falls back on language priors. We introduce COAST (COntrastive Adaptive Semantic Token Pruning), a training-free pruning framework that casts compression as adaptive semantic routing. COAST uses native cross-modal attention to identify query-specific anchors and estimate contextual dispersion via attention entropy, then adapts the retention trade-off between semantic evidence and spatial context. It further uses a contrastive routing score to preserve both anchor-aligned evidence and complementary spatial context. Across seven benchmarks, COAST reduces visual tokens by 77.8% and achieves a 2.15x latency speedup while retaining 98.64% of the original average performance. Beyond a single backbone or compression setting, COAST consistently outperforms strong pruning baselines across token budgets and generalizes across multiple LVLM families, showing that adaptive semantic routing is a robust alternative to one-shot scalar pruning

Automated grading of diabetic retinopathy (DR) faces several critical challenges: subtle inter-grade visual distinctions in fine-grained lesion patterns, distributional discrepancies induced by heterogeneous imaging devices and acquisition conditions, and the inherent inability of purely visual approaches to exploit clinical semantic knowledge. In this paper, we propose CLIP-Guided Semantic Diffusion (CGSD), a DR grading framework that synergistically integrates vision-language pretraining with diffusion probabilistic modeling. We adopt a domain-specific vision-language model tailored for DR grading as the semantic guidance module and adapt it to the target domain via Low-Rank Adaptation (LoRA), effectively bridging the distributional gap between the pretrained model and the target dataset with only a minimal number of trainable parameters. Building on this foundation, we construct a cross-modal semantic conditioning vector by computing the dot product between image features and the text description features of each DR grade, yielding a joint representation that simultaneously encodes visual content and clinical-grade semantics. This vector serves as the conditioning signal for the diffusion denoising network, replacing the structurally complex dual-branch visual prior employed in existing diffusion-based classification methods. Experiments on the APTOS 2019 dataset demonstrate that the proposed approach achieves an accuracy of 87.5% and a macro-averaged F1 score of 0.731, outperforming a variety of representative methods. Ablation studies further validate the independent contribution of each constituent module.

In recent years, pre-trained visual-linguistic models have demonstrated tremendous potential, becoming a crucial foundational framework for numerous downstream tasks. However, the information density between text and images is not uniformly distributed. Existing methods often overlook the inherent and dynamic differences in information density and semantic scope between text tags and image blocks. These common uniform alignment strategies result in coarse-grained cross-modal interactions and loss of fine semantic details. Moreover, pursuing finer alignment typically requires substantial computational overhead, limiting practical model deployment. To address this challenge, this paper proposes a novel framework for dynamic cross-modal alignment with continuous detail introduction. First, we design a dynamically adaptive cross-modal matching mechanism that uses a learnable matching function to dynamically assign varying numbers and sizes of image tags to text tags of the same size but different information density, enabling more precise attention interaction. Second, we develop a continuous detail introduction module to progressively incorporate high-resolution visual feature enhancement into the alignment process. Extensive experiments across multiple benchmarks demonstrate significant improvements in the accuracy of various downstream tasks while reducing computational overhead.

Vision-Language Models (VLMs) exhibit systematic bias toward visual illusions, recalling memorized facts rather than perceiving actual visual differences. This paper presents a training-free framework for the 5th DataCV Challenge Task 1 at CVPR 2026, addressing this perception-versus-memory conflict through three complementary strategies:(1) illusion-aware image preprocessing that weakens illusion-inducing context via type-specific transformations (edge extraction, color isolation, morphological processing, and reference-line overlay), (2) anti-illusion prompt engineering guiding VLMs toward qualitative visual comparison, and (3) multi-vote ensemble that further improves robustness. Our method achieves 90.48% accuracy on the official 630-image test set using Claude (claude-opus-4-6) with 5-vote majority ensemble, and 98.41% on a human-verified subset. The approach requires no finetuning, relying solely on visual manipulation and prompt design. Our solution secured 2nd place in the challenge, only 0.47% behind the 1st-place solution. Code is available at https://github.com/jasminezz/sf-illusion-aware-vlm.git.

End-to-end autonomous driving requires models to understand traffic scenes, infer driving intent, and generate executable motion plans. Recent vision-language-action (VLA) models inherit semantic priors from large-scale vision-language pretraining, yet still face a coupling trade-off: fully shared backbones preserve multimodal interaction but may entangle language reasoning and trajectory prediction, whereas decou pled reasoning-action pipelines reduce task conflict but weaken semantic-motion coupling. We propose VECTOR-DRIVE, a tightly coupled VLA framework built on Qwen2.5-VL-3B. VECTOR-DRIVE keeps all tokens coupled through shared self attention and routes feed-forward computation according to token semantics. Vision and language tokens are processed by a Vision-Language Expert to preserve semantic priors, while target-point, ego-state, and noisy action tokens are routed to a Trajectory Expert for motion-specific computation. On the action-token pathway, a flow-matching planner refines noisy action tokens into future waypoints and speed profiles. This design couples semantic reasoning and motion planning within a single multimodal Transformer while separating task-specific FFN computation. On Bench2Drive, VECTOR-DRIVE achieves 88.91 Driving Score and outperforms representative end-to end and VLA-based baselines. Qualitative results and ablations further validate the benefits of shared attention, semantic-aware expert routing, progressive training, and flow-based action de coding.

Recent "Thinking with Video" approaches use Video Generation Models (VGMs) for visual reasoning by producing temporally coherent Chain-of-Frames as reasoning artifacts. Even strong VGMs, however, exhibit two recurring failure modes on goal-directed tasks: long-horizon drift on multi-step tasks and mid-clip simulation errors that compound. Both stem from the absence of explicit reasoning built upon the VGM's short-horizon visual prior, a role naturally filled by Vision-Language Models (VLMs), but where to place the VLM is non-trivial: upfront plans commit before any frame is generated and post-hoc critiques over whole videos intervene too late. We propose VLM-VGM Collaborative Video Reasoning (CollabVR), a closed-loop framework that couples the VLM with the VGM at step-level granularity: the VLM plans the immediate next action, inspects the clip the VGM generates, and folds the verifier's diagnosis directly into the next action prompt to repair detected failures. On Gen-ViRe and VBVR-Bench, CollabVR improves both open-source and closed-source VGMs over single-inference, Pass@$k$, and prior test-time scaling baselines at matched compute, with the largest gains on the hardest tasks. It also yields further improvements on top of a reasoning-fine-tuned VGM, indicating that step-level VLM supervision is orthogonal to and stackable with reasoning-oriented fine-tuning. We provide video samples and additional qualitative results at our project page: https://joow0n-kim.github.io/collabvr-project-page.

Platform content moderation applies explicit policy rules and context-dependent conditions to decide whether user content is allowed, restricted, or removed. A correct moderation outcome must therefore depend on which rules a case activates, how those rules interact, and whether the available evidence is sufficient. Current multimodal safety benchmarks largely reduce moderation to matching predefined final labels, leaving this underlying rule structure untested. As a result, a high benchmark score reveals little about whether a model applies the policy correctly or arrives at the correct label through superficial cues. To evaluate this rule-governed process, we introduce RuleSafe-VL, a benchmark for rule-conditioned decision reasoning in vision-language content moderation. Derived from publicly available platform moderation policies, RuleSafe-VL formalizes 93 atomic rules and 92 typed rule relations, yielding 2,166 context-sensitive image-text cases across three high-risk policy families. Its four diagnostic tasks decompose moderation into a rule-conditioned decision chain. They identify activated rules, recover rule interactions, judge decision sufficiency, and resolve outcomes once missing context is supplied. Experiments on 10 frontier, open-source, and safety-oriented VLMs reveal rule-relation recovery as the dominant bottleneck, where the best model reaches only 64.8 Macro-F1 and some safety-oriented models fall below 7 Macro-F1. Decision-state prediction also remains unreliable, peaking at 64.5 Macro-F1. RuleSafe-VL shifts moderation evaluation from final-label scoring toward diagnostic assessment of rule-conditioned decision reasoning.

Multimodal RAG systems increasingly rely on vision-language retrievers to ground visual queries in external textual evidence. Existing adversarial studies on RAG mainly manipulate the retrieval corpus or memory, while attacks on vision-language and remote sensing models typically target end-task predictions. Input-space threats to the evidence retrieval stage of remote sensing multimodal RAG remain underexplored. To address this gap, we introduce CloudWeb, an atmospheric retrieval hijacking attack that modifies only the input image while keeping the retriever, generator, and knowledge base fixed at deployment. CloudWeb overlays parameterized cloud- and haze-like patterns on remote sensing images and optimizes them with a retrieval-oriented objective that pulls adversarial image embeddings toward target atmospheric evidence, suppresses source-scene evidence, enforces rank separation, and regularizes naturalness and coverage. To the best of our knowledge, this is the first study of retrieval-stage atmospheric evidence hijacking in remote sensing multimodal RAG. We evaluate CloudWeb on a seven-dataset remote sensing RAG benchmark with five CLIP-style retrievers, including GeoRSCLIP, RemoteCLIP, OpenAI CLIP, and OpenCLIP, together with downstream vision-language generators. Across retrievers, CloudWeb consistently outperforms clean retrieval, handcrafted atmospheric baselines, random cloud perturbations, and fixed variants in injecting weather-related evidence into top-ranked results. On GeoRSCLIP ViT-B/32, Weather@5 increases from 0.71% to 43.29%. Downstream generation further shows measurable weather hallucination and semantic shift, indicating that retrieval-stage hijacking can propagate to the final RAG response. These findings reveal a practical failure mode: natural-looking atmospheric changes can compromise evidence retrieval before generation begins.

Vision-language models (VLMs) have shown strong potential for scientific image understanding, but general-purpose models often lack the domain-specific visual knowledge required for reliable materials characterization. In this work, we fine-tuned an open-source VLM (Qwen3-VL-32B-Instruct) for fracture-surface image analysis using a curated dataset of 13,168 open-source, literature-mined fracture-surface images. Morphology annotations were generated by GPT-5.2-Reasoning (high) from both the images and relevant excerpts of their source papers, and the dataset was further enriched with targeted manual collection and rotation-based augmentation. The resulting specialist model outperforms flagship proprietary multimodal models on a benchmark of 100 manually annotated images. It achieves a precision of 0.92, compared to 0.35 for the base Qwen3-VL-32B-Instruct, 0.58 for GPT-5.5-Reasoning (high), and 0.78 for Gemini 3.1 Pro-Reasoning (high). Dataset ablations show that manual collection of rare-feature images and augmentation via image rotation are both beneficial to improve recognition of less common fracture morphology features. We further discuss integrated use of the fine-tuned model with proprietary models to combine fracture-specific visual accuracy with broader multimodal reasoning for autonomous fractography. Although focused on fracture-surface images, this work demonstrates how VLMs can be adapted through targeted collection and fine-tuning on novel feature images to recognize those features and support downstream decision-making in autonomous microscopy workflows.

Open-world referring segmentation requires grounding unconstrained language expressions to precise pixel-level regions. Existing multimodal large language models (MLLMs) exhibit strong open-world visual grounding, but their outputs remain limited to sparse bounding-box coordinates and are insufficient for dense visual prediction. Recent MLLM-based segmentation methods either directly predict sparse contour coordinates, struggling to reconstruct continuous object boundaries, or rely on external segmentation foundation models such as the Segment Anything Model (SAM), introducing substantial architectural and deployment overhead. We present Qwen3-VL-Seg, a parameter-efficient framework that treats the MLLM-predicted box as a semantically grounded structural prior and decodes it into pixel-level referring segmentation. At its core, a lightweight box-guided mask decoder combines multi-scale spatial feature injection, spatial-semantic query construction, box-guided high-resolution pixel fusion, and iterative mask-aware query refinement, introducing only 17M parameters (about 0.4% of the base model). For scalable open-world training, we construct SA1B-ORS, an SA-1B-derived dataset with two subsets: SA1B-CoRS (category-oriented samples) and SA1B-DeRS (descriptive, instance-specific samples). For evaluation, we curate ORS-Bench, a manually screened benchmark with in-distribution and out-of-distribution subsets covering diverse referring expression types. Extensive experiments on referring expression segmentation, visual grounding, and ORS-Bench show that Qwen3-VL-Seg performs strongly across closed-set and open-world settings, with clear advantages on language-intensive instructions and strong out-of-distribution generalization. Evaluations on general multimodal benchmarks further show that the model broadly preserves general-purpose multimodal competence after segmentation-oriented adaptation.

Multimodal Large Language Models (MLLMs) have made remarkable progress on vision-language reasoning, yet most methods still compress visual evidence into discrete textual thoughts, creating an information bottleneck for fine-grained perception. Recent latent visual reasoning methods attempt to reason in continuous hidden states, but we find that they suffer from insufficient manifold compatibility: latent trajectories drift away from pretrained reasoning circuits, collapse into instance-agnostic patterns, and are often bypassed during answer generation. To address these issues, we propose RIS (Retrieve, Integrate, and Synthesize), a spatial-semantic grounded framework that develops latent reasoning as a compatible extension of pretrained MLLM computation. We first construct a step-wise grounded reasoning dataset with bounding boxes and region-specific semantic descriptions. Built on this supervision, RIS anchors latent tokens to both spatial and semantic evidence, enforces their causal role through a progressive attention bottleneck, and introduces short language transition tokens to bridge synthesized latent states back to vocabulary-aligned decoding. Experiments on V*, HRBench4K, HRBench8K, MMVP, and BLINK show consistent improvements over closed/open-source and latent reasoning baselines. Further analyses demonstrate that RIS learns diverse, interpretable, and progressively integrated latent trajectories, offering a practical path toward faithful internal visual reasoning in MLLMs.

How to achieve vision-language (VL) tracking using natural language descriptions from a video sequence \textbf{without relying on any bounding-box ground truth}? In this work, we achieve this goal by tackling \textit{self-supervised VL tracking}, which aims to evaluate tracking capabilities guided by natural language descriptions. We introduce \textbf{\tracker}, a novel self-supervised VL tracker that is capable of tracking any referred object by a language description. Unlike traditional methods that equally fuse all language and visual tokens, we propose an efficient Dynamic Token Aggregation Module, which treats each visual token \textbf{unequally}. The module consists of three main steps: i) Based on an anchor token, it selects multiple important target tokens from the template frame. ii) The selected target tokens are merged according to their attention scores and aggregated into the language tokens, thereby eliminating redundant visual token noise and enhancing semantic alignment. iii) Finally, the fused language tokens serve as guiding signals to extract potential target tokens from the search frame and propagate them to subsequent frames, enhancing temporal prompts and encouraging the tracker to autonomously learn instance tracking from unlabeled videos. This new modeling approach enables the effective self-supervised learning of language-guided tracking representations without the need for large-scale bounding box annotations. Extensive experiments on VL tracking benchmarks show that {\tracker} surpasses SOTA self-supervised methods.

Systematic reviews and meta-analyses frequently require numerical data that authors report only as figures, yet manual digitisation is slow and does not scale. We present PlotPick, an open-source tool that uses vision-language models (VLMs) to batch-extract structured tabular data from scientific figures. We evaluate six VLMs from three providers on two established chart-to-table benchmarks (ChartX and PlotQA) and compare against the dedicated chart-to-table model DePlot. All six VLMs outperform DePlot on both benchmarks. On ChartX (restricted to bar charts, line charts, box plots, and histograms; n=300), VLMs achieve 88-96% recall versus 71% for DePlot. On PlotQA (n=529), VLMs achieve 86-99% RMSF1 versus 94% for DePlot. The gap is largest on chart types absent from the dedicated models' training data: on box plots, DePlot achieves 24% RMSF1 while VLMs achieve 83-97%. PlotPick is available at https://plotpick.streamlit.app.

The condition monitoring (CM) of synthetic fibre ropes (SFRs) used in offshore, maritime, and industrial settings demands more than a classifier: inspectors need continuous severity estimates, maintenance recommendations, anomaly flags, deterioration timelines, and automated reports, all from a single inspection image. We present DART (Damage Assessment via Rope Transformer), a vision-language foundation model that addresses the full rope inspection workflow through a unified multi-task architecture. DART extends the Joint-Embedding Predictive Architecture (JEPA) to the cross-modal domain by coupling a Vision Transformer (ViT-H/14) with Llama-3.2-3B-Instruct via a Severity-Conditioned Cross-Modal Fusion (SC-CMF) module. Three architectural innovations drive the model's versatility: (1) HD-MASK, a saliency-guided masking strategy that focuses self-supervised reconstruction on damage-dense patches; (2) per-class learnable severity gates that adaptively weight language grounding by damage category; and (3) a Contrastive Damage Disentanglement (CDD) loss that shapes the embedding space to simultaneously encode damage type, severity ordering, and cross-modal semantics. Trained once on 4,270 images spanning 14 fine-grained rope damage classes, the frozen DART backbone supports downstream tasks without any task-specific fine-tuning: damage classification (93.22 % accuracy, 91.04 % macro-F1, +38.5 pp over a vision-only baseline), continuous severity regression (Spearman rho = 0.94, within-1-ordinal accuracy 99.6 %), few-shot recognition (89.2 % macro-F1 at 20 shots). These results demonstrate that DART functions as a general-purpose CM backbone that goes well beyond classification, providing actionable inspection intelligence from a single shared representation.

A pervasive intuition holds that vision-language models (VLMs) are most trustworthy when their attention maps look sharp: concentrated attention on the queried region should imply a confident, calibrated answer. We test this Attention-Confidence Assumption directly. We instrument three open-weight VLM families (LLaVA-1.5, PaliGemma, Qwen2-VL; 3-7B parameters) with a unified mechanistic pipeline -- the VLM Reliability Probe (VRP) -- that compares attention structure, generation dynamics, and hidden-state geometry against a single correctness label. Three results emerge. (i) Attention structure is a near-zero predictor of correctness (R_pb(C_k,y)=0.001, 95% CI [-0.034,0.036]; R_pb(H_s,y)=-0.012, [-0.047,0.024] on a pooled n=3,090 split), even though attention remains causally necessary for feature extraction (top-30% patch masking drops accuracy by 8.2-11.3 pp, p<0.001). (ii) Reliability becomes legible later in the computation: a single hidden-state linear probe reaches AUROC>0.95 on POPE for two of three families, and self-consistency at K=10 is the strongest behavioral predictor we measure at 10x inference cost (R_pb=0.43). (iii) Causal neuron-level ablations expose a sharp architectural split with direct monitor-design implications: late-fusion LLaVA concentrates reliability in a fragile late bottleneck (-8.3 pp object-identification accuracy after top-5 probe-neuron ablation), whereas early-fusion PaliGemma and Qwen2-VL distribute it widely and absorb destruction of ~50% of their peak-layer hidden dimension with <=1 pp degradation. The takeaway is narrow but consequential: in 3-7B VLMs, reliability is read more reliably off hidden-state geometry, layer-wise margin formation, and sparse late-layer circuits than off attention-map sharpness.

With advances in multimodal research and deep learning, Multimodal Large Language Models (MLLMs) have emerged as a powerful paradigm for a wide range of multimodal tasks. As a core problem in vision-language research, Visual Question Answering (VQA) has increasingly employed MLLMs to improve performance, particularly in open-domain settings where external knowledge is essential. In this work, we aim to further enhance retrieval-based VQA by more effectively integrating MLLMs with structured reasoning and knowledge acquisition. We introduce a logical prompting strategy that fuses Chain-of-Thought (CoT) reasoning with Visual Question Decomposition (VQD), termed CoVQD, to guide retrieval toward more accurate and relevant knowledge for MLLM inference. Building on this idea, we propose a new framework, CoVQD-guided RAG (CgRAG), which enables MLLMs to access more comprehensive and coherent external knowledge while benefiting from structured visual-text reasoning guidance, thereby improving generalization and reliability in complex cross-domain VQA scenarios. Extensive experiments on E-VQA, InfoSeek, and OKVQA benchmarks demonstrate the effectiveness of the proposed method.

Source-Free Domain Adaptation (SFDA) adapts source models to target domains without accessing source data, addressing privacy and transmission issues. However, existing methods still initialize from a source pre-trained model and thus are not truly source-free. Recent works have introduced Vision-Language (ViL) models to guide the adaptation process, in these methods, we observe that for the same target domain, different source models yield minimal variation in final results, indicating the source model itself has limited impact. Motivated by this, we propose ViL-Only Domain Adaptation (VODA) , a stricter setting that eliminates all dependencies on source domain, relying solely on a randomly initialized model, a ViL model, and unlabeled target data. We analyze the adaptation dynamics of VODA and introduce Two-Stage Denoised-Region Distillation (TS-DRD) , a two-stage framework that first warms up the model with ViL guidance, then seek a Denoised-Region inherent in both the ViL and adapting model, yielding cleaner supervision for distillation. Experiments on Office-Home, VisDA, and DomainNet-126 show that under VODA, TS-DRD achieves competitive or superior performance to existing SFDA methods that still use source models, demonstrating its effectiveness and the potential of the VODA setting.

A hierarchical cross-modal fusion model is proposed for vision-language question answering (VLQA) in industrial robotics, targeting the challenges of semantic ambiguity, complex environmental layouts, and domain-specific terminology common in modern manufacturing. The framework integrates advanced object detection, multi-scale visual encoding, syntactic parsing, and task-aware semantic attention to unite vision and language signals into a joint reasoning space. Region-based deep networks extract visual features, weighted embeddings aggregate, and recurrent neural parsing encodes sentence structures. Through fine-grained semantic alignment driven by adaptive fusion and cross-attention mechanisms, the system can handle operational queries, instruction steps, and anomaly detection with higher reliability. Compared to the existing VLQA benchmarks, validation experiments conducted on the IVQA and RIF benchmarks indicate improvements in semantic alignment, Top-1 accuracy, and robustness to ambiguous or procedural task queries. Ablation studies further quantify the impact of each architectural module, confirming the necessity of multi-level feature integration and context-driven gating for dependable industrial deployment. The technical advancements reported here provide core methodologies to improve the interpretability and operational effectiveness of industrial robots faced with diverse human-robot interaction tasks.

Traditional Shot Boundary Detection (SBD) inherently struggles with complex transitions by formulating the task around isolated cut points, frequently yielding corrupted video shots. We address this fundamental limitation by formalizing the Shot Transition Detection (STD) task. Rather than searching for ambiguous points, STD explicitly detects the continuous temporal segments of transitions. To tackle this, we propose TransVLM, a Vision-Language Model (VLM) framework for STD. Unlike regular VLMs that predominantly rely on spatial semantics and struggle with fine-grained inter-shot dynamics, our method explicitly injects optical flow as a critical motion prior at the input stage. Through a simple yet effective feature-fusion strategy, TransVLM directly processes concatenated color and motion representations, significantly enhancing its temporal awareness without incurring any additional visual token overhead on the language backbone. To overcome the severe class imbalance in public data, we design a scalable data engine to synthesize diverse transition videos for robust training, alongside a comprehensive benchmark for STD. Extensive experiments demonstrate that TransVLM achieves superior overall performance, outperforming traditional heuristic methods, specialized spatiotemporal networks, and top-tier VLMs. This work has been deployed to production. For more related research, please visit HeyGen Research (https://www.heygen.com/research) and HeyGen Avatar-V (https://www.heygen.com/research/avatar-v-model). Project page: https://chence17.github.io/TransVLM/

As Vision-Language Models (VLMs) become increasingly integrated into decision-making systems, it is essential to understand how visual inputs influence their behavior. This paper investigates the effects of visual priming on VLMs' cooperative behavior using the Iterated Prisoner's Dilemma (IPD) as a test scenario. We examine whether exposure to images depicting behavioral concepts (kindness/helpfulness vs. aggressiveness/selfishness) and color-coded reward matrices alters VLM decision patterns. Experiments were conducted across multiple state-of-the-art VLMs. We further explore mitigation strategies including prompt modifications, Chain of Thought (CoT) reasoning, and visual token reduction. Results show that VLM behavior can be influenced by both image content and color cues, with varying susceptibility and mitigation effectiveness across models. These findings not only underscore the importance of robust evaluation frameworks for VLM deployment in visually rich and safety-critical environments, but also highlight how architectural and training differences among models may lead to distinct behavioral responses-an area worthy of further investigation.

Long-context LLM serving is bottlenecked by the cost of attending over ever-growing KV caches. Dynamic sparse attention promises relief by accessing only a small, query-dependent subset of the KV state per decoding step and extending the KV storage to CPU memory. In practice, however, these algorithmic savings rarely translate into end-to-end system-level gains because sparse methods typically operate at different granularities and thus rely on ad hoc, per-algorithm implementations. At the same time, hierarchical KV storage introduces a new systems bottleneck: retrieving fine-grained, irregular KV subsets across the GPU-CPU boundary can easily erase the benefits of sparsity. We present SPIN, a sparse-attention-aware inference framework that co-designs the execution pipeline with hierarchical KV storage through three techniques: (1) a unified partition abstraction that maps different sparsity granularities onto a shared page-based KV substrate; (2) a locality-aware KV cache manager that dynamically sizes per-request HBM budgets and uses a GPU-friendly bucketed LRU policy to cut PCIe round-trips; and (3) a two-level hierarchical metadata layout sized to the active working set rather than the worst-case address space. Built on vLLM with three representative sparse attention algorithms, SPIN delivers 1.66-5.66x higher end-to-end throughput and 7-9x lower TTFT than vLLM, and reduces TPOT by up to 58% over the original sparse-attention implementations.

Vision-language models have shown strong performance, but they often generalize poorly to specialized domains. While semi-supervised vision-language learning mitigates this limitation by leveraging a small set of labeled image-text pairs together with abundant unlabeled images, existing methods remain fundamentally pairwise and fail to model the global structure of multimodal representation manifolds. Existing topology-based alignment methods rely on persistence diagram matching, which neither guarantees geometric alignment nor utilizes the image-text pairing information central to vision-language learning. We propose Topology-Aware Multimodal Representation Alignment (ToMA), a framework that uses persistent homology to identify topologically salient edges and aligns them across modalities through available cross-modal correspondences. ToMA leverages both H_0-death edges and lightweight H_1-birth edges, allowing it to capture both connectivity and cycle structure without constructing 2-simplices. Experiments show that ToMA yields stable gains, with clear improvements on remote sensing and modest but consistent benefits on fashion retrieval. Additional analysis shows that ToMA is more stable than alternative topology-based objectives and that lightweight H_1-birth edges provide useful higher-order structural signals.

Vision-Language Models (VLMs) exhibit strong performance in instruction following and open-ended vision-language reasoning, yet they frequently generate fluent outputs that are weakly grounded in visual evidence. Prior works have shown that instruction prompting further worsens this issue by amplifying language priors, especially when the visual signal is uncertain or ambiguous. To address this challenge, we propose a decoding framework that explicitly balances linguistic informativeness and visual faithfulness during generation. Our method, Instruction-Evidence Contrastive Dual-Stream Decoding (IECD$^2$), maintains two parallel probability distribution of tokens at each decoding step: an instruction-driven stream that promotes expressive and informative responses, and an evidence-driven stream that enforces strict grounding in the image. These two streams are adaptively fused using a symmetric KL-based contrastive gate, which suppresses tokens favored by language priors but unsupported by visual evidence, while preserving them when both distributions agree. We evaluate IECD$^2$ on multiple datasets spanning various generative vision-language reasoning tasks such as captioning and visual question answering on multiple datasets such as, POPE, MME, VQAv2, AMBER, and MSCOCO. IECD$^2$ demonstrates consistent improvements in task accuracy and reasoning performance with substantial reduction in hallucination compared to state-of-the-art decoding approaches.

Vision-Language-Navigation (VLN) models exhibit excellent navigation accuracy but incur high computational overhead. Token caching has emerged as a promising training-free strategy to reduce this cost by reusing token computation results; however, existing token caching approaches rely on visual domain methods for cacheable token selection, leading to challenges when adapted to VLN models. 1) Visual domain methods become invalid when there is viewpoint migration. 2) Visual domain methods neglect critical edge information without the aid of additional algorithms. 3) Visual domain methods overlook the temporal variation of scenarios and lack adjustability in cache budgets. In this paper, we develop detailed analyses and find that the impacts of these challenges exhibit invariance and analyzability in the frequency domain. Based on these, we propose a frequency-guided token caching framework, called FreqCache. Utilizing the inherent properties of the frequency domain, FreqCache achieves optimal token cache establishment, refreshment, and adaptive adjustment. Experiments show that FreqCache achieves 1.59x speedup with ignorable overhead, showing the effect of integrating frequency domain methods in VLN token caching.

Vision-language models (VLMs) are increasingly deployed as evaluators in tasks requiring nuanced image understanding, yet their reliability in scoring alignment between images and text descriptions remains underexplored. We investigate whether small, open-weight VLMs exhibit \emph{sycophantic} behavior when evaluating image-text alignment: assigning high scores without grounding their judgments in visual evidence. To quantify this phenomenon, we introduce the \emph{Bluffing Coefficient} (\bc), a metric that measures the mismatch between a model's score and its evidence recall. We evaluate six open-weight VLMs ranging from 450M to 8B parameters on a benchmark of 173,810 AI-generated character portraits paired with detailed textual descriptions. Our analysis reveals a significant inverse correlation between model size and sycophancy rate ($r = -0.96$, $p = 0.002$), with smaller models exhibiting substantially higher rates of unjustified high scores. The smallest model tested (LFM2-VL, 450M) produced sycophantic evaluations in 22.3% of cases, compared to 6.0% for the largest (LLaVA-1.6, 7B). These findings have direct implications for the deployment of small, open-weight VLMs as automated evaluators within attribute-rich, synthetic image evaluation tasks, where the gap between assigned scores and cited visual evidence is both measurable and consequential.

Large (vision-)language models exhibit remarkable capability but remain highly susceptible to jailbreaking. Existing safety training approaches aim to have the model learn a refusal boundary between safe and unsafe, based on the user's intent. It has been found that this binary training regime often leads to brittleness, since the user intent cannot reliably be evaluated, especially if the attacker obfuscates their intent, and also makes the system seem unhelpful. In response, frontier models, such as GPT-5, have shifted from refusal-based safeguards to safe completion, that aims to maximize helpfulness while obeying safety constraints. However, safe completion could be exploited when a user pretends their intention is benign. Specifically, this intent inversion would be effective in multi-turn conversation, where the attacker has multiple opportunities to reinforce their deceptively benign intent. In this work, we introduce a novel multi-turn jailbreaking method that exploits this vulnerability. Our approach gradually builds conversational trust by simulating benign-seeming intentions and by exploiting the consistency property of the model, ultimately guiding the target model toward harmful, detailed outputs. Most crucially, our approach also uncovered an additional class of model vulnerability that we call para-jailbreaking that has been unnoticed up to now. Para-jailbreaking describes the situation where the model may not reveal harmful direct reply to the attack query, however the information that it reveals is nevertheless harmful. Our contributions are threefold. First, it achieves high success rates against frontier models including GPT-5-thinking and Claude-Sonnet-4.5. Second, our approach revealed and addressed para-jailbreaking harmful output. Third, experiments on multimodal VLM models showed that our approach outperformed state-of-the-art models.

Chart-to-code generation demands strict visual precision and syntactic correctness from Vision-Language Models (VLMs). However, existing approaches are fundamentally constrained by data-centric limitations: despite the availability of growing chart-to-code datasets, simply scaling homogeneous chart-code pairs conflates visual perception with program logic, preventing models from fully leveraging the richness of multimodal supervision. We present CharTide, a novel data-centric framework that systematically redesigns both training and alignment data for chart-to-code generation. First, we construct a 2M-sample dataset via a Tri-Perspective Tuning strategy, explicitly decoupling training into visual perception, pure-text code logic, and modality fusion streams, enabling a 7B model to surpass specialized baselines using only supervised data. Second, we reformulate alignment as a data verification problem rather than a heuristic scoring task. To this end, we introduce an Inquiry-Driven RL framework grounded in the principle of information invariance: a downstream model should yield consistent answers to identical visual queries across both original and generated charts. Moving beyond rigid rule matching or VLM scoring, we employ a frozen Inspector to objectively verify generated charts through atomic QA tasks, providing verifiable reward signals based on answer accuracy. Experiments on ChartMimic, Plot2Code, and ChartX show that CharTide-7B/8B significantly outperforms open-source baselines, surpasses GPT-4o, and is competitive with GPT-5.

Recent advances in embodied Vision-Language Agentic Systems (VLAS), powered by large vision-language models (LVLMs), enable AI systems to perceive and reason over real-world scenes. Within this context, environmental signals such as traffic lights are essential in-band signals that can and should influence agent behavior. However, similar signals could also be crafted to operate as misleading visual injections, overriding user intent and posing security risks. This duality creates a fundamental challenge: agents must respond to legitimate environmental cues while remaining robust to misleading ones. We refer to this tension as trust boundary confusion. To study this behavior, we design a dual-intent dataset and evaluation framework, through which we show that current LVLM-based agents fail to reliably balance this trade-off, either ignoring useful signals or following harmful ones. We systematically evaluate 7 LVLM agents across multiple embodied settings under both structure-based and noise-based visual injections. To address these vulnerabilities, we propose a multi-agent defense framework that separates perception from decision-making to dynamically assess the reliability of visual inputs. Our approach significantly reduces misleading behaviors while preserving correct responses and provides robustness guarantees under adversarial perturbations. The code of the evaluation framework and artifacts are made available at https://anonymous.4open.science/r/Visual-Prompt-Inject.

Chain-of-Thought (CoT) reasoning has become a powerful driver of trajectory prediction in VLA-based autonomous driving, yet its autoregressive nature imposes a latency cost that is prohibitive for real-time deployment. Latent CoT methods attempt to close this gap by compressing reasoning into continuous hidden states, but consistently fall short of their explicit counterparts. We suggest that this is due to purely linguistic latent representations compressing a symbolic abstraction of the world, rather than the causal dynamics that actually govern driving. Thus, we present OneVL (One-step latent reasoning and planning with Vision-Language explanations), a unified VLA and World Model framework that routes reasoning through compact latent tokens supervised by dual auxiliary decoders. Alongside a language decoder that reconstructs text CoT, we introduce a visual world model decoder that predicts future-frame tokens, forcing the latent space to internalize the causal dynamics of road geometry, agent motion, and environmental change. A three-stage training pipeline progressively aligns these latents with trajectory, language, and visual objectives, ensuring stable joint optimization. In inference, the auxiliary decoders are discarded, and all latent tokens are prefilled in a single parallel pass, matching the speed of answer-only prediction. Across four benchmarks, OneVL becomes the first latent CoT method to surpass explicit CoT, delivering superior accuracy at answer-only latency. These results show that with world model supervision, latent CoT produces more generalizable representations than verbose token-by-token reasoning. Code has been open-sourced to the community. Project Page: https://xiaomi-embodied-intelligence.github.io/OneVL

Vision-language modeling is rapidly increasing in popularity with an ever expanding list of available models. In most cases, these vision-language models have parameters in the tens of billions, which is necessary for some needs, but in many cases smaller models are necessary (e.g., on edge devices or independent robotic platforms). Unfortunately, there is little research in producing light-weight models or in training them with small datasets. Inspired by the language learning progression and data sparsity in child development, in this paper, we address both of these goals in a systematic fashion. We show that two-tower encoder models are superior to one-tower encoders in low-resource settings for discriminative English tasks. We show also that incorporating traditional convolutional networks into the two-tower transformer architecture can help produce parameter efficient vision-language models. Finally, we show that the cross-modal fusion module of two-tower encoders can vary significantly in shape and size while producing the same results. In addition, we present ESsEN, a compact vision-language model that can be trained end-to-end with relatively few resources that performs as well on several tasks with only a fraction of the parameters compared to other models. The experimental results and the tools we present here make vision-language modeling more accessible to a wider variety of researchers.

Vision Language Models (VLMs) achieved rapid progress in the recent years. However, despite their growth, VLMs development is heavily grounded on English, leading to two main limitations: (i) the lack of multilingual and multimodal datasets for training, and (ii) the scarcity of comprehensive evaluation benchmarks across languages. In this work, we address these gaps by introducing a new comprehensive suite of resources for VLMs training and evaluation spanning five European languages (English, French, German, Italian, and Spanish). We adopt a regeneration-translation paradigm that produces high-quality cross-lingual resources by combining curated synthetic generation and manual annotation. Specifically, we build Multi-PixMo, a training corpus obtained regenerating examples from Pixmo pre-existing datasets with permissively licensed models: PixMo-Cap, PixMo-AskModelAnything, and CoSyn-400k. On the evaluation side, we construct a set of multilingual benchmarks derived translating widely used English datasets (MMbench, ScienceQA, MME, POPE, AI2D). We assess the quality of these resources through qualitative and quantitative human analyses, measuring inter-annotator agreement. Additionally, we perform ablation studies to demonstrate the impact of multilingual data, with respect to English only, in VLMs training. Experiments, comprising 3 different models show that using multilingual, multimodal examples for training VLMs aids is consistently beneficial on non-English benchmarks, with positive transfer to English as well.

Vision-Language-Action (VLA) models have gained much attention from the research community thanks to their strength in translating multimodal observations with linguistic instructions into desired robotic actions. Despite their advancements, VLAs often overlook explicit reasoning and learn the functional input-action mappings, omitting crucial logical steps, which are especially pronounced in interpretability and generalization for complex, long-horizon manipulation tasks. In this work, we propose ReFineVLA, a multimodal reasoning-aware framework that fine-tunes VLAs with teacher-guided reasons. We first augment robotic datasets with reasoning rationales generated by an expert teacher model, guiding VLA models to learn to reason about their actions. Then, we fine-tune pre-trained VLAs with the reasoning-enriched datasets with ReFineVLA, while maintaining the underlying generalization abilities and boosting reasoning capabilities. We also conduct attention map visualization to analyze the alignment among visual observation, linguistic prompts, and to-be-executed actions of ReFineVLA, reflecting the model is ability to focus on relevant tasks and actions. Through this additional step, we explore that ReFineVLA-trained models exhibit a meaningful agreement between vision-language and action domains, highlighting the enhanced multimodal understanding and generalization. Evaluated across a suite of simulated manipulation benchmarks on SimplerEnv with both WidowX and Google Robot tasks, ReFineVLA achieves state-of-the-art performance, in success rate over the second-best method on the both the WidowX benchmark and Google Robot Tasks.

The reliability of VLM-as-a-Judge is critical for the automatic evaluation of vision-language models (VLMs). Despite recent progress, our analysis reveals that VLM-as-a-Judge often pays limited attention to the image when making decisions. Instead, they often blindly favor the more informative answer, even when they can recognize it conflicts with the image content. We call this problem informativeness bias, which significantly undermines judge reliability. To address it, we propose BIRCH (Balanced Informativeness and CoRrectness with a Truthful AnCHor), a judging paradigm that first corrects inconsistencies with the image content in candidate answers, and then compares the answers against this corrected version. This shifts the judge's focus from informativeness to image-grounded correctness. Experiments on multiple models and benchmarks show that BIRCH reduces informativeness bias by up to 17%, resulting in performance gains of up to 9.8%. Our work reveals an overlooked but fundamental flaw in current VLM-as-a-Judge systems and highlights the need for more principled designs.

Source-Free Domain Adaptation (SFDA) seeks to adapt a source model, which is pre-trained on a supervised source domain, for a target domain, with only access to unlabeled target training data. Relying on pseudo labeling and/or auxiliary supervision, conventional methods are inevitably error-prone. To mitigate this limitation, in this work we for the first time explore the potentials of off-the-shelf vision-language (ViL) multimodal models (e.g., CLIP) with rich whilst heterogeneous knowledge. We find that directly applying the ViL model to the target domain in a zero-shot fashion is unsatisfactory, as it is not specialized for this particular task but largely generic. To make it task-specific, we propose a novel DIFO++ approach. Specifically, DIFO++ alternates between two steps during adaptation: (i) Customizing the ViL model by maximizing the mutual information with the target model in a prompt learning manner, (ii) Distilling the knowledge of this customized ViL model to the target model, centering on gap region reduction. During progressive knowledge adaptation, we first identify and focus on the gap region, where enclosed features are entangled and class-ambiguous, as it often captures richer task-specific semantics. Reliable pseudo-labels are then generated by fusing predictions from the target and ViL models, supported by a memory mechanism. Finally, gap region reduction is guided by category attention and predictive consistency for semantic alignment, complemented by referenced entropy minimization to suppress uncertainty. Extensive experiments show that DIFO++ significantly outperforms the state-of-the-art alternatives. Our code and data are available at https://github.com/tntek/DIFO-Plus.

Implicit spatial relations and deep semantic structures encoded in object attributes are crucial for procedural planning in embodied AI systems. However, existing approaches often over rely on the reasoning capabilities of vision language models (VLMs) themselves, while overlooking the rich structured semantic information that can be mined from multimodal inputs. As a result, models struggle to effectively understand functional spatial relationships in complex scenes. To fully exploit implicit spatial relations and deep semantic structures in multimodal data, we propose GaLa, a vision language framework for multimodal procedural planning. GaLa introduces a hypergraph-based representation, where object instances in the image are modeled as nodes, and region-level hyperedges are constructed by aggregating objects according to their attributes and functional semantics. This design explicitly captures implicit semantic relations among objects as well as the hierarchical organization of functional regions. Furthermore, we design a TriView HyperGraph Encoder that enforces semantic consistency across the node view, area view, and node area association view via contrastive learning, enabling hypergraph semantics to be more effectively injected into downstream VLM reasoning. Extensive experiments on the ActPlan1K and ALFRED benchmarks demonstrate that GaLa significantly outperforms existing methods in terms of execution success rate, LCS, and planning correctness.

Imitation learning has enabled robots to acquire complex visuomotor manipulation skills from demonstrations, but deployment failures remain a major obstacle, especially for long-horizon action-chunked policies. Once execution drifts off the demonstration manifold, these policies often continue producing locally plausible actions without recovering from the failure. Existing runtime monitors either require failure data, over-trigger under benign feature drift, or stop at failure detection without providing a recovery mechanism. We present Rewind-IL, a training-free online safeguard framework for generative action-chunked imitation policies. Rewind-IL combines a zero-shot failure detector based on Temporal Inter-chunk Discrepancy Estimate (TIDE), calibrated with split conformal prediction, with a state-respawning mechanism that returns the robot to a semantically verified safe intermediate state. Offline, a vision-language model identifies recovery checkpoints in demonstrations, and the frozen policy encoder is used to construct a compact checkpoint feature database. Online, Rewind-IL monitors self-consistency in overlapping action chunks, tracks similarity to the checkpoint library, and, upon failure, rewinds execution to the latest verified safe state before restarting inference from a clean policy state. Experiments on real-world and simulated long-horizon manipulation tasks, including transfer to flow-matching action-chunked policies, demonstrate that policy-internal consistency coupled with semantically grounded respawning offers a practical route to improved reliability in imitation learning. Supplemental materials are available at https://sjay05.github.io/rewind-il

Multimodal large language models (MLLMs) have achieved impressive progress on vision language benchmarks, yet their capacity for visual cognitive and visuospatial reasoning remains less understood. We introduce "Mind's Eye", a multiple-choice benchmark of eight visuo-cognitive tasks inspired by classic human intelligence tests and organized under a novel "A-R-T" taxonomy: Abstraction, Relation, and Transformation. The tasks probe core processes of fluid intelligence such as pattern induction, analogical relation mapping, and mental transformation. We evaluate a diverse suite of closed-source and open-source MLLMs and compare their performance with human participants. Humans achieve 80% accuracy, while top performing MLLMs remain below 50%. Error analysis reveals failures in: (i) visual attention allocation, (ii) internal perceptual manipulation, and (iii) weak abstraction of underlying visual concepts. Our findings suggest that current MLLMs exhibit limited visuospatial reasoning capabilities, when compared with human participants, highlighting the need for more cognitively grounded evaluation frameworks.

Vision Language Models (VLMs) have been applied to several specific domains and have shown strong problem-solving capabilities. However, astronomical imaging, a quite complex problem involving multidisciplinary knowledge and several subtasks, has not been adequately studied. Due to the complexity of the astronomical imaging process, both world-class astronomical organizations, such as NASA, and expert enthusiasts devote a great deal of time and effort. This is because the processes in astronomical imaging have complex underlying correlations that significantly influence one another, making the quality diagnosis and error localization of astronomical images challenging. To address this problem, we propose AstroVLM, a collaborative multi-agent system for diagnosing the quality of astronomical images. Experiment results show that AstroVLM outperforms all baselines on real-world astronomical imaging quality diagnosis tasks, providing a reference for language models to handle complicated multi-process tasks.

Large Vision-Language Models (LVLMs) consistently require new arenas to guide their expanding boundaries, yet their capabilities with hypergraphs remain unexplored. In the real world, hypergraphs have significant practical applications in areas such as life sciences and social communities. Recent advancements in LVLMs have shown promise in understanding complex topologies, yet there remains a lack of a benchmark to delineate the capabilities of LVLMs with hypergraphs, leaving the boundaries of their abilities unclear. To fill this gap, in this paper, we introduce $\texttt{HyperGVL}$, the first benchmark to evaluate the proficiency of LVLMs in hypergraph understanding and reasoning. $\texttt{HyperGVL}$ provides a comprehensive assessment of 12 advanced LVLMs across 84,000 vision-language question-answering (QA) samples spanning 12 tasks, ranging from basic component counting to complex NP-hard problem reasoning. The involved hypergraphs contain multiscale synthetic structures and real-world citation and protein networks. Moreover, we examine the effects of 12 textual and visual hypergraph representations and introduce a generalizable router $\texttt{WiseHyGR}$ that improves LVLMs in hypergraph via learning adaptive representations. We believe that this work is a step forward in connecting hypergraphs with LVLMs.

Large Language Model (LLM) deployment is increasingly shifting to cost-efficient accelerators like Google's Tensor Processing Units (TPUs), prioritizing both performance and total cost of ownership (TCO). However, existing LLM inference kernels and serving systems remain largely GPU-centric, and there is no well-established approach for efficiently mapping LLM workloads onto TPU architectures--particularly under the dynamic and ragged execution patterns common in modern serving. In this paper, we present Ragged Paged Attention (RPA), a high-performance and flexible attention kernel for TPUs, implemented using Pallas and Mosaic. RPA addresses these challenges through three key techniques: (1) fine-grained tiling to enable efficient dynamic slicing over ragged memory, (2) a custom software pipeline that fuses KV cache updates with attention computation, and (3) a distribution-aware compilation strategy that generates specialized kernels for decode, prefill, and mixed workloads. Evaluated on Llama 3 8B on TPU7x, RPA achieves up to 86% memory bandwidth utilization (MBU) in decode and 73% model FLOPs utilization (MFU) in prefill. Integrated as the primary TPU backend in vLLM and SGLang, RPA provides a production-grade foundation for efficient TPU inference and offers practical insights into kernel design.

Over the past year, spatial intelligence has drawn increasing attention. Many prior works study it from the perspective of visual-spatial intelligence, where models have access to visuospatial information from visual inputs. However, in the absence of visual information, whether linguistic intelligence alone is sufficient to endow models with spatial intelligence, and how models perform relevant tasks with text-only inputs still remain unexplored. Therefore, in this paper, we focus on a fundamental and critical capability in spatial intelligence from a linguistic perspective: viewpoint rotation understanding (VRU). Specifically, LLMs and VLMs are asked to infer their final viewpoint and predict the corresponding observation in an environment given textual description of viewpoint rotation and observation over multiple steps. We find that both LLMs and VLMs perform poorly on our proposed dataset while human can easily achieve 100% accuracy, indicating a substantial gap between current model capabilities and the requirements of spatial intelligence. To uncover the underlying mechanisms, we conduct a layer-wise probing analysis and head-wise causal intervention. Our findings reveal that although models encode viewpoint information in the hidden states, they appear to struggle to bind the viewpoint position with corresponding observation, resulting in a hallucination in final layers. Finally, we selectively fine-tune the key attention heads identified by causal intervention to improve VRU performance. Experimental results demonstrate that such selective fine-tuning achieves improved VRU performance while avoiding catastrophic forgetting of generic abilities. Our dataset and code will be released at https://github.com/Young-Zhen/VRU_Interpret .

Understanding emotions is a fundamental ability for intelligent systems to be able to interact with humans. Vision-language models (VLMs) have made tremendous progress in the last few years for many visual tasks, potentially offering a promising solution for understanding emotions. However, it is surprising that even the most sophisticated contemporary VLMs struggle to recognize human emotions or to outperform even specialized vision-only classifiers. In this paper we ask the question "Why do VLMs struggle to recognize human emotions?", and observe that the inherently continuous and dynamic task of facial expression recognition (DFER) exposes two critical VLM vulnerabilities. First, emotion datasets are naturally long-tailed, and the web-scale data used to pre-train VLMs exacerbates this head-class bias, causing them to systematically collapse rare, under-represented emotions into common categories. We propose alternative sampling strategies that prevent favoring common concepts. Second, temporal information is critical for understanding emotions. However, VLMs are unable to represent temporal information over dense frame sequences, as they are limited by context size and the number of tokens that can fit in memory, which poses a clear challenge for emotion recognition. We demonstrate that the sparse temporal sampling strategy used in VLMs is inherently misaligned with the fleeting nature of micro-expressions (0.25-0.5 seconds), which are often the most critical affective signal. As a diagnostic probe, we propose a multi-stage context enrichment strategy that utilizes the information from "in-between" frames by first converting them into natural language summaries. This enriched textual context is provided as input to the VLM alongside sparse keyframes, preventing attentional dilution from excessive visual data while preserving the emotional trajectory.

Vision-Language Models (VLMs) have shown remarkable capabilities in joint vision-language understanding, but their large scale poses significant challenges for deployment in resource-constrained scenarios. Knowledge Distillation (KD) offers a viable way to improve model capabilities without increasing model size or data requirements, making deployment more efficient. However, applying KD to VLMs is challenged by modality-specific supervision: although multimodal knowledge in VLMs is fused within the language space, current methods supervise each modality separately without explicitly addressing multimodal alignment, leading to inconsistent multimodal knowledge transfer. To address this, we propose Switch-KD, a visual-switch distillation framework that unifies vision-language knowledge transfer within a shared text-probability space. Switch-KD comprises two key components: (1) Visual-Switch Distillation, which switches the student's visual outputs into the teacher's language pathway to construct cross-modal probabilistic references for implicit visual knowledge transfer; and (2) Dynamic Bi-directional Logits Difference (DBiLD) loss, which adaptively aligns informative probability regions while preserving the distributional structures of teacher and student through bidirectional supervision. Guided by Switch-KD, a 0.5B TinyLLaVA effectively distills rich multimodal knowledge from its 3B teacher, yielding an average improvement of 3.6 points across 10 multimodal benchmarks without any architectural modification.

Human perception of visual similarity is inherently adaptive and subjective, depending on the users' interests and focus. However, most image retrieval systems fail to reflect this flexibility, relying on a fixed, monolithic metric that cannot incorporate multiple conditions simultaneously. To address this, we propose CLAY, an adaptive similarity computation method that reframes the embedding space of pretrained Vision-Language Models (VLMs) as a text-conditional similarity space without additional training. This design separates the textual conditioning process and visual feature extraction, allowing highly efficient and multi-conditioned retrieval with fixed visual embeddings. We also construct a synthetic evaluation dataset CLAY-EVAL, for comprehensive assessment under diverse conditioned retrieval settings. Experiments on standard datasets and our proposed dataset show that CLAY achieves high retrieval accuracy and notable computational efficiency compared to previous works.

Vision-Language Models (VLMs) have been applied to a wide range of reasoning tasks, yet it remains unclear whether they can reason robustly under distribution shifts. In this paper, we study covariate shifts in which the perceptual input distribution changes while the underlying prediction rules do not. To investigate this question, we consider visual deductive reasoning tasks, where a model is required to answer a query given an image and logical rules defined over the object concepts in the image. Empirically, we find that VLMs fine-tuned through gradient-based end-to-end training can achieve high in-distribution accuracy but fail to generalize under such shifts, suggesting that fine-tuning does not reliably induce the underlying reasoning function. This motivates a neuro-symbolic perspective that decouples perception from reasoning. However, we further observe that recent neuro-symbolic approaches that rely on black-box components for reasoning can still exhibit inconsistent robustness across tasks. To address this issue, we propose VLC, a neuro-symbolic method that combines VLM-based concept recognition with circuit-based symbolic reasoning. In particular, task rules are compiled into a symbolic program, specifically a circuit, which executes the rules exactly over the object concepts recognized by the VLM. Experiments on three simple visual deductive reasoning tasks with distinct rule sets show that VLC consistently achieves higher task accuracy on out-of-distribution data than other reasoning paradigms. Code is available at https://github.com/uiuctml/VLC.

Material classification has emerged as a critical task in computer vision and graphics, supporting the assignment of accurate material properties to a wide range of digital and real-world applications. While traditionally framed as an image classification task, this domain faces significant challenges due to the scarcity of annotated data, limiting the accuracy and generalizability of trained models. Recent advances in vision-language foundation models (VLMs) offer promising avenues to address these issues, yet existing solutions leveraging these models still exhibit unsatisfying results in material recognition tasks. In this work, we propose a novel framework that effectively harnesses foundation models to overcome data limitations and enhance classification accuracy. Our method integrates two key innovations: (a) a robust image generation and auto-labeling pipeline that creates a diverse and high-quality training dataset with material-centric images, and automatically assigns labels by fusing object semantics and material attributes in text prompts; (b) a prior incorporation strategy to distill information from VLMs, combined with a joint fine-tuning method that optimizes a pre-trained vision foundation model alongside VLM-derived priors, preserving broad generalizability while adapting to material-specific features. Extensive experiments demonstrate significant improvements on multiple datasets. We show that our synthetic dataset effectively captures the characteristics of real world materials, and the integration of priors from vision-language models significantly enhances the final performance. The source code and dataset will be released.

Active perception - the ability of a robot to proactively select viewpoints to acquire task-relevant information - is essential for robust operation in real-world environments. However, existing approaches are typically limited to fixed objectives or constrained settings, and struggle to generalize to open-ended perception intents specified in natural language. We propose I-Perceive, a foundation model for language-conditioned active perception in large-scale indoor environments. Given a query image, a set of context images, and a natural language instruction, I-Perceive predicts a 6D camera pose that fulfills the specified perception intent. The model integrates a vision-language pathway for semantic grounding with a geometric reasoning pathway for multi-view 3D understanding, connected via multi-layer semantic fusion to enable language-conditioned geometric reasoning. To support scalable training, we construct a large-scale dataset of language-viewpoint pairs from both real-world scene-scanning data and simulated environments using an automated pipeline. Extensive experiments demonstrate that I-Perceive significantly outperforms strong baselines on prediction accuracy, viewpoint feasibility, and instructions alignment. The model exhibits strong zero-shot generalization to unseen scenes and instructions, and enables closed-loop active perception, progressively refining viewpoints over sequential interactions.

While open sourced Vision-Language Models (VLMs) have proliferated, selecting the optimal pretrained model for a specific downstream task remains challenging. Exhaustive evaluation is often infeasible due to computational constraints and data limitations in few shot scenarios. Existing selection methods fail to fully address this: they either rely on data-intensive proxies or use symmetric textual descriptors that neglect the inherently directional and model-specific nature of transferability. To address this problem, we propose a framework that grounds model selection in the internal functional dynamics of the visual encoder. Our approach represents each task via layer wise conductance and derives a target-conditioned block importance distribution through entropy regularized alignment. Building on this, we introduce Directional Conductance Divergence (DCD), an asymmetric metric that quantifies how effectively a source task covers the target's salient functional blocks. This allows for predicting target model rankings by aggregating source task ranks without direct inference. Experimental results on 48 VLMs across 21 datasets demonstrate that our method outperforms state-of-the-art baselines, achieving a 14.7% improvement in NDCG@5 over SWAB.

Multimodal large language models (MLLMs) have achieved impressive progress in vision-language reasoning, yet their ability to understand temporally unfolding narratives in videos remains underexplored. True narrative understanding requires grounding who is doing what, when, and where, maintaining coherent entity representations across dynamic visual and temporal contexts. We introduce NarrativeTrack, the first benchmark to evaluate narrative understanding in MLLMs through fine-grained entity-centric reasoning. Unlike existing benchmarks limited to short clips or coarse scene-level semantics, we decompose videos into constituent entities and examine their continuity via a Compositional Reasoning Progression (CRP), a structured evaluation framework that progressively increases narrative complexity across three dimensions: entity existence, entity changes, and entity ambiguity. CRP challenges models to advance from temporal persistence to contextual evolution and fine-grained perceptual reasoning. A fully automated entity-centric pipeline enables scalable extraction of temporally grounded entity representations, providing the foundation for CRP. Evaluations of state-of-the-art MLLMs reveal that models fail to robustly track entities across visual transitions and temporal dynamics, often hallucinating identity under context shifts. Open-source general-purpose MLLMs exhibit strong perceptual grounding but weak temporal coherence, while video-specific MLLMs capture temporal context yet hallucinate entities' contexts. These findings uncover a fundamental trade-off between perceptual grounding and temporal reasoning, indicating that narrative understanding emerges only from their integration. NarrativeTrack provides the first systematic framework to diagnose and advance temporally grounded narrative comprehension in MLLMs.

Spatial reasoning is a critical capability for intelligent robots, yet current vision-language models (VLMs) still fall short of human-level performance in video-based spatial reasoning. This gap mainly stems from two challenges: a semantic-geometric misalignment that prevents consistent 3D understanding, and the absence of persistent memory to retain 3D representation and understanding across frames. To address these limitations, we present VLM$^2$, a Vision-Language Model with persistent Memory for spatial reasoning with a view-consistent, 3D-aware representation purely from 2D videos. Specifically, we incorporate a dual-memory module consisting of a working memory that operates as a sliding window to focus on immediate context, and an episodic memory that consolidates and stores critical information across frames. This design enables bounded and efficient spatial reasoning under a fixed computational cost. Extensive experiments on multiple benchmarks show that VLM$^2$ achieves state-of-the-art performance among video-based models, significantly advancing the frontier of visual-spatial intelligence.

Semantic scene understanding in robotics requires representations that are both metric-accurate and queryable via natural language in real-time. While recent Vision-Language Models enable powerful 2D image-text alignment, their integration into real-time 3D mapping systems remains challenging due to their requirements on ground truth poses, computational cost, and memory constraints. We present VLEM (Vision-Language Embedding Mapping), a real-time framework for integrating pixel-aligned 2D vision-language embeddings from various backends into a globally consistent, metric-accurate 3D representation, requiring only a raw RGB-D stream. Compared to ConceptFusion, Open-Fusion, and RayFronts, VLEM provides better open-set segmentation performance and a more compact representation. We further demonstrate VLEM's versatility in interactive real-time robotic manipulation tasks and mobile mapping scenarios.

We investigate whether socioeconomic indicators, like household wealth, leave recoverable informational imprints in both satellite imagery (capturing features like buildings and roads) and Internet-sourced text (reflecting historical, cultural, and narratives of neighborhoods). Using DHS data from African neighborhoods (clusters), we pair high-resolution Landsat images with textual descriptions generated by LLMs conditioned on location/year, plus text retrieved by an LLM-driven AI Search Agent from web sources. We develop a multimodal framework that predicts household wealth (International Wealth Index; IWI) via five pipelines: (i) a vision model on satellite images, (ii) an LLM using only location and year, (iii) an AI agent that searches and synthesizes web text, (iv) a joint image-text encoder, and (v) an ensemble of all signals. Our framework yields three contributions. First, evaluations show that fusing vision and agent/LLM-generated text improves on vision-only baselines in wealth prediction (e.g., R-squared of 0.77 vs. 0.63 on out-of-sample splits), with LLM-internal knowledge (artificial neural memory) proving surprisingly predictive in out-of-country/time generalization. Second, we find suggestive evidence of partial representational alignment: fused embeddings from vision and language modalities correlate moderately (median cosine similarity across modalities of about 0.60 after alignment). This pattern is broadly consistent with the Platonic Representation Hypothesis, but does not by itself establish convergence to a single shared latent representation. Because agent-retrieved data yields only marginal and unstable gains across splits, our evidence for the Agent-Induced Novelty Hypothesis is limited. Third, we release a large-scale multimodal dataset of about 60,000 DHS clusters, each linked to satellite images, LLM-generated descriptions, and AI-agent-retrieved texts.

As Visual Language Models (VLMs) become increasingly embedded in everyday applications, ensuring they can recognise and appropriately handle privacy-sensitive content is thus essential to protect users. To this end, we conduct a comprehensive evaluation of twelve state-of-the-art VLMs and identify limitations in their understanding of visual privacy. However, existing privacy-related datasets often suffer from label inconsistencies, limiting their reliability. To address this, we introduce two compact, high-quality benchmarks, PrivBench and PrivBench-H, that focus on commonly recognised visual privacy categories aligned with the General Data Protection Regulation (GDPR). Additionally, we present PrivTune, an instruction-tuning dataset specifically curated to improve privacy sensitivity. We obtain multiple Privacy VLMs by fine-tuning off-the-shelf VLMs on only a few hundred samples from PrivTune, which leads to substantial gains on all benchmarks, surpassing even GPT-4, while maintaining strong performance on other tasks. Our findings show that privacy-awareness in VLMs can be substantially improved with minimal data and careful dataset design, setting the stage for safer, more privacy-aligned AI systems.