Large Vision Language Models

Large Visual Language Models (LVLMs) have achieved remarkable success in vision tasks. However, the significant differences between industrial and natural scenes make applying LVLMs challenging. Existing LVLMs rely on user-provided prompts to segment objects. This often leads to suboptimal performance due to the inclusion of irrelevant pixels. In addition, the scarcity of data also makes the application of LVLMs in industrial scenarios remain unexplored. To fill this gap, this paper proposes an open industrial dataset and a Refined Text-Visual Prompt (RTVP) for zero-shot industrial defect detection. First, this paper constructs the Multi-Modal Industrial Open Dataset (MMIO) containing 80K+ samples. MMIO contains diverse industrial categories, including 6 super categories and 18 subcategories. MMIO is the first large-scale multi-scenes pre-training dataset for industrial zero-shot learning, and provides valuable training data for open models in future industrial scenarios. Based on MMIO, this paper provides a RTVP specifically for industrial zero-shot tasks. RTVP has two significant advantages: First, this paper designs an expert-guided large model domain adaptation mechanism and designs an industrial zero-shot method based on Mobile-SAM, which enhances the generalization ability of large models in industrial scenarios. Second, RTVP automatically generates visual prompts directly from images and considers text-visual prompt interactions ignored by previous LVLM, improving visual and textual content understanding. RTVP achieves SOTA with 42.2% and 24.7% AP in zero-shot and closed scenes of MMIO.

Large-scale Visual-Language Models (LVLMs) have achieved remarkable success in natural visual tasks, yet their application to industrial defect detection remains challenging due to two fundamental limitations: (i) the scarcity of large-scale industrial datasets that cover diverse defect categories across multiple domains, and (ii) the reliance on manual prompts (points, boxes, masks) that introduce subjective noise and lack text-visual interaction for fine-grained understanding. To address these challenges, we introduce a Large-Scale Multi-Modal Industrial Open-Closed benchmark (MMIOC-1M) containing over one million samples across $14$ super-categories, $29$ industrial scenes, and $351$ defect subcategories. To our knowledge, MMIOC-1M is the first unified largest benchmark supporting both open-vocabulary and closed-set industrial detection, providing valuable pre-training data for LVLMs in industrial scenarios. Furthermore, we propose a Refined Text-Visual Prompt Network (RTVPNet) that incorporates three key innovations: (1) an expert-assisted domain projection mechanism that enables rapid adaptation of general vision models to industrial domains, (2) an energy-based sparse sampling strategy that automatically generates refined visual prompts without manual intervention, and (3) a bidirectional text-visual interaction module that enhances cross-modal semantic alignment and understanding. Extensive experiments demonstrate that RTVPNet achieves state-of-the-art performance on MMIOC-1M, LVIS, and COCO benchmarks while maintaining computational efficiency. The dataset and code are available at https://github.com/hellozzk/MMIO.

Different vision neural networks -- trained to classify, contrast, reconstruct, or match images to text -- should have correspondingly different internal representations. We report that they do not. After training, the top sixteen principal directions of variation inside thirteen modern vision encoders converge to the same sixteen-dimensional geometric object. We call this the cross-architecture substrate and study it with PCA, centred kernel alignment (CKA), and Pang 2026 calibration. The substrate transports across four visual domains (natural photographs, medical CT, satellite, microscopy) at median Procrustes-CKA 0.679, and across eight domains (adding sketches, depth, thermal infrared, astronomy) at 0.604, every pair >0.40. It survives Pang calibration globally (7.4x disc-vs-MAE separation, n=13,394) and locally (4.82-5.30, p<10^{-44}). It is not pixel statistics (0.263), not Gabor features (0.31), not a random projection (0.041), and emerges in the first 10% of training while accuracy keeps climbing. We deliver four applications: a label-free transferability filter beating LogME (3x faster, +0.15 Kendall-tau); a four-way domain detector (99.6% accuracy); a frozen low-shot probe (16 dims beat 768-dim DINOv2 by 3.78pp at N=50 labels per class); and a teacher-free distillation auxiliary matching trained-teacher KD on 33 pairs (7.56pp peak gain at 10% label fraction). The substrate does not cross modalities, does not help cross-paradigm distillation, and does not predict transfer quality (rho=0.08 against transfer accuracy).

Recent vision-language models (VLMs) excel at multimodal understanding and reasoning, yet their fine-grained visual perception remains underexplored. A natural extension of ``How many r are there in Strawberry?'' asks: how small a visual pattern can a VLM reliably perceive? As such, we introduce FineSightBench, a new benchmark that systematically probes this limit by separating perception tasks (pixel-level recognition of letters, shapes, objects) from reasoning tasks (spatial reasoning, counting, ordering over small targets) across controlled scales of 4--48px. Through comprehensive experiments and detailed failure mode analysis on state-of-the-art models, we reveal a sharp dissociation: perception saturates around 12px, while reasoning remains limited even at larger scales, with persistent numeracy and sequence errors. These findings expose fundamental deficiencies in VLMs' fine-scale visual reasoning that demand more rigorous evaluation.

Recent breakthroughs in self-supervised learning (SSL), such as the Latent-Euclidean Joint-Embedding Predictive Architecture (LeJEPA), alongside successes in integrating visual encoders with language models, have driven the demand for adaptable, high-capacity vision encoders in Computed Tomography (CT). In this work, we explore 2D slice-based architectures as a flexible alternative to native 3D models for processing volumetric CT data. Using the CT-RATE dataset, we trained DALE-CT (Depth-Aware Latent-Euclidean Computed Tomography), a 2D model family built entirely from scratch using LeJEPA, and compared its performance against a continually pre-trained DINOv2 baseline. To enhance representation quality, we developed a novel 3D depth-aware pre-training strategy anchored by dense auxiliary supervision from both automated anatomical masks and human-annotated abnormalities. Under linear probe evaluation with Multiple Instance Learning (MIL) for multi-abnormality detection, the frozen backbone of this dual-supervised model (DALE-CT-2S) achieves a Macro AUROC of 0.833. This performance demonstrates near-parity with state-of-the-art 3D vision-language models, achieved entirely from scratch with significantly less data and no textual supervision. To ensure reproducibility, all training code, evaluation scripts, and model weights have been made publicly available.

Current Vision-Language Models struggle with hours-long videos because processing full-length visual sequences induces prohibitive token explosion and attention dilution. To overcome this, we introduce MemDreamer to decouple perception and reasoning, shifting long-video understanding into an agentic exploration process. As a plug-and-play framework, it incrementally streams videos to construct a Hierarchical Graph Memory, a top-down three-tier architecture for semantic abstraction, anchored by a foundational graph capturing spatiotemporal and causal relations. During inference, the reasoning model employs agentic tool-augmented retrieval, navigating hierarchies, searching nodes, and traversing logical edges via an Observation-Reason-Action loop. Experiments show MemDreamer achieves SOTA results across four mainstream benchmarks, narrowing the gap with human experts to only 3.7 points. It constrains the reasoning context window to merely 2% of full-context ingestion while delivering a 12.5 point absolute accuracy gain. Furthermore, statistical analysis uncovers a strong positive linear correlation between an VLM's performance on logic reasoning and long-video understanding benchmarks, establishing agentic capability scaling as a new paradigm for multimodal comprehension.

Vision-language driving models increasingly use reasoning supervision to bridge perception, prediction, and planning, but existing driving rationales are often free-form and expensive to generate with frontier models. We present VeriDrive, a framework for constructing planning-oriented, verifiable counterfactual supervision. VeriDrive converts driving reasoning into a structured Perception-Evaluation-Revision chain that grounds key objects in future motion, evaluates alternative ego trajectories with rule-checkable evidence, revises risky intent toward expert behavior, and produces final planning targets. To scale data construction, VeriDrive combines local generation with validator-guided selective correction, escalating only invalid or difficult samples. We build the VeriDrive dataset on nuScenes and train under the Omni-Q protocol. Controlled open-loop experiments show that VeriDrive improves L2, Collision, and Intersection over OmniDrive while reducing logged token usage, generation time, and actual paid LLM/VLM cost. These results show that auditable intermediate fields and structured revision targets can improve vision-language planning supervision under realistic annotation budgets. Code, prompts, and validator scripts are coming soon and will be released after the review process.

Geospatial understanding is a critical yet underexplored dimension in the development of machine learning systems for tasks such as image geolocation and spatial reasoning. In this work, we analyze the geospatial representations acquired by three model families: vision-only architectures (e.g., ViT), vision-language models (e.g., CLIP), and large-scale multimodal foundation models (e.g., LLaVA, Qwen, and Gemma). By evaluating across image clusters, including people, landmarks, and everyday objects, grouped based on the degree of localizability, we reveal systematic gaps in spatial accuracy and show that textual supervision enhances the learning of geospatial representations. Our findings suggest the role of language as an effective complementary modality for encoding spatial context and multimodal learning as a key direction for advancing geospatial AI.

Vision-Language Models (VLMs) have demonstrated strong performance across multimodal tasks, yet their safety robustness remains an open challenge. While prior work has shown that structured visual prompts such as flowcharts can effectively jailbreak VLMs, existing studies are largely limited to English-centric settings. In this paper, we introduce MLingualFC, a multilingual multimodal benchmark designed to evaluate jailbreak vulnerabilities of VLMs across diverse languages using structured flowchart representations. MLingualFC encodes harmful instructions into flowchart images across five languages (Hindi, Punjabi, Spanish, Romanian, and German). We evaluate state-of-the-art multilingual VLMs, including Qwen2.5-VL, Gemma-4, and Pangea, under a black-box threat model. Our results reveal significant multilingual safety gaps. Flowchart-based attacks achieve high attack success rates (ASR) in case of Latin script languages, demonstrating that visual encoding of harmful content effectively bypasses safety alignment across languages. In contrast, non-Latin script languages such as Punjabi exhibit substantially lower ASR, suggesting potential limitations in visual text recognition rather than stronger safety alignment. These findings highlight that current VLM safety mechanisms fail to generalize across languages and modalities. Resources are available at https://github.com/Rishabhpm23/MLingualFC

Continual Object Detection (COD) requires a detector to acquire new categories over time while preserving previously learned ones. This goal is closely related to open-vocabulary detection, since both settings require reasoning over categories that are not fully covered by the annotations available at the current training stage. Recent CLIP-based open-vocabulary detectors have shown strong zero-shot generalization, and frameworks such as F-ViT demonstrate that vision-language pretraining can provide powerful zero-shot detection ability for unseen categories. However, real-world deployments cannot remain purely zero-shot: once these detectors are continually updated on newly introduced categories, they suffer severe catastrophic forgetting and quickly lose their previously calibrated detection ability. We therefore propose CL-CLIP, a CLIP-based COD framework that equips open-vocabulary detectors with better continual learning ability through cost-volume-guided category decoupling. Specifically, following CAT-Seg, we compute a CLIP image-text similarity cost volume, defined as dense category-wise response maps between visual tokens and class text embeddings. This zero-shot spatial prior decomposes shared region features into class-specific pathways, which are then processed by a Multi-Expert RoI head. Extensive experiments on PASCAL VOC and MS-COCO show that CL-CLIP substantially improves the F-ViT baseline under continual fine-tuning and achieves competitive performance with existing continual object detectors, especially in adapting to newly introduced categories while preserving competitive base-class performance.

As Earth Observation data generation outpaces downlink bandwidth and human-in-the-loop processing, a widening gap has emerged between onboard collection and actionable ground intelligence. This paper presents NAVI-Orbital, a software system deployed on a Low Earth Orbit (LEO) spacecraft. On April 16, 2026, NAVI-Orbital achieved what is, to the authors' knowledge, the first in-orbit demonstration of a vision-language model performing autonomous multi-modal inference entirely onboard. NAVI-Orbital uses a local vision-language model (Gemma 3) to classify each captured scene, produce a text description of its content and the relationships between its features, and respond to operator follow-up via natural-language dialogue. The system is re-tasked through plain-English prompts in place of conventional command sequences, and is orchestrated by a graph-based state machine (LangGraph) coordinating dedicated agents for detection and dialogue. Results across ground benchmarking (88.16% accuracy on the 7,960-image curated AID benchmark), Flatsat validation, and live in-orbit captures of newly acquired, previously unseen Earth imagery (including uncorrected YAM-9 imagery, processed onboard with hardware-accelerated GPU inference and no fine-tuning for the flight instrument) demonstrate the feasibility of running foundation models on satellite-class edge computers to invert the conventional acquire-then-downlink-everything bandwidth profile through semantic compression of Earth observations in-orbit.

We present the Membership Inference Test (MINT) Demo 2, a framework designed to improve transparency in machine learning training processes. MINT is a technique for experimentally determining whether specific data were used during machine learning model training. We establish the theoretical framework and propose multiple architectures for MINT depending on the amount of information known about the models that are being audited. Experimental results using a popular face recognition model, 4 state-of-the-art LLMs, and multiple, diverse, and large-scale public image and text databases achieve promising accuracy levels in the detection of training data of up to 90%. Building on these results, we introduce a comprehensive web platform1 that expands these capabilities to image and text modalities. The platform integrates a diverse technological stack, including MINT, aMINT, and gMINT, allowing users to audit a wide range of models. This demonstrator aims to promote AI transparency and provides a practical tool to foster compliance with emerging AI regulations.

The rapid evolution of generative image models challenges existing AI-generated image detectors, particularly in open-world settings with unseen generators. Recent training-free approaches measure robustness gaps in frozen vision foundation models (VFMs), detecting fakes via perturbation-induced embedding drift. However, these methods rely on fixed invariance geometry inherited from pretraining and lack principled adaptation to the detection task. We instead formulate AI-generated image detection as learning a structured invariance manifold of real images under one-class supervision. Building upon a frozen VFM, we introduce lightweight projection heads that decompose representation space into complementary robust and fragile subspaces. The robust subspace is explicitly trained to suppress variations induced by physically plausible imaging transformations, approximating tangent directions of a real-image manifold, while the fragile subspace retains sensitivity to edit-like perturbations. A structured ordering margin enforces hierarchical separation between physical invariance and edit-induced variability, enabling detection as a margin-violation test relative to the learned manifold. At inference, multi-scale patch-wise drift under both transformation families yields a dual-channel invariance signature and interpretable localization. Extensive experiments demonstrate strong open-world generalization across unseen generators and resolutions, consistently outperforming training-free robustness-based baselines while providing interpretable invariance-violation maps.

Vision-Language Models (VLMs) frequently rely on language priors, producing confident answers that are weakly grounded in visual evidence. While this behavior is widely observed, its internal mechanisms and its impact on benchmark evaluation remain insufficiently understood. In this work, we study language-prior reliance from both mechanistic and behavioral perspectives. Internally, we combine counterfactual layer replacement with supervised layer-wise MLP probing to trace how ground-truth visual semantics and language-prior semantics compete across the language decoder. Our analysis reveals a multi-stage bottleneck: intermediate layers often fail to effectively retrieve visual information, while later layers can further suppress surviving visual signals in favor of text-space biases. Externally, we introduce a progressive visual decay metric based on multi-step Gaussian blurring, which identifies instances whose answers remain invariant even as visual content is increasingly destroyed. Across twelve visual question-answering benchmarks and three representative VLMs, we find that a substantial fraction of examples remain answerable under severe or total visual obfuscation, indicating that current benchmarks can inadvertently reward visual ignorance. These findings demonstrate that language-prior reliance is a systematic routing failure affecting both model internals and benchmark validity. Finally, we outline critical pathways for future research, highlighting the necessity of designing training distributions and evaluation protocols built on structurally isolated or counterfactual data to enforce genuine cross-modal grounding.

Language-guided UAV agents must execute long-horizon semantic instructions while producing smooth, physically feasible continuous flight commands, yet existing Vision-Language Navigation (VLN) benchmarks typically use discrete or coarse actions and existing UAV Vision-Language-Action (VLA) tasks focus on short, atomic maneuvers. To address this gap in UAV task settings, we introduce \textbf{FLIGHT}, a \textbf{F}ine-grained \textbf{L}ong-horizon \textbf{I}nstruction-\textbf{G}uided benchmark for \textbf{H}ybrid UAV navigation and reasoning \textbf{T}asks, which combines multi-stage instructions with dense 6-DoF trajectory annotations across two dataset splits: Fine-grained VLN and Long-horizon Flow. To endow the UAV agent with the capability of real-time in-flight reasoning over task execution status and mission planning, while simultaneously accommodating high-frequency, real-time precise control, we further propose \textbf{FLIGHT VLA}, an asynchronous architecture that decouples a low-frequency Streaming Pilot Vision-Language Model (VLM) for task-state reasoning from a high-frequency diffusion action model for continuous control, supervised by explicit \textbf{Pilot Reasoning} texts that summarize the current flight state and anticipate the next subgoal. In closed-loop evaluation, FLIGHT VLA consistently surpasses representative VLN and VLA baselines on our FLIGHT benchmarks, achieving stronger multi-stage completion, subgoal adherence, and terminal control. Its trained Streaming Pilot Reasoning VLM further improves UAV video reasoning, validating the effectiveness of our design.

Recent advances in 3D multimodal large language models (3D-MLLMs) have enabled unified solutions for 3D scene understanding tasks, including visual question answering, captioning, and referring segmentation. However, existing 3D-MLLMs remain largely object-centric, limiting their ability to model fine-grained part structures that are essential for embodied interaction with 3D environments. In this work, we present PAR3D, a unified part-aware 3D-MLLM framework that enables models to understand, reason about, and ground both objects and their parts in 3D scenes. To enable training and evaluation of part-aware 3D scene understanding, we introduce ScenePart, a synthetic 3D scene dataset with part-level annotations and language instructions. We further develop Part-Aware 3D Representation Learning to enrich 3D visual representations with fine-grained part-level semantics, and propose Hierarchical Segmentation Query Generation to ground part targets via hierarchical object-part queries. Extensive experiments show that our method substantially improves part-level question answering and referring segmentation, while also achieving strong performance across object-level vision-language tasks.

Medical imaging artificial intelligence has achieved strong performance in isolated image interpretation, but remains poorly aligned with radiological practice, where diagnosis and follow-up rely on comparison across prior studies and analogous reference cases. Here we formulate radiological comparison as an entity-aware cross-image reasoning problem and introduce a framework that supports both reference-case retrieval and temporal comparative interpretation. We construct MedReCo-DB, a large-scale comparative imaging resource derived from routine image-report pairs, comprising more than 690,000 images from over 160,000 patients across eight institutions, four countries and seven imaging modalities. Reports are decomposed into anatomical structures, abnormal findings and pathological conditions to provide supervision for entity-conditioned retrieval and comparative visual question answering. Using this resource, we develop MedReCo, an entity-aware visual encoder for controllable retrieval of clinically analogous cases, and MedReCo-VLM, a vision--language extension for generative interpretation of interval change. Across internal, external and cross-center evaluations, MedReCo achieved the highest Recall@1 in all 12 internal retrieval settings and improved external retrieval by a mean of 6.0 percentage points. In clinically confusable differential groups, it consistently outperformed the strongest baselines. MedReCo-VLM achieved the best performance across all comparative generation evaluations and improved longitudinal follow-up accuracy by 14.5-46.5 percentage points on chest radiographs and 13.0-27.9 percentage points on CT. These findings suggest that entity-aware comparative reasoning can be learned from routine clinical data at scale and may provide a more clinically aligned foundation for medical imaging AI.

While vision-language models excel at general multimodal understanding, they still struggle with visual spatial planning. We attribute this to a perception-reasoning modality gap: visual planning requires models to infer latent state structures from pixels and then reason over the recovered structure to produce valid actions, whereas symbolic planning directly leverages explicit objects and constraints. This creates dual bottlenecks in visual state recovery and multi-step planning. To address this, we propose MGSD, a two-stage modality-gap-aware self-distillation framework. First, a cold-start grounding stage equips the visual student with reliable state representations, minimizing early perception noise. Second, a privileged teacher transfers planning capabilities via on-policy distillation, using explicit symbolic states to supervise the student's own visual rollout prefixes. Crucially, symbolic data is used strictly during training, leaving inference purely visual. Experiments on visual planning benchmarks show that MGSD consistently improves visual planning across both 4B and 8B backbones, raising the macro average by 19.3% and 18.4%, respectively. The resulting models narrow the gap to symbolic-input upper bounds, while ablations and diagnostics confirm that the improvement comes from both visual state recovery and optimal-path reasoning. These results suggest that modality-gap-aware self-distillation improves not only how models perceive actionable states, but also how they plan over the inferred structure. Code is available at https://github.com/Oranger-l/MGSD.

Large Vision-Language Models have achieved significant reasoning performance in various tasks. However, there are few studies on text-to-3D indoor scene generation with LVLMs. The main challenge is that prevailing LVLM-based methods employ chain-of-thought sequential decision mechanisms that cannot revise earlier decisions, causing error propagation. In this paper, we consider the task as a planning problem constrained by spatial and layout commonsense. To solve this problem, we model it as a tree search problem with global and local trees, which differs from existing sequential decision-making approaches. In the global tree, we place each object iteratively and explore multiple attempts like humans furnishing a room, where the problem space is represented as a tree. To effectively search the tree, we propose a hierarchical scene representation and a PRM-guided MCTS method. This representation abstracts a scene into room level, region level, floor object level, and supported object level. The PRM-guided MCTS method uses the PRM to prune unnecessary branches and the MCTS algorithm to balance exploration and exploitation to get an optimal solution with fewer attempts. In the local tree, it further decomposes the placement of each object into finer sub-steps, including the specific placement parameters. To make the whole appearance of the scene consistent, we leverage pre-trained diffusion image generative models to predict textures for all the objects in the scene. As existing benchmarks for text-to-3D indoor scene generation remain limited in scale and diversity, we collect a new large-scale diverse dataset that contains 65 scene types and 3250 instructions with diverse sizes, layouts, and styles, named 3DTindo-bench, to better assess the capability of the state-of-the-art models. Our experiments show that our method generates more realistic 3D scenes than state-of-the-art methods.

Understanding and reasoning about the physical world is the foundation of intelligent behavior, yet state-of-the-art vision-language models (VLMs) still fail at causal physical reasoning, often producing plausible but incorrect answers. To address this gap, we introduce CausalPhys, a benchmark of over 3,000 carefully curated video- and image-based questions spanning four domains: Perception, Anticipation, Intervention, and Goal Orientation. Each question is paired with an expert-annotated causal graph capturing object-attribute-event dependencies, enabling interpretable and fine-grained evaluation of causal understanding. Building on this, we formulate a causal-graph-grounded metric that quantitatively measures how well a model's chain-of-thought reasoning aligns with the correct causal relations, moving beyond answer-only accuracy and enabling systematic diagnosis of VLMs' causal reasoning failures. Using this metric, we conduct a comprehensive analysis of leading VLMs, revealing systematic gaps in capturing causal dependencies and underscoring the need for causality-aware learning. To address these limitations, we further propose Causal Rationale-informed Fine-Tuning (CRFT), which explicitly aligns VLM reasoning with causal structures. Extensive experiments demonstrate that CRFT substantially enhances both reasoning accuracy and interpretability across multiple model backbones. By unifying dataset curation, causal evaluation, and causality-informed learning, CausalPhys establishes a strong foundation for advancing modern VLMs toward causally grounded physical reasoning.

Deploying Video Anomaly Detection (VAD) in real-world surveillance faces a fundamental tension between the demand for high-level semantics to ensure effectiveness and the limited computational resources of edge devices. Vision-Language Models (VLMs) provide rich open-vocabulary semantics, but their latency and computational cost preclude on-device deployment. To address the challenge, we propose MemoVAD, an edge-cloud collaborative framework that selectively incorporates VLM semantics into streaming VAD. MemoVAD runs most inference on the edge with a lightweight detector and a causal Temporal Context Encoder (TCE) to model temporal dependencies. Specifically, we introduce an Uncertainty-Aware Gating (UAG) policy grounded in Subjective Logic to model perceived uncertainty and query the cloud-based VLM only for high-uncertainty and semantically novel clips. Besides, a Dynamic Semantic Memory (DSM) is designed to cache VLM-verified prototypes for efficient retrieval, enabling the edge model to progressively incorporate VLM-level semantics via a semantic adapter. Experiments on UCF-Crime and XD-Violence datasets via a real edge device show that MemoVAD substantially reduces communication overhead while surpassing state-of-the-art performance.

Latent visual reasoning (LVR) inserts supervised latent tokens between perception and answer generation in vision-language models (VLMs). The field uses alignment between these latents and their visual targets, i.e., cosine similarity or mean squared error (MSE), as both the training loss and the quality metric, assuming that better alignment yields a better answer. We test this with a designed matrix of five LVR variants and find the assumption inverted: cosine alignment is negatively correlated with accuracy across all five (r=-0.94). To explain this, we introduce PRISM, a pair of inference-time diagnostics: a linear probe that asks where the answer is decodable, and a corruption test that asks whether the latent is load-bearing. The supervised latents are largely bypassed. Corrupting them shifts accuracy by at most four points. The answer is decodable downstream of the latent but not at it, and the size of this decodability gap predicts how much each variant relies on its latent under perturbation. Consistent with an Information Bottleneck reading of the loss, the auxiliary objective reshapes the language model via shared parameters rather than via the latent variable it nominally optimizes.

Recent advancements in Vision-Language Models (VLMs) have significantly enhanced their ability to interpret complex visual semantics, yet their capacity for chronological reasoning remains under-explored. In this paper, we introduce a novel benchmark specifically designed to evaluate how VLMs perceive and reason about chronological information within and across images. Unlike existing video-based benchmarks that focus on frame sequencing, our work delves into the underlying logic of chronological judgment and the expansion toward multimodal integration. To facilitate this, we construct three specialized datasets: one containing visually similar objects spanning long historical durations, another categorized by diverse event and object types, and a third pairing images with time-sensitive news text for cross-modal alignment. Through extensive experiments, we analyze whether models exhibit performance disparities across categories and, crucially, explore whether they rely on ``incorrect shortcuts'', such as image color rather than genuine chronological features. Our results reveal that while VLMs show promise, they frequently exploit superficial cues like grayscale versus color filters to bypass authentic chronological reasoning. By providing these high-quality datasets and a rigorous evaluation framework, we offer a diagnostic tool to identify current limitations and guide the development of more robust, logically grounded multimodal models. The source code is shown in https://github.com/LuoRenqiang/ChronoVision.

We present T-SAR-JEPA, a self-supervised framework for temporal anomaly detection in SAR amplitude stacks via latent prediction. A ViT-Base/16 encoder from SAR-JEPA is domain-adapted on 39,300 Capella patches using local masked reconstruction with gradient feature prediction. A temporal transformer with sinusoidal time encoding forecasts future latent states from K=7 acquisitions, with progressive unfreezing substantially reducing validation loss. The model operates on amplitude alone; InSAR coherence serves exclusively as independent pseudo-ground-truth. On the DFC 2026 dataset (300 time-series, three AOIs), T-SAR-JEPA achieves ROC-AUC of 77.0% on the Hawaii eruption window, outperforming RX, PaDiM, Linear AR, and LSTM baselines (~50%). Spatial coherence of 99.9% (p < 0.001, permutation test) confirms structured detections. Code: https://github.com/TerraLatent/t-sar-jepa

Given the inherently multimodal nature of human experience, vision-language models (VLMs) hold substantial promise for modeling human cognition as it grows and develops with experience. Realizing their potential requires tools for comparing VLMs with human cognitive development across tasks, ages, and populations. We present LEVANTE-bench, a benchmark based on tasks and data from the Learning Variability Network (LEVANTE), which distributes open-source tasks and data measuring children's cognition across languages and cultures. In LEVANTE-bench, we systematically assess VLMs on six tasks, comparing their alignment with children aged 5-12 ($N$ = 1547) across three countries. We compare models at multiple scales, assessing their overall accuracy, their task- and item-level alignment with children, and how well they match children's trial-level error distributions. Alignment was heterogeneous across scales: at the level of tasks and items, more capable models aligned better with humans. However, match to human error distributions varied widely across tasks, and for several tasks, smaller models matched younger children's errors better. In addition, even the best-performing VLMs struggled on matrix reasoning and mental rotation tasks. Thus, current VLM architectures align only partially with the cognitive abilities of children.

Recent studies have explored Vision-Language Models (VLMs) for food analysis. However, most existing methods rely primarily on supervised fine-tuning (SFT), which often limits reasoning and generalization capabilities. Moreover, high-quality large-scale nutritional annotations remain scarce. To address these issues, we introduce CalorieBench-80K, a large-scale benchmark with curated calorie labels and dietary advice annotations. To the best of our knowledge, it is the first food image benchmark to incorporate Chain-of-Thought (CoT) annotations for calorie reasoning. We also propose Food-R1, a unified food VLM trained in a multi-task learning paradigm to equip the model with broad capabilities. Food-R1 undergoes CoT-based cold-start instruction tuning, followed by reinforcement fine-tuning (RFT) using Group Relative Policy Optimization (GRPO) to improve reasoning and performance. Experiments on CalorieBench-80K and representative benchmarks show that Food-R1 consistently outperforms strong baselines across food-related tasks. The code, model weights, and benchmark annotations are available at the project repository.

Current prompt-based and adapter-based tuning of vision-language models (VLMs) is attractive for medical imaging, where clinical data sensitivity favors frozen backbones and annotations are limited. However, these methods typically optimize only the ground-truth class, treating all other classes as equally incorrect, ignoring clinically meaningful class relations and yielding unstable decision boundaries in limited-supervision settings. We propose Omni-Geometry Knowledge Distillation (OGKD), a new framework that injects class-relation structure into the teacher to produce directional targets that preserve the ground truth while respecting inter-class geometry. Using these targets, we develop two distillation losses: Global Geometry-Aware Distillation (GAD) operates on the global image token, and Label-Guided Geometry Distillation (LGD) applies the same geometry to attentive patch tokens to improve fine-grained alignment. Across comprehensive experiments and analyses on 11 widely-used medical datasets for base-to-novel and few-shot evaluations, our OGKD achieves substantially better performance, consistently improving accuracy by an average absolute gain of 1.7%-2.8% over all prior state-of-the-art VLM adaptation counterparts. It also robustly generalizes to unseen classes and yields more reliable predictions than other approaches. Our code is available at https://github.com/tientrandinh/OGKD.

Vision-language models (VLMs) are increasingly used in multi-image, multi-turn agentic settings where decisions depend on visual changes. However, in existing open-weight VLMs, visual comparisons happen only inside the language model, while the visual encoder itself remains stateless: each image is encoded independently, without access to the prior visual context. As a result, small but task-critical changes may be attenuated before the language model has a chance to compare them, especially when those changes do not affect the high-level semantics of the scene. We introduce a Stateful Visual Encoder, which conditions each visual representation on prior visual features. Under supervised finetuning, VLMs equipped with stateful encoders achieve consistent improvements on controlled tasks involving cross-image spatial aggregation, multi-object visual differencing, and visual trajectory behavior cloning. These improvements are consistent across input resolutions, language model sizes, and VLM backbones. Finally, we validate our model on real-world tasks, including longitudinal radiology, fine-grained image comparison, and remote sensing, where stateful encoders consistently improve generalist VLM baselines and can match or surpass specialized models in selected domains. Project page: https://statefulvisualencoders.github.io/

Foundation models have driven rapid progress in computer vision, yet the two dominant paradigms, vision-language foundation models (VLMs) and vision-only foundation models (VFMs), remain only partially compatible. VLMs offer language-grounded semantic alignment but are often visually coarse, while VFMs learn discriminative perceptual geometry but lack semantic grounding. We propose GPUA (Geometry-Preserving Unsupervised Alignment), a framework that integrates the complementary strengths of VFMs and VLMs. Inspired by cross-lingual alignment, GPUA treats VFM features as a visual language and learns an orthogonal mapping that translates the VFM space into the VLM semantic space, preserving geometry and narrowing the modality gap without labels or model parameter updates. GPUA is task-agnostic and requires only feature-level access to pretrained models. Experiments across diverse benchmarks demonstrate improved cross-model compatibility and strong gains in downstream zero-shot recognition and segmentation with negligible overhead. Code is available at https://github.com/Yuteam14/GPUA

Multimodal large language models (MLLMs) are predominantly evaluated on free-form vision-language tasks such as visual question answering, captioning, and summarization. However, their practical use is rapidly expanding to more structured computer vision settings, where users prompt models to perform localization-centric tasks such as object detection, often within larger agentic or decision-making systems. Despite this shift, there is currently no standardized benchmark that systematically evaluates these capabilities at scale. In this work, we introduce the first comprehensive benchmark specifically designed to assess the promptable localization abilities of generalist MLLMs. Our benchmark spans four core task categories: object detection, referring expression detection, instance-level detection, and video-based detection. To enable consistent and fair evaluation, we develop a unified framework that standardizes inputs, enforces parsable bounding box outputs, and defines transparent evaluation protocols across tasks. Using this suite, we evaluate a diverse set of open-source and proprietary MLLMs, providing an in-depth analysis of their performance and limitations. Beyond accuracy, we examine models' ability to adhere to output format specifications, showing that current systems are highly sensitive to formatting constraints and often fail to generalize even to minor variations. Our results highlight both the strengths and shortcomings of state-of-the-art MLLMs in localization settings, and point toward important directions for improving multimodal model design and evaluation.

Recent years have seen remarkable progress in unified vision-language models handling both multimodal understanding and generation within a single architecture. While autoregressive VLMs can reason across modalities, they fail to generate high-quality images. In contrast, diffusion models produce photorealistic visuals yet struggle to generate coherent text, making it challenging to develop a single unified model that can seamlessly handle both visual and text generation. Recent advances suggest that language can be effectively embedded within visual representations, allowing models to reason about textual semantics directly from images. To this end, we propose UniCanvas, a first attempt that unifies diffusion models to generate interleaved multimodal contents through text-in-image generation. Diffusion models naturally capture transformations on a shared pixel canvas, which can be viewed as world models of visual change. Instead of producing discrete text tokens, the model learns to represent language as visual patterns inside images, leveraging its inherent multimodal embedding space. This design allows the model to "draw" text naturally within a single pixel canvas during image synthesis, achieving seamless multimodal generation. Experiments demonstrate that UniCanvas improves performance over previous unified models, positioning text-in-image generation with diffusion models as a promising unified multimodal generation paradigm.

Curating training data is among the most consequential yet labor-intensive parts of modern AI development: practitioners iteratively propose, implement, evaluate, and revise data policies against noisy benchmark feedback. We ask whether generalist coding agents can automate this data-curation loop. We introduce Curation-Bench, an agent-centric benchmark that fixes the model, training recipe, and evaluation suite while giving agents command-line access to inspect data, implement policies, submit them to a fixed training/evaluation pipeline, and revise. In a vision-language instruction-tuning instantiation, out-of-the-box agents reach strong published data-selection baselines within ten iterations. However, trajectory analysis reveals a persistent execution-research gap: agents mainly tune local policy variants rather than explore new policy families, even when given strategy guides and paper references. Scaffolds requiring each iteration to cite, instantiate, and adapt a prior method shift agents toward method-guided exploration. The scaffolded agent autonomously composes -- without human design input -- a data-selection policy that outperforms strong published baselines at one-tenth their data budget. Overall, current agents can run the curation loop, but reliable data research requires scaffolded method adaptation, not open-ended prompting alone. Code and benchmark are open-sourced.

Vision-language foundation models such as CLIP and SigLIP provide widely used representations for multimodal learning systems. While these models are typically compared through downstream performance, such evaluations often do not explain how their representations differ structurally. In this work, we study this problem through the task of Contrastive Embedding Clustering: identifying sample subsets that are weakly clustered under one representation but strongly clustered under another. We propose \emph{Kernel Optimization for Discrepancy Analysis (KODA)}, a kernel-based framework for contrastive representation comparison and alignment. KODA constructs unified multimodal kernels through modality-wise kernel composition and formulates discrepancy discovery as a constrained optimization problem that searches for coherent structures in one representation while suppressing coherence in a reference representation. This yields interpretable discrepancy directions associated with specific sample subsets and modality interactions. To scale KODA to large vision-language datasets, we develop randomized low-dimensional approximations of joint kernels using random projections, including Random Fourier Features for shift-invariant kernels. Empirically, KODA identifies consistent and interpretable discrepancy structures across vision-language representations and provides sample subsets for representation alignment. The code is available at https://github.com/yokiwuuu/KODA.

As foundation models scale toward fusing more heterogeneous visual streams, understanding how diverse encoders interact under joint training becomes a prerequisite for principled design. Yet large vision-language models (LVLMs) currently lack the tools to do so, and parameter-efficient encoder configurations remain hard to identify before training. To re-examine encoder roles under joint training, on the 16-benchmark Cambrian-1 suite we retrain and evaluate all 31 non-empty subsets of five common vision encoders under a unified pipeline (~20k GPU-hours total), and report three findings. First, retraining each subset from scratch reveals encoder rankings that differ from those obtained by masking encoders on a fixed checkpoint, including which encoder ranks first overall. Second, we decompose each encoder's contribution into two axes, Capacity, the score an encoder reaches on its own, and Necessity, the drop when it is removed from the full pool. The two axes are not interchangeable. Pairing the two highest-Capacity encoders is suboptimal, while pairing a high-Capacity anchor with an adaptive complement matches the full five-encoder model. Adding further encoders beyond this pair yields only marginal gains. Third, at fixed parameter count, per-encoder pre-projector effective rank explains the residual score variation. The strongest pairs combine an anchor whose rank survives joint training with a complement whose rank expands under it, suggesting that higher-rank, less-collapsed projector inputs correspond to a more favorable optimization regime at the encoder-projector interface. Together, the Capacity-Necessity decomposition and the pre-projector rank analysis, along with comprehensive evaluation through retraining, expose a methodological gap in multi-encoder LVLM design, and offer concrete primitives for closing it.

Visual instruction tuning effectively adapts a pre-trained Large Language Model (LLM) to process image information alongside text. Yet, it remains unclear how visual features are embedded into the layer-wise hierarchy of abstractions of the LLM backbone. Across a diverse set of vision-language architectures, we show that instruction tuning primarily serves as a bridge, embedding visual features directly into the intermediate semantic layers of the LLM, bypassing the early layers devoted to unimodal processing. With probing analyses and causal interventions, we show that these intermediate layers are the semantic core of vision-language processing and play a critical role in the performance on a broad set of multimodal benchmarks. In addition, by comparing the geometry of semantically equivalent visual and textual representations, we find that fine-tuning extends and strengthens the existing abstraction phase, aligning visual features with pre-existing textual ones. Finally, we confirm the functional role of this localized alignment by restricting fine-tuning to intermediate layers alone: this strategy preserves the performance of full fine-tuning on vision-centric benchmarks while reducing training time. Our results suggest that multimodal integration is a localized phenomenon driven by the repurposing of the internal abstraction engine of the LLM.

Vision-language models map visual features into a shared embedding space through learned projection layers, yet it remains unclear how these transformations alter the structure of visual information. This study examines changes in representation through spatial-frequency accessibility, measured by the linear recoverability of band-limited Fourier energy from model representations. To isolate effects beyond dimensionality reduction, we introduce Residual Spectral Loss (RSL), which evaluates changes relative to a dimension-matched random projection baseline. To reduce confounding effects from optimization, the analysis uses pretrained models with all parameters frozen. The experimental results show consistent frequency-dependent changes in accessibility across CLIP and DINOv2 on ImageNet and MS-COCO datasets. Spectral accessibility follows a non-monotonic trajectory across depth, peaking at intermediate layers before decreasing toward the output representation. The final transformation differs across architectures: CLIP's learned projection is spectrally neutral, with changes explained by compression, whereas DINOv2's [CLS] pooling induces a structured loss across the spectrum. These findings identify intermediate layers and pooling mechanisms as primary drivers of spectral transformation in modern vision encoders.

Vision-language models (VLMs) have been explored for visual programming, where they generate code to solve visual tasks. However, most prior work focuses on visual programming for productivity; it remains unclear how well current VLMs perform on education-oriented visual programming and what factors limit their performance. To bridge this gap, we introduce TurtleAI, a benchmark containing 823 tasks curated based on real-world visual programming tasks in the Turtle Graphics domain. Solving these tasks requires models to perceive geometric patterns, reason about spatial relationships, and synthesize Python code that faithfully reproduces geometric patterns. We evaluate 20+ VLMs, including GPT-5, GPT-4o, and Qwen2-VL-72B, and find that they struggle significantly, with most achieving success rates below 30%. To address these limitations, we propose a data generation technique that requires only a small set of seed samples. Fine-tuning Qwen2-VL-72B on the resulting synthetic data yields an improvement of about 20% on real-world tasks. Our failure analysis reveals that GPT-4o struggles with spatial reasoning and precise visual replication, whereas fine-tuning primarily improves the alignment between visual reasoning and code implementation.

When asked what a meme or sarcastic post means, Large Vision Language Models (LVLMs) tend to describe what the image shows rather than what the author is trying to communicate. Standard instruction tuning entangles a post's literal content with its pragmatic meaning, letting surface-level details contaminate the final response. We reframe meme understanding as a problem of literal-pragmatic decomposition and propose \textbf{Intent Projection}, a framework that separates the two signals at the representation, output, and objective levels within a single LVLM backbone. At the representation level, an orthogonal projection module removes dominant unimodal directions from the fused image-text representation, retaining only the pragmatic residual, while a surface-real affect classifier anchors the decoder with a discrete tag that names the polarity gap. At the output level, the model externalizes a structured reasoning chain, and at the objective level a contrastive reward explicitly penalizes answers that restate the literal description. Across six multimodal benchmarks, Intent Projection consistently outperforms open-source baselines and narrows the gap to proprietary models, with the largest gains on high-divergence posts where literal collapse is most damaging.

Engineering diagrams pose a distinct challenge for vision-language models: unlike natural images or general documents, they encode information through dense spatial layouts, domain-specific symbols, and cross-references between visual callouts and structured parts tables. Despite their centrality to service, repair, and design workflows, there is no public benchmark for measuring VLM capabilities in this domain; existing datasets primarily focus on flowcharts, scientific figures, or business documents. To address this gap, we introduce Enginuity, the first open dataset and benchmark for evaluating VLMs on complex engineering diagrams. We define two tasks over a corpus of U.S. military service and repair manuals: structured parts-table extraction (Task 1) and free-form visual diagram question answering (VQA)(Task 2) for benchmarking. We evaluate four frontier VLMs (GPT-5.2 Chat, Claude Opus 4.7, Gemma 4, Qwen3-VL-32B-Instruct) under zero-shot and chain-of-thought prompting. On Task 1, models reach Recall@all of 0.61-0.87 but Token F1pen of only 0.03-0.18, exposing a systematic gap between part identification and description fidelity. Task 2 reveals a consistent factual-reasoning gap across all models. A supporting analysis shows that token-overlap metrics under-report model capability on technical descriptions by 2-6x relative to semantic similarity, motivating LLM-as-judge calibration for domain-specific evaluation. We release the dataset, annotations, evaluation harness, and per-sample model outputs to support a reproducible study of VLM capability on engineering content.

Hallucination has recently garnered significant research attention in Large Vision-Language Models (LVLMs). Direct Preference Optimization (DPO) aims to learn directly from the corrected preferences provided by humans, thereby addressing the hallucination issue. Despite its success, this paradigm has yet to specifically target the perceptual bottleneck in attended regions or address insufficient Visual Robustness against image degradation. Furthermore, existing preference pairs are often vision-agnostic and their inherently off-policy nature limits their effectiveness in guiding model learning. To address these challenges, we propose Perceptual Processing Direct Preference Optimization (P$^2$-DPO), a novel training paradigm in which the model generates and learns from its own preference pairs, thereby directly addressing the identified visual bottlenecks while inherently avoiding the issues of vision-agnostic and off-policy data. It introduces: (1) an on-policy preference pairs construction method targeting Focus-and-Enhance perception and Visual Robustness, and (2) a well-designed Calibration Loss to precisely align visual signals with the causal generation of text. Experimental results demonstrate that with a comparable amount of training data and cost, P$^2$-DPO outperforms strong baselines that rely on costly human feedback on benchmarks. Furthermore, evaluations on Attention Region Fidelity (ARF) and image degradation scenarios validate the effectiveness of P$^2$-DPO in addressing perceptual bottleneck in attended regions and improving Visual Robustness against degraded inputs.

We present P-Topics (Perception Topics) modeling, a novel problem for understanding how images are perceived affectively and across cultures. The goal is to (1) discover and model the different perception experiences in a dataset of images and captions, where each experience is defined by an objective factual and a subjective affective aspect, and (2) associate images to their relevant perception experiences. We introduce PercepT (Perception topic Transformer), a two-stage architecture that tackles P-Topics modeling. In the formation stage, percepT discovers P-Topics as visual-textual clusters using an unsupervised training objective, and dynamically selects the number of clusters to match the perceptual richness of the dataset. In the mapping stage, it learns P-Topic mapping functions via attention pooling to associate images to their respective clusters. On ArtELingo, PercepT achieves a silhouette score of 0.97 compared to 0.37 from the closest baseline reflecting better perceptual clusters. PercepT also achieves an AUC score of 0.94 compared to 0.77 showing better mapping to perceptual clusters. Human evaluation confirms that PercepT captures semantically meaningful perception experiences and significantly outperforms existing methods. Our implementation will be made public.

Large vision language models (LVLMs) have made rapid advancements and are deployed across various applications, yet hallucinations remain a major challenge. Activation steering is appealing due to its minimal training overhead and controllability at inference time. However, we found that during autoregressive decoding, visual conditioning affects token prediction sparsely and locally across decoding steps, and many existing methods that average image-versus-no-image differences over the entire sequence dilute these critical signals, yielding low signal-to-noise ratio steering directions. Additionally, many existing methods apply a fixed steering strength, which misallocates the intervention budget, over-perturbs non-critical tokens, and can cause instability. To address these limitations, we propose Token-Level Visual-Sensitivity Steering (TLVS) for hallucination mitigation. Our approach first extracts token-level steering vectors and refines them, and then applies fine-grained, visual-sensitivity-adaptive steering only where it matters. This lightweight, plug-and-play mechanism requires only minimal training for calibration and can be applied across diverse vision-language models. It modulates the steering strength at each decoding step, selectively suppressing hallucination-prone spans while preserving evidence-grounded content. We evaluate TLVS on several benchmarks, including POPE, AMBER, CHAIR (COCO), MMHal, and HallusionBench, demonstrating consistent improvements over previous steering methods.

In embodied vision-language decision making tasks such as robotic manipulation and navigation, Vision-Language and Vision-Language-Action Models (VLMs & VLAs) are powerful tools with different benefits: VLMs are better at long-term planning, while VLAs are better at reactive control. However, their performance is limited by the same perceptual bottleneck: visual hallucinations arise due to the models' inability to distinguish task-relevant objects from distractors. In principle, accurate identification and focus on critical objects while filtering out irrelevant ones is the key to break this limitation. A straightforward solution is one-step focus: directly attending to essential objects. However, this approach proves ineffective because effective focus inherently requires deep scene understanding. To this end, we propose SceneDiver, a coarse-to-fine focus plan generation method for VLMs leveraging their long-term planning abilities, that first constructs a holistic scene graph to establish initial comprehension, then progressively decomposes the task into simpler sub-problems through an iterative cycle of recognition, understanding, and analysis. To enable reactive control, we also design a lightweight adapter for distilling the deliberate focus ability into VLAs. Evaluations on standard embodied AI benchmarks confirm that our method substantially reduces visual hallucinations for both VLMs and VLAs, while preserving computational efficiency in tasks requiring fast execution. Our code and data are released at: https://future-item.github.io/SceneDiver.

Group Relative Policy Optimization (GRPO) and its variants, originally developed for Large Language Models (LLMs), have recently been applied to Multimodal LLMs and produced strong results. However, their coarse-grained holistic credit assignment from a single scalar advantage underfits vision-language (VL) tasks, where outputs are often long-form responses grounded in semantically rich images. To address this limitation, we exploit a structured signal that single-scalar formulations discard: the natural segmentation of long-form VL outputs. Concretely, we propose Segment-Decomposed GRPO (SD-GRPO), which z-normalizes verifiable per-segment rewards across the rollout group, yielding a vector of per-segment advantages in place of a single scalar. We evaluate SD-GRPO across three settings spanning controlled and real-world long-form VL generation, organized by increasing semantic entanglement across segments. On a controlled multi-panel dense-captioning task constructed from DOCCI, where segments are semantically independent, SD-GRPO consistently outperforms the GRPO baseline, with larger gains at higher segment counts. Extending to a controlled multi-chart long-form VQA task constructed from MultiChartQA, we show both theoretically and empirically that rollout-level rewards suffer from cross-segment credit misattribution that scales with output length. On a real-world scientific figure captioning task on the MMSci dataset, where subfigure captions share context across the figure, blending holistic and per-segment rewards further improves on both, suggesting per-segment normalization alone is insufficient when segments are semantically entangled. Finally, by integrating SD-GRPO into Dr. GRPO, we confirm that it can be applied to any GRPO framework with minimal implementation overhead to enhance long-form VL generation.

We study preference alignment for image inpainting. Rather than proposing yet another method, we revisit the problem from first principles and reassess its core challenges. We adopt the widely used direct preference optimization framework and construct preference training data with publicly available reward models. Our empirical study spans nine reward models, two benchmarks, and two baseline inpainting models that differ in architecture and generative mechanism. Our main findings are: (1) Most reward models provide valid signals for preference data construction, although some are unreliable as evaluators. (2) Across models and benchmarks, preference data exhibits consistent trends under both candidate and sample scaling. (3) Reward models display pronounced biases--particularly in brightness, composition, and color scheme--that make them prone to inducing reward hacking. (4) A simple ensemble of reward models mitigates such biases and yields robust, generalizable performance. {\color{rebuttal_blue}(5) Preference alignment is transferable to the object removal task, where the goal shifts from open-ended creative generation to coherent background completion. (6) Further analysis reveals that a calibrated ensemble method further mitigates hacking and improves robustness.} Without modifying model architectures or introducing additional datasets, our models substantially outperform prior state-of-the-art models on standard metrics, large vision-language model evaluations, and human assessments. Our code is available at: https://github.com/shenytzzz/Follow-Your-Preference.

Vision-language models (VLMs) for radiology have emerged as a scalable paradigm by leveraging image-report pairs naturally produced in clinical workflows. However, this pairing reveals a mismatch in scale: each finding occupies only a small region of the image, yet supervision is provided only at the global image-report level. This poses a central challenge: prior approaches spread weight densely across all patches rather than concentrating on the sparse subset relevant to a given query. To address this, we present GLINT (Gated Language-Image alignmeNT), a framework that explicitly models this sparse correspondence. On the alignment side, we introduce Sparsely Gated Alignment, a novel architecture in which a sigmoid gate over a separate gate embedding space activates only the patches relevant to each textual query, enforcing explicit sparsity. On the representation side, we add Dense Feature Regularization, which anchors the trainable encoder's intermediate features to a frozen self-supervised learning (SSL) teacher, preserving the fine-grained patch features that the gate relies on. The same recipe applies to both 2D chest X-ray (CXR) and 3D chest computed tomography (CT), built with DINOv3 and V-JEPA 2.1, respectively. GLINT enables zero-shot classification, grounding, and segmentation from free-text queries, and to our knowledge is the first to demonstrate zero-shot segmentation on 3D CT volumes without mask supervision. Notably, the most pronounced gains arise on zero-shot grounding and segmentation, where sparse, query-specific localization is required, consistent with our design intent. In downstream evaluation, GLINT outperforms both SSL encoders and medical VLMs on classification, report generation, and segmentation.

Large Vision-Language Models (LVLMs) show strong visualization interpretation, yet it is unclear whether their responses reflect genuine reasoning over visual evidence or factual priors learned during training. Current evaluations mix these two sources, obscuring when correct visual interpretation is overridden by memorized facts. We present a framework that isolates visual correctness from factual correctness, revealing validity limitations in existing visualization literacy assessments. Across three experiments with 15 state-of-the-art LVLMs: (1) several models reach human-level performance on standard tests (VLAT), but this may reflect factual recall rather than visual understanding, while randomized-data tests (reVLAT) underestimate literacy when correct visual interpretation is superseded by factual priors. (2) Using our Counterfactual Visualization Literacy Assessment Test (CVLAT) with capability-normalized arbitration metrics, we classify models by the sign of their visual-factual reliance index (VFRI), revealing a visualization-oriented majority and a factual knowledge-oriented minority, though several near-zero cases warrant caution. A human baseline (N=30) on the same counterfactual items confirms that people overwhelmingly follow the chart under conflict, providing a human reference point. (3) Prompt-based intervention can shift prioritization, but its effectiveness is highly model-dependent and direction-asymmetric, and high chart-reading capability does not predict prompt-controllability. Overall, high visualization accuracy is not sufficient evidence of faithful visual reasoning: reliable integration into visual analytics requires evaluating not only visualization literacy but also how models arbitrate between visual evidence and factual priors when the two diverge. Benchmark and code: https://github.com/JaeyoungKim-HCIL/CVLAT

Recently, zero-shot 3D scene understanding via 2D Vision-Language Models (VLMs) has gained increasing research interest due to their promising spatial reasoning capabilities. Typically, multiple 2D views are sampled from a 3D point cloud and fed into pre-trained VLMs to answer a given question. This paradigm highlights the critical role of input context quality and raises the challenge of retaining as many task-relevant 3D details as possible under a limited input budget. We propose \texttt{KeyVT}, a hierarchical approach for input context collection at both the view and token levels. Specifically, we combine pixel features with camera parameters and assess view importance based on both semantic content and geometric position, resulting in spatially consistent and task-relevant views. Furthermore, we address redundancy among patches across selected views by identifying representative tokens under the optimal transport (OT) framework, where view tokens and key tokens are formulated as two discrete distributions in the embedding space. These key tokens are expected to cover all view features by minimizing the OT distance. We evaluate our framework on three widely used benchmarks, demonstrating significant improvements over existing tuning-free methods and performance comparable to training-based approaches.

Vision-Language Models (VLMs) inherit the auto-regressive generation paradigm and cache the keys and values (KV) of all previous tokens to accelerate inference, resulting in memory consumption that scales linearly with context length. This issue is particularly pronounced in VLMs due to substantial redundancy in the visual modality. Although KV cache eviction approaches can effectively reduce inference memory, they often incur significant performance degradation in VLMs, as most are designed for language models and overlook the inherent gap between text and vision. By systematically analyzing the modality gap in VLMs in this work, we argue that the importance of visual information should be grounded in textual guidance and accordingly propose a Text-Grounded KV Eviction method for VLMs (TGV-KV). TGV-KV comprises three submodules: (1) Text-Vision Budgeting (TVB) assigns budget to each layer based on the mutual information interaction. (2) Text-Weighted Ranking (TWR) assesses the priority of text and ranks vision importance based on weighted text-image attention. (3) Text-Prioritised Retention (TPR) policy strategically preserves text KV to avoid acute information loss. We evaluate TGV-KV across five models with different sizes and architectures, showing that TGV-KV preserves 99.2% full-KV accuracy on the VizWiz-VQA task with LLaVA-NeXT and boosts end-to-end throughput by 52.6% with an extreme retention budget of 5%. Code is available at https://github.com/Danielement321/TGV-KV.

Deploying language-driven agents in robotics requires evaluations that reflect real-world task demands: natural-language instructions with reproducible outcomes. Such agents must connect language models to callable perception and control tools, and be assessed using deployment-critical metrics including latency, accuracy, and error modes. We present SCOPE (Simulation and Camera Operations for Perception and Evaluation), a modular agent for natural-language, open-vocabulary pan-tilt-zoom (PTZ) camera control and visual scene understanding, designed explicitly for edge deployment. SCOPE operates both in a Blender-based simulation environment and on a physical PTZ camera, executing all perception, planning, and control locally at the deployment site using edge-accessible compute. We release a 536-task benchmark spanning QA, single- and multi-step commands, counting, spatial reasoning, descriptions, and optical character recognition in a Blender-based simulation environment that exposes realistic PTZ control affordances. Execution traces are combined with an LM-as-Judge to evaluate latency, accuracy, and error modes. We evaluate 19 planner-perception model combinations pairing Qwen3 small language models (SLMs) with Moondream and Qwen vision-language models (VLMs). Stronger SLMs substantially reduce hallucinations and improve tool routing, leading to more reliable closed-loop behavior. Once a sufficiently capable SLM is used, perception becomes the dominant performance bottleneck. Mixture-of-Experts models on both the planning and perception side consistently match or exceed dense alternatives at latencies and memory footprints comparable to much smaller networks. Quantization provides additional efficiency gains with minimal accuracy degradation, identifying a practical, sim-to-real validated design point for real-time, edge-feasible language-driven PTZ control.

Supervised fine-tuning is the predominant approach for adapting autoregressive vision-language models to downstream tasks. Recent work has shown that this paradigm is highly vulnerable to backdoor attacks, and that existing defenses are ineffective in open-ended generation settings. In response, we propose BYORn, a backdoor-robust fine-tuning framework motivated by the observation that poisoned target responses are often semantically implausible given the corresponding image-text inputs and a pretrained model. BYORn identifies such misaligned responses and dynamically replaces them with alternative responses generated by the model, thereby breaking the correlation between triggers and target outputs. The resulting objective gradient corresponds to the gradient of the empirical estimate of the population risk upper bound over the clean data distribution. Empirically, BYORn consistently improves robustness to backdoor attacks while preserving clean-task performance, establishing a new trade-off frontier between generalization and attack success rate. Finally, we demonstrate that BYORn remains effective against adaptive attacks specifically designed to circumvent the proposed defense.

Vision-language models (VLMs) are playing an increasingly important role across multiple domains. In many applications, such as robotics, it is crucial to quantify the uncertainty in the output of these models. } We develop FUSE, a probabilistic framework for capturing two complementary sources of uncertainty in vision-language modeling: (i) aleatoric embedding-level uncertainty derived from input data vision-language ambiguity, and (ii) epistemic model-level uncertainty estimated from the semantic response diversity of VLMs. Our approach formulates a Bayesian fusion mechanism that analytically combines these uncertainty sources to produce a scalar measure of uncertainty. This measure can be used to reliably predict the model's output correctness for downstream applications. We demonstrate that our method outperforms baselines and achieves SOTA uncertainty calibration.

Assessing built-environment interaction, such as wheelchair accessibility, is difficult because real-world mobility is shaped by distributed, context-dependent, and temporary barriers that are hard to capture at scale. To support scalable assessment, this paper examines whether vision-language models (VLMs) can identify accessibility barriers from Google Street View (GSV) imagery. We propose an expert-guided retrieval-augmented framework that combines GSV images, ADA-informed guidance, and expert-derived rubrics to evaluate accessibility dimensions. We collect a campus-scale dataset at the University of Florida, linking 407 unique GSV locations with GPS-derived wheelchair dwell behavior as a mobility-friction signal. Results show that VLM ratings are both negatively correlated and distributionally similar with dwell time, indicating partial but consistent alignment with a behavioral proxy for mobility friction. Visual cue analysis shows that certain environmental objects, such as curb ramps and crosswalks, are associated with higher VLM accessibility scores, while alignment remains limited for subtle surface conditions, transient obstructions, and viewpoint-dependent barriers. Overall, our findings show the potential of expert-guided VLMs for scalable accessibility assessment aligning with sensor-derived indicators of real-world wheelchair navigation.

Large-scale generative models have demonstrated remarkable capabilities across image generation and editing tasks. However, their performance in low-level vision tasks, which require pixel-wise control, remains insufficiently studied. To address this gap, we introduce \textbf{LL-Bench}, a comprehensive \textbf{Benchmark} for evaluating the capabilities of large-scale generative models on \textbf{L}ow-\textbf{L}evel vision tasks. The benchmark comprises 2,469 real-world degraded images covering 16 low-level degradation tasks, and 28,919 restored images produced by 10 state-of-the-art large-scale generative models and 21 conventional restoration models, which are annotated with 152,020 expert-level pairwise human preferences and 28,334 quality scores. Built upon LL-Bench, we present a systematic diagnosis that reveals the performance boundaries and unique failure modes of large-scale generative models across diverse low-level vision tasks, compared with conventional representative restoration approaches. Moreover, we investigate the effectiveness of current quality evaluation metrics on LL-Bench, which exhibit significant discrepancy with human ratings. To better align restored-image quality assessment with human preferences, we further propose \textbf{LL-Score}, an MLLM-based evaluator that captures both restoration quality and hallucination existence. Extensive experiments demonstrate that LL-score not only outperforms existing image quality assessment metrics, but also serves as a promising reward model for training generative models on low-level vision tasks.

Enabling Vision-Language Models (VLMs) to perform spatial reasoning remains challenging. Existing approaches treat VLMs as passive observers, which is difficult for real-world applications. Moreover, reinforcement learning methods rely on sparse rewards, limiting their effectiveness for complex reasoning tasks. Inspired by pigeons' building and exploiting cognitive maps for navigation, we propose a novel agentic pipeline for spatial reasoning. First, we introduce a new \emph{dynamic cognitive map} parameterizing scene layout as object positions and orientations, serving as persistent memory for new observations. Second, we propose a novel \emph{Spatial Assertion Codes (SAC)}, Python expressions programmatically describing spatial relationships. By collaborating with the dynamic cognitive map, SAC enables verification of intermediate reasoning steps, providing dense reward signals. We optimize the model via supervised and reinforcement finetuning. Experiments on the MindCube benchmark demonstrate state-of-the-art performance with \emph{80.5%} overall accuracy, outperforming the best current method by \emph{29.5} accuracy points (a relative improvement of \emph{53.2%}) on the challenging \textsc{Rotation} subset. Our code and data are open-sourced at https://github.com/dw-dengwei/active-spatial-reasoning.git.

Understanding subtle driver actions is essential for building reliable driver monitoring systems. Existing visionlanguage models (VLMs) are trained on general datasets and struggle to recognize fine distinctions in driver behaviors. This paper addresses this limitation by creating a detailed natural language version of the Drive&Act dataset. We evaluate three VLMs on our new benchmark using LLM-based scoring methods. Their performance on the new benchmark shows that they cannot reliably generate accurate fine-grained driver activity descriptions. Based on the labeled Drive&Act dataset we create a new Drive&Act description dataset containing finegrained descriptions to train VLMs on driver activity understanding. Cross dataset evaluation on the Driver Monitoring Dataset (DMD) shows that the VLM fine-tuned on our new Drive&Act description dataset generalizes well to actions in the DMD dataset. The VLM fine-tuned on our Drive&Act description dataset achieves an ACCR score of 76 outperforming the zero-shot VLM baseline with an ACCR score of 66. These findings demonstrate that adapting VLMs with richly described driver actions can significantly improve their ability to interpret driver behavior while also highlighting the need for more diverse datasets to support broader generalization in future applications. Our Drive&Act description dataset and code will be publicly available on GitHub.

The joint optimization of image-based (I2I) and text-based (T2I) person re-identification (ReID) is hindered by modality discrepancies and conflicting training objectives, leading to suboptimal shared representations. While I2I ReID focuses on identity-level invariance across images of the same person, T2I ReID is driven by instance-specific textual descriptions tied to unique visual traits. This paper explores the fundamental difference between two ReID tasks and their optimization processes for effective training. Since I2I and T2I ReID are often studied separately, the loss functions optimized for one retrieval setting may negatively affect the representation quality required by the other. Motivated by these findings, we propose a decoupled two-stage training pipeline for learning a shared representation across image and text modalities. The pipeline is based on a single vision encoder that supports both I2I and T2I retrieval while avoiding cross-task interference during training. We provide extensive experiments across multiple configurations, varying domain mixing procedures, learning strategies, and task objectives. We observed that I2I ReID pre-training positively impacts the generalization ability to T2I data. Besides, we find that incorporating textual supervision during the vision encoder training stage enhances both I2I and T2I performance. We believe our insights provide a meaningful step toward unified ReID systems and cross-modal retrieval overall.

Blind image quality assessment (BIQA) predicts perceived image quality without access to a pristine reference and is fundamental to applications such as image compression, transmission, and restoration. Recent BIQA methods increasingly rely on large vision-language models (VLMs). Although frozen VLMs provide an efficient alternative to computationally expensive full fine-tuning, it remains unclear how much performance is sacrificed by not adapting the backbone and, more importantly, under what conditions such adaptation is truly beneficial. Answering this question, however, is complicated by the widespread use of image-level splitting on synthetic-distortion benchmarks, where distorted versions of the same reference image can appear in both training and test partitions. This content overlap artificially inflates the apparent performance of frozen representations, masking their true generalization ability and potentially leading to incorrect conclusions about the value of backbone adaptation. We therefore address these two issues jointly. We develop an efficient BIQA framework that fuses a natural-scene-statistics descriptor with frozen SigLIP and CLIP-H embeddings through a lightweight regression head, and then apply parameter-efficient Low-Rank Adaptation (LoRA) to the SigLIP backbone, training only $0.23\%$ of its parameters. Evaluating both frozen and adapted models across six datasets under image-level and reference-level protocols, we find that image-level splitting inflates frozen-feature SROCC by up to $0.44$ and masks wide variation in true difficulty, which reference-level evaluation reveals. Under this content-independent protocol, LoRA adaptation recovers performance in proportion to the exposed difficulty, with the largest gains where frozen features generalize poorly (up to $+0.357$ SROCC on TID2013) and little benefit where they are already strong.

The way our eyes move while reading provides valuable insights into both the reader's cognitive processes and the properties of the text. In particular, eye-tracking-while-reading data has shown to be highly beneficial in various technological applications, such as enhancing and interpreting language models and inferring a reader's characteristics. However, these applications often rely on large-scale, data-driven models, which demand extensive eye-tracking datasets that are challenging to obtain due to the resource-intensive nature of data collection. To address the challenge of data scarcity, we develop Eyettention II, an end-to-end trained deep-learning model capable of generating realistic scanpaths consisting of a complete set of fixation attributes in chronological order, including fixation location, within-word landing position, and fixation duration. Our model is lightweight, efficiently trainable on limited GPU resources, and closely aligned with cognitive theories. We demonstrate that Eyettention II surpasses state-of-the-art models in scanpath prediction and mirrors human-like gaze behavior by capturing key psycholinguistic phenomena. With its robust performance, Eyettention II holds the potential to drive advancements in natural language processing, facilitate piloting the materials of psycholinguistic experiments, and uncover new insights beyond what is explicitly encoded in theoretical cognitive models.

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