Vision-Language Grounding

Vision-and-Language Navigation requires agents to follow natural-language instructions in visually changing environments. A central challenge is the dynamic entanglement between language and observations: the meaning of instruction shifts as the agent's field of view and spatial context evolve. However, many existing models encode the instruction as a static global representation, limiting their ability to adapt instruction meaning to the current visual context. We therefore model instruction understanding as an Instruction-as-State variable: a decision-relevant, token-level instruction state that evolves step by step conditioned on the agent's perceptual state, where the perceptual state denotes the observation-grounded navigation context at each step. To realize this principle, we introduce State-Entangled Environment-Guided Instruction Understanding (S-EGIU), a coarse-to-fine framework for state-conditioned segment activation and token-level semantic refinement. At the coarse level, S-EGIU activates the instruction segment whose semantics align with the current observation. At the fine level, it refines the activated segment through observation-guided token grounding and contextual modeling, sharpening its internal semantics under the current observation. Together, these stages maintain an instruction state that is continuously updated according to the agent's perceptual state during navigation. S-EGIU delivers strong performance on several key metrics, including a +2.68% SPL gain on REVERIE Test Unseen, and demonstrates consistent efficiency gains across multiple VLN benchmarks, underscoring the value of dynamic instruction--perception entanglement.

Diffusion models have emerged as powerful tools for a wide range of vision tasks, including text-guided image generation and editing. In this work, we explore their potential for object grounding in remote sensing imagery. We propose a hybrid pipeline that integrates diffusion-based localization cues with state-of-the-art segmentation models such as RemoteSAM and SAM3 to obtain more accurate bounding boxes. By leveraging the complementary strengths of generative diffusion models and foundational segmentation models, our approach enables robust and adaptive object localization across complex scenes. Experiments demonstrate that our pipeline significantly improves localization performance, achieving over a 14% increase in [email protected] compared to existing state-of-the-art methods.

Video Large Language Models (Vid-LLMs) have demonstrated remarkable performance in video understanding tasks, yet their robustness under conversational interaction remains largely underexplored. In this paper, we identify spatiotemporal sycophancy, a failure mode in which Vid-LLMs retract initially correct, visually grounded judgments and conform to misleading user feedback under negation-based gaslighting. Rather than merely changing their answers, the models often fabricate unsupported temporal or spatial explanations to justify incorrect revisions. To systematically investigate this phenomenon, we propose a negation-based gaslighting evaluation framework and introduce GasVideo-1000, a curated benchmark designed to probe spatiotemporal sycophancy with clear visual grounding and temporal reasoning requirements. We evaluate a broad range of state-of-the-art open-source and proprietary Vid-LLMs across diverse video understanding tasks. Extensive experiments reveal that vulnerability to negation-based gaslighting is pervasive and severe, even among models with strong baseline performance. While prompt-level grounding constraints can partially mitigate this behavior, they do not reliably prevent hallucinated justifications or belief reversal. Our results indicate that current Vid-LLMs lack robust mechanisms for maintaining grounded spatiotemporal beliefs under adversarial conversational feedback.

Manual annotation of high-quality visual question answering with grounding (VQA-G) datasets, which pair visual questions with evidential grounding, is crucial for advancing vision-language models (VLMs), but remains unscalable. Existing automated methods are often hindered by two key issues: (1) inconsistent data fidelity due to model hallucinations; (2) brittle verification mechanisms based on simple heuristics. To address these limitations, we introduce AutoVQA-G, a self-improving agentic framework for automated VQA-G annotation. AutoVQA-G employs an iterative refinement loop where a Consistency Evaluation module uses Chain-of-Thought (CoT) reasoning for fine-grained visual verification. Based on this feedback, a memory-augmented Prompt Optimization agent analyzes critiques from failed samples to progressively refine generation prompts. Our experiments show that AutoVQA-G generates VQA-G datasets with superior visual grounding accuracy compared to leading multimodal LLMs, offering a promising approach for creating high-fidelity data to facilitate more robust VLM training and evaluation. Code: https://github.com/rohnson1999/AutoVQA-G

Vision-Language Models (VLMs) excel at photorealistic generation, yet often struggle to represent abstract meaning such as idiomatic interpretations of noun compounds. To study whether high visual fidelity interferes with idiomatic compositionality under visual abstraction, we introduce DIVA, a controlled benchmark that replaces high-fidelity visual detail with schematic iconicity by generating paired, sense-anchored visualizations for literal and idiomatic readings. We further propose Semantic Alignment Gap ($Δ$), an architecture-agnostic metric that quantifies divergence between literal and idiomatic visual grounding. We additionally introduce a directional signed bias $b(t)$ to separately measure the direction and strength of literal preference. Evaluating 8 recent VLMs, we reveal a consistent Literal Superiority Bias: model scale alone does not resolve literal preference, and increased visual fidelity is associated with weaker symbolic alignment, suggesting cognitive interference from hyper-realistic imagery. Our findings suggest that improving compositional understanding requires iconographic abstraction of visual input and anchoring interpretation and generation in intended meaning.

Vision-language models enable open-vocabulary object grounding through natural language queries, under the implicit assumption that semantically equivalent descriptions yield consistent outputs. We examine this assumption using a controlled pipeline combining DETR for object proposals with CLIP for language-conditioned selection on 263 COCO val2017 images. We find that overlapping prompts such as "a person," "a human," and "a pedestrian" frequently select different instances, with mean instability of 2.11 distinct selections across six prompts. PCA analysis shows this variability is structured and directional, not random. Prompt ensembling does not improve quality and often shifts selections toward generic regions. We further show that text embedding proximity explains only 34% of grounding disagreement (r = -0.58), confirming that instability arises from the argmax selection mechanism rather than text-level distances alone.

Object level hallucination remains a central reliability challenge for vision language models (VLMs), particularly in binary object existence verification. Existing benchmarks emphasize aggregate accuracy but rarely disentangle whether errors stem from perceptual limitations or from the influence of contextual textual priors, leaving underlying failure mechanisms ambiguous. We introduce DO-Bench, a controlled diagnostic benchmark that isolates these sources through structured multimodal interventions. Rather than evaluating models in unconstrained settings, DO-Bench probes two complementary dimensions: the Prior Override dimension progressively strengthens contextual textual priors while holding visual evidence constant to assess resistance to prior pressure, and the Perception-Limited dimension incrementally enhances visual evidence from full-scene context to localized object crops to measure perceptual grounding strength. This paired design enables attribution of errors to prior suppression, perceptual insufficiency, or their interaction. We further define two diagnostic metrics, PriorRobust and PerceptionAbility, to quantify these behaviors consistently. Evaluations across diverse open- and closed-source VLMs reveal systematic differences in prior sensitivity and perceptual reliability, demonstrating that object hallucination reflects heterogeneous, mechanism dependent failure patterns beyond aggregate accuracy.

Multi-step zoom-in pipelines are widely used for GUI grounding, yet the intermediate predictions they produce are typically discarded after coordinate remapping. We observe that these intermediate outputs contain a useful confidence signal for free: zoom consistency, the distance between a model's step-2 prediction and the crop center. Unlike log-probabilities or token-level uncertainty, zoom consistency is a geometric quantity in a shared coordinate space, making it directly comparable across architecturally different VLMs without calibration. We prove this quantity is a linear estimator of step-1 spatial error under idealized conditions (perfect step-2, target within crop) and show it correlates with prediction correctness across two VLMs (AUC = 0.60; Spearman rho = -0.14, p < 10^{-6} for KV-Ground-8B; rho = -0.11, p = 0.0003 for Qwen3.5-27B). The correlation is small but consistent across models, application categories, and operating systems. As a proof-of-concept, we use zoom consistency to route between a specialist and generalist model, capturing 16.5% of the oracle headroom between them (+0.8%, McNemar p = 0.19). Code is available at https://github.com/omxyz/zoom-consistency-routing.

GUI grounding, which localizes interface elements from screenshots given natural language queries, remains challenging for small icons and dense layouts. Test-time zoom-in methods improve localization by cropping and re-running inference at higher resolution, but apply cropping uniformly across all instances with fixed crop sizes, ignoring whether the model is actually uncertain on each case. We propose UI-Zoomer, a training-free adaptive zoom-in framework that treats both the trigger and scale of zoom-in as a prediction uncertainty quantification problem. A confidence-aware gate fuses spatial consensus among stochastic candidates with token-level generation confidence to selectively trigger zoom-in only when localization is uncertain. When triggered, an uncertainty-driven crop sizing module decomposes prediction variance into inter-sample positional spread and intra-sample box extent, deriving a per-instance crop radius via the law of total variance. Extensive experiments on ScreenSpot-Pro, UI-Vision, and ScreenSpot-v2 demonstrate consistent improvements over strong baselines across multiple model architectures, achieving gains of up to +13.4%, +10.3%, and +4.2% respectively, with no additional training required.

Large vision-language models (VLMs) achieve strong performance on multimodal tasks but often suffer from hallucination and poor grounding in knowledge-intensive reasoning. We propose SmoGVLM, a small, graph-enhanced VLM that integrates structured knowledge with visual and textual modalities, using Graph Neural Networks. We investigate the effects of our method across a range of model sizes, from tiny (1.3B) to large (13B) models. Our results demonstrate that, when trained using our approach, a small model can achieve performance gains upto 16.24%, and surpass its larger counterparts, outperforming larger VLMs and strong fine-tuned baselines. These results highlight the potential of structured knowledge augmentation for efficient, smaller-scale multimodal reasoning systems.

Multimodal Large Language Models frequently suffer from inference hallucinations, partially stemming from language priors dominating visual evidence. Existing training-free mitigation methods either perturb the visual representation and deviate from the natural image distribution, or enforce intrusive manipulations that compromise the model's inherent generative fluency. We introduce a novel perspective that multimodal hallucination manifests as the hypersensitivity of visual grounding to textual phrasing during the decoding phase. Building on this insight, we propose Decoding by Perturbation (DeP), a training-free framework mitigating prior-induced hallucinations via controlled textual interventions. DeP employs a dynamic probe applying multi-level textual perturbations to elicit latent language priors. Leveraging attention variance, it enhances stable evidence regions while suppressing suspicious noise in the feature space. Furthermore, it constructs an interpretable prior drift direction using logits statistics to counteract probability biases from textual co-occurrences. Extensive experiments confirm DeP effectively reduces hallucinations and achieves superior performance across multiple benchmarks.

World models such as DINO-WM and LeWM specify the goal with an image, which is difficult to obtain in advance for novel tasks. We present the Grounded World Model (GWM), a latent world model that enables zero-shot planning in the real world from language goals alone. Given a candidate action sequence and the current observation, GWM predicts the future in the visual space of a pretrained video-language embedding model. The frozen readout of this embedding model maps this imagined future and the task description into the same embedding space, where their negative cosine similarity serves as the planning cost. Training GWM requires only offline and task-agnostic video-action pairs and no language labels. In simulated experiments on WISER, planning with GWM, which executes the candidate action of lowest cost, solves 87% of 288 tasks with unseen instructions and visual signals, while ten fine-tuned VLAs average 22%. We then scale GWM up with real robot data, and use it for zero-shot planning in realistic simulation and real scenes. In the IsaacSim evaluation, planning with GWM completes all 70 trials across 14 tasks that require reasoning over referring expressions, matching a modular planner grounded by a frontier VLM, while pi0.5 reaches 37/70. Deployed on a real Franka, the same stack completes 55/60 separately evaluated pick-and-place sub-tasks, comparable to the modular planner's 52/60, with the full system running locally on a single consumer GPU. Project website: https://quanyili.github.io/gwm-wiser/.

Robot grasping of desktop object is widely used in intelligent manufacturing, logistics, and agriculture.Although vision-language models (VLMs) show strong potential for robotic manipulation, their deployment in low-level grasping faces key challenges: scarce high-quality multimodal demonstrations, spatial hallucination caused by weak geometric grounding, and the fragility of open-loop execution in dynamic environments. To address these challenges, we propose Closed-Loop Asynchronous Spatial Perception(CLASP), a novel asynchronous closed-loop framework that integrates multimodal perception, logical reasoning, and state-reflective feedback. First, we design a Dual-Pathway Hierarchical Perception module that decouples high-level semantic intent from geometric grounding. The design guides the output of the inference model and the definite action tuples, reducing spatial illusions. Second, an Asynchronous Closed-Loop Evaluator is implemented to compare pre- and post-execution states, providing text-based diagnostic feedback to establish a robust error-correction loop and improving the vulnerability of traditional open-loop execution in dynamic environments. Finally, we design a scalable multi-modal data engine that automatically synthesizes high-quality spatial annotations and reasoning templates from real and synthetic scenes without human teleoperation. Extensive experiments demonstrate that our approach significantly outperforms existing baselines, achieving an 87.0% overall success rate. Notably, the proposed framework exhibits remarkable generalization across diverse objects, bridging the sim-to-real gap and providing exceptional robustness in geometrically challenging categories and cluttered scenarios.

While Vision-Language Models (VLMs) demonstrate remarkable zero-shot recognition capabilities across a diverse spectrum of multimodal tasks, it yet remains an open question whether these architectures genuinely comprehend geometric structure or merely exploit RGB textures and contextual priors as statistical shortcuts. Existing evaluations fail to isolate this mechanism, conflating semantic reasoning with texture mapping and relying on imprecise annotations that inadvertently leak environmental cues. To address this gap, we introduce $\textbf{BareBones}$, a zero-shot benchmark designed to stress-test pure geometric shape comprehension. We curate pixel-level silhouettes of geometrically distinct classes across six datasets: five established segmentation sources (ImageNet-S, DIS5K, ThinObject5K, PASCAL VOC, CUB-200) and our novel flagship collection, WTP-Bench, establishing a noise-free geometric taxonomy. WTP-Bench is an extreme, fine-grained visual puzzle that forces models to identify inter-class geometric concepts from boundary contours alone. Our evaluation of 26 state-of-the-art proprietary and open-weight VLMs (eg. GPT-4.1, Gemini, Claude Sonnet 4.5, LLaVA) reveals a consistent, severe performance collapse under RGB deprivation, a phenomenon we term the $\textit{Texture Bias Cliff}$. By documenting universal structural blindspots, BareBones establishes a rigorous yardstick for genuine geometric grounding. Project Page: https://eternal-f1ame.github.io/WTP-Bench/

Vision language action (VLA) models enable generalist robotic agents but often exhibit language ignorance, relying on visual shortcuts and remaining insensitive to instruction changes. We present Prospective Grounding and Alignment VLA (ProGAL-VLA), which constructs a 3D entity-centric graph (GSM), uses a slow planner to produce symbolic sub-goals, and aligns them with grounded entities via a Grounding Alignment Contrastive (GAC) loss. All actions are conditioned on a verified goal embedding $g_t$, whose attention entropy provides an intrinsic ambiguity signal. On LIBERO-Plus, ProGAL-VLA increases robustness under robot perturbations from 30.3 to 71.5 percent, reduces language ignorance by 3x-4x, and improves entity retrieval from 0.41 to 0.71 Recall@1. On the Custom Ambiguity Benchmark, it reaches AUROC 0.81 (vs. 0.52), AUPR 0.79, and raises clarification on ambiguous inputs from 0.09 to 0.81 without harming unambiguous success. The verification bottleneck increases mutual information of language-actions, the GAC loss imposes an entity-level InfoNCE bound, and attention entropy yields calibrated selective prediction, indicating that explicit verified grounding is an effective path toward instruction-sensitive, ambiguity-aware agents.

Large foundation models have made significant advances in embodied intelligence, enabling synthesis and reasoning over egocentric input for household tasks. However, VLM-based auto-labeling is often noisy because the primary data sources lack accurate human action labels, chain-of-thought (CoT), and spatial annotations; these errors are amplified during long-horizon spatial instruction following. These issues stem from insufficient coverage of minute-long, daily household planning tasks and from inaccurate spatial grounding. As a result, VLM reasoning chains and world-model synthesis can hallucinate objects, skip steps, or fail to respect real-world physical attributes. To address these gaps, we introduce EgoTL. EgoTL builds a think-aloud capture pipeline for egocentric data. It uses a say-before-act protocol to record step-by-step goals and spoken reasoning with word-level timestamps, then calibrates physical properties with metric-scale spatial estimators, a memory-bank walkthrough for scene context, and clip-level tags for navigation instructions and detailed manipulation actions. With EgoTL, we are able to benchmark VLMs and World Models on six task dimensions from three layers and long-horizon generation over minute-long sequences across over 100 daily household tasks. We find that foundation models still fall short as egocentric assistants or open-world simulators. Finally, we finetune foundation models with human CoT aligned with metric labels on the training split of EgoTL, which improves long-horizon planning and reasoning, step-wise reasoning, instruction following, and spatial grounding.

When a Vision-Language Model (VLM) sees a blue banana and answers "yellow", is the problem of perception or arbitration? We explore the question in ten VLMs with various sizes and reveal an Encoding-Grounding Dissociation: models that fail to report what they see (and thus provide a wrong answer) still encode the visual evidence as strongly as models that provide the correct answer. Using Multimodal Arbitration Crossover (MAC) analysis with layer-by-layer Logit Lens probing, we track the competition between visual and prior signals across every layer of each model. We show that visual attributes can be linearly decodable from early layers (AUC > 0.86). The accuracy remains nearly identical for both successful and failed samples. However, the gap in the final-layer logit - not the strength of encoding - better predicts grounding outcomes with a correlation of $ρ=$ 0.847. After having studied when VLMs base their answers on image clues rather than prior knowledge, we want to understand the causal relationships. We establish causality through full-sequence activation patching. The standard last-token interventions in LLM interpretability do not affect VLMs. In contrast, replacing the full token sequence at layers identified by MAC alters 60 to 84% of outputs. Partial-token decomposition shows that image tokens carry almost all of the causal impact, while text tokens have none. Scaling addresses the remaining architectural differences to achieve perfect retention. Moving from diagnosis to intervention, we show that training-free activation steering - both linear and sparse autoencoder-guided - in early layers can improve visual grounding by up to +3.8% with degrading performance in some setups. Overall, these findings lead to a clear conclusion: VLMs already see well, but the challenge is acting on what they see. Targeted interventions can help to bridge this gap.

Multimodal large language models (MLLMs) frequently hallucinate objects that are absent from the visual input, often because attention during decoding is disproportionately drawn to visually dominant or frequently occurring content. We observe that this inequity in attention allocation is a root cause of object hallucination: when rare, small, or contextually peripheral objects receive insufficient attention, the model fails to ground its generation in the full visual scene. We argue that every object in an image, regardless of its size, frequency or visual salience, deserves equal representational opportunity during decoding. To this end, we propose DOP-OBC, a training-free and architecture-agnostic decoding strategy built on the principle of equitable attention. Two complementary object-aware signals work in tandem: a Dominant Object Penalty (DOP) that softly suppresses attention over-concentration on visually dominant regions, and an Outlier Boost Coefficient (OBC) that amplifies attention toward rare yet confidently detected objects. These signals are injected as per-row logit modulations within the causal attention mask, requiring no weight updates and preserving autoregressive decoding properties. Extensive experiments across image and video MLLMs demonstrate consistent reductions in object hallucination on CHAIR and POPE benchmarks, alongside improvements in GPT-4o assessed captioning quality across correctness, consistency, detail, context and temporal dimensions. DOP-OBC establishes that fairness in attention allocation is not merely a design principle but a practical and effective path toward more faithful multimodal generation.

Despite rapid progress, pretrained vision-language models still struggle when answers depend on tiny visual details or on combining clues spread across multiple regions, as in documents and compositional queries. We address this by framing grounding as test-time evidence retrieval: given a query, the model should actively identify where to look next to resolve ambiguity. To this end, we propose a training-free, model-intrinsic grounding method that uses uncertainty as supervision. Specifically, we compute the entropy of the model's next-token distribution and backpropagate it to the visual token embeddings to obtain an entropy-gradient relevance map, without auxiliary detectors or attention-map heuristics. We then extract and rank multiple coherent regions to support multi-evidence queries, and introduce an iterative zoom-and-reground procedure with a spatial-entropy stopping rule to avoid over-refinement. Experiments on seven benchmarks across four VLM architectures demonstrate consistent improvements over existing methods, with the largest gains on detail-critical and high-resolution settings, while also producing more interpretable evidence localizations.

Current vision-language pre-training (VLP) paradigms excel at global scene understanding but struggle with instance-level reasoning due to global-only supervision. We introduce InstAP, an Instance-Aware Pre-training framework that jointly optimizes global vision-text alignment and fine-grained, instance-level contrastive alignment by grounding textual mentions to specific spatial-temporal regions. To support this, we present InstVL, a large-scale dataset (2 million images, 50,000 videos) with dual-granularity annotations: holistic scene captions and dense, grounded instance descriptions. On the InstVL benchmark, InstAP substantially outperforms existing VLP models on instance-level retrieval, and also surpasses a strong VLP baseline trained on the exact same data corpus, isolating the benefit of our instance-aware objective. Moreover, instance-centric pre-training improves global understanding: InstAP achieves competitive zero-shot performance on multiple video benchmarks, including MSR-VTT and DiDeMo. Qualitative visualizations further show that InstAP localizes textual mentions to the correct instances, while global-only models exhibit more diffuse, scene-level attention.

Spatio-Temporal Video Grounding requires jointly localizing target objects across both temporal and spatial dimensions based on natural language queries, posing fundamental challenges for existing Multimodal Large Language Models (MLLMs). We identify two core challenges: entangled spatio-temporal alignment, arising from coupling two heterogeneous sub-tasks within the same autoregressive output space, and dual-domain visual token redundancy, where target objects exhibit simultaneous temporal and spatial sparsity, rendering the overwhelming majority of visual tokens irrelevant to the grounding query. To address these, we propose Bridge-STG, an end-to-end framework that decouples temporal and spatial localization while maintaining semantic coherence. While decoupling is the natural solution to this entanglement, it risks creating a semantic gap between the temporal MLLM and the spatial decoder. Bridge-STG resolves this through two pivotal designs: the Spatio-Temporal Semantic Bridging (STSB) mechanism with Explicit Temporal Alignment (ETA) distills the MLLM's temporal reasoning context into enriched bridging queries as a robust semantic interface; and the Query-Guided Spatial Localization (QGSL) module leverages these queries to drive a purpose-built spatial decoder with multi-layer interactive queries and positive/negative frame sampling, jointly eliminating dual-domain visual token redundancy. Extensive experiments across multiple benchmarks demonstrate that Bridge-STG achieves state-of-the-art performance among MLLM-based methods. Bridge-STG improves average m_vIoU from $26.4$ to $34.3$ on VidSTG and demonstrates strong cross-task transfer across various fine-grained video understanding tasks under a unified multi-task training regime.

Vision-Language Models often struggle with complex visual reasoning due to the visual information loss in textual CoT. Existing methods either add the cost of tool calls or rely on localized patch-based embeddings that are insufficient to extract semantics in multi-step reasoning. We propose "Decompose, Look, and Reason" (DLR), a reinforced latent reasoning framework that dynamically decomposes queries into textual premises, extracts premise-conditioned continuous visual latents, and deduces answers through grounded rationales. We introduce a three-stage training pipeline and propose a novel Spherical Gaussian Latent Policy, to enable effective exploration in the latent space. Extensive experiments on vision-centric benchmarks show that DLR consistently outperforms strong baselines, including text-only, interleaved multimodal CoT, and latent reasoning methods, while providing superior stepwise interpretability.

Cultural values are expressed not only through language but also through visual scenes and everyday social practices. Yet existing evaluations of cultural values in language models are almost entirely text-only, leaving it unclear whether culture-conditioned judgments remain stable when response options are visualized. We introduce ValueGround, a benchmark for evaluating culture-conditioned visual value grounding in multimodal large language models (MLLMs). Built from World Values Survey questions, ValueGround uses minimally contrastive image pairs to represent opposing response options while controlling irrelevant variation. Given a country, a question, and an image pair, a model must choose the image that best matches the country's value tendency without access to the original response-option texts. Experiments across six MLLMs and 13 countries show that models perform substantially worse with visualized response options than with the original textual options, with average accuracy dropping from 72.8% to 62.6%. Our benchmark provides a controlled testbed for studying cross-modal transfer of culture-conditioned value judgments.

Video Anomaly Detection (VAD) has traditionally been framed as binary classification or outlier detection, providing neither interpretable reasoning nor precise spatial localization of anomalous events. While Vision-Language Models (VLMs) offer rich scene understanding, they struggle with reliable spatial grounding - often producing hallucinated or geometrically invalid bounding boxes when asked to localize objects. We propose VANGUARD (Video Anomaly Understanding through Reasoning and Grounding), a framework that unifies anomaly classification, spatial grounding, and chain-of-thought reasoning within a single VLM. VANGUARD introduces a three-stage curriculum that progressively layers training objectives: (1) classifier warmup on frozen backbone features, (2) LoRA-adapted spatial grounding, and (3) chain-of-thought generation. To overcome the sparse annotation typical of VAD benchmarks, we employ a teacher-student annotation pipeline in which a VLM (Qwen3-VL-4B) generates structured per-subclip reasoning trajectories based on manual annotations available from the UCA Dataset. Further, GroundingDINO provides bounding box supervision. On UCF-Crime, VANGUARD achieves 94% ROC-AUC with 84% F1 while simultaneously producing interpretable chain-of-thought explanations and spatial grounding of anomalous objects - capabilities absent from prior VAD methods. Ablations confirm that staged training outperforms monolithic optimization, and that structured reasoning acts as an implicit regularizer yielding more balanced predictions than classification-only fine-tuning. Zero-shot transfer to XD-Violence and ShanghaiTech demonstrates cross-domain generalization without target-domain adaptation.

Diffusion large language models (dLLMs) are emerging as promising alternatives to autoregressive (AR) LLMs. Recently, this paradigm has been extended to multimodal tasks, leading to the development of diffusion multimodal large language models (dMLLMs). These models are expected to retain the reasoning capabilities of LLMs while enabling faster inference through parallel generation. However, when combined with Chain-of-Thought (CoT) reasoning, dMLLMs exhibit two critical issues. First, we observe that dMLLMs often generate the final answer token at a very early timestep. This trend indicates that the model determines the answer before sufficient reasoning, leading to degraded reasoning performance. Second, during the initial timesteps, dMLLMs show minimal dependency on visual prompts, exhibiting a fundamentally different pattern of visual information utilization compared to AR vision-language models. In summary, these findings indicate that dMLLMs tend to generate premature final answers without sufficiently grounding on visual inputs. To address these limitations, we propose Position and Step Penalty (PSP) and Visual Reasoning Guidance (VRG). PSP penalizes tokens in later positions during early timesteps, delaying premature answer generation and encouraging progressive reasoning across timesteps. VRG, inspired by classifier-free guidance, amplifies visual grounding signals to enhance the model's alignment with visual evidence. Extensive experiments across various dMLLMs demonstrate that our method achieves up to 7.5% higher accuracy while delivering more than 3x speedup compared to reasoning with four times more diffusion steps.

It is critical for vision-language models (VLMs) to comprehensively understand visual, temporal, and textual cues. However, despite rapid progress in multimodal modeling, video understanding performance still lags behind text-based reasoning. In this work, we find that progress is even worse than previously assumed: commonly reported long video understanding benchmarks contain 40-60% of questions that can be answered using text cues alone. Furthermore, we find that these issues are also pervasive in widely used post-training datasets, potentially undercutting the ability of post-training to improve VLM video understanding performance. Guided by this observation, we introduce VidGround as a simple yet effective solution: using only the actual visually grounded questions without any linguistic biases for post-training. When used in tandem with RL-based post-training algorithms, this simple technique improves performance by up to 6.2 points relative to using the full dataset, while using only 69.1% of the original post-training data. Moreover, we show that data curation with a simple post-training algorithm outperforms several more complex post-training techniques, highlighting that data quality is a major bottleneck for improving video understanding in VLMs. These results underscore the importance of curating post-training data and evaluation benchmarks that truly require visual grounding to advance the development of more capable VLMs. Project page: http://vidground.etuagi.com.

Recent advances in multimodal large language models (MLLMs) have enabled impressive progress in vision-language understanding, yet their high computational cost limits deployment in resource-constrained scenarios such as personal assistants, document understanding, and smart cameras. Most existing methods rely on Transformer-based cross-attention, whose quadratic complexity hinders efficiency. Moreover, small vision-language models often struggle to precisely capture fine-grained, task-relevant visual regions, leading to degraded performance on fine-grained reasoning tasks that limit their effectiveness in the real world. To address these issues, we introduce Firebolt-VL, an efficient vision-language model that replaces the Transformer-based decoder with a Liquid Foundation Model (LFM) decoder. To further enhance visual grounding, we propose a Token-Grid Correlation Module, which computes lightweight correlations between text tokens and image patches and modulates via the state-space model with FiLM conditioning. This enables the model to selectively emphasize visual regions relevant to the textual prompt while maintaining linear-time inference. Experimental results across multiple benchmarks demonstrate that Firebolt-VL achieves accurate, fine-grained understanding with significantly improved efficiency. Our model and code are available at: https://fireboltvl.github.io

Vision-language models (VLMs) have achieved remarkable success across diverse tasks. However, concerns about their trustworthiness persist, particularly regarding tendencies to lean more on textual cues than visual evidence and the risk of producing ungrounded or fabricated responses. To address these issues, we propose Saliency-R1, a framework for improving the interpretability and faithfulness of VLMs reasoning. Specifically, we introduce a novel saliency map technique that efficiently highlights critical image regions contributing to generated tokens without additional computational overhead. This can further be extended to trace how visual information flows through the reasoning process to the final answers, revealing the alignment between the thinking process and the visual context. We use the overlap between the saliency maps and human-annotated bounding boxes as the reward function, and apply Group Relative Policy Optimization (GRPO) to align the salient parts and critical regions, encouraging models to focus on relevant areas when conduct reasoning. Experiments show Saliency-R1 improves reasoning faithfulness, interpretability, and overall task performance.

Vision-language models (VLMs) have achieved impressive performance across a wide range of multimodal tasks. However, they often fail on tasks that require fine-grained visual perception, even when the required information is still present in their internal representations. Prior work has attributed this ``hidden-in-plain-sight'' gap to the language model, but the cause remains unexplained. In this work, we demonstrate that this gap arises from the language model's lack of semantic labels for fine-grained visual details: when visual entities can be mapped to known concepts, VLMs bypass visual comparison and reason through language; when they cannot, VLMs resort to brittle and hallucinated descriptions. We verify this across semantic correspondence, synthetic shape matching, and face matching, and find that VLMs perform much better when the relevant entities are nameable than when they are unnamable. Mechanistically, Logit Lens analysis confirms that VLMs explicitly recover semantic labels for nameable entities and surface more unique tokens compared to unnameable entities. Furthermore, we show that this limitation can be addressed: teaching completely arbitrary names for unknown entities improves performance. More importantly, task-specific finetuning yields even stronger generalization without relying on language priors, i.e., through real visual perception. Our findings suggest that current VLM failures on visual tasks reflect a learned shortcut rather than a fundamental limitation of multimodal reasoning. Code and datasets are available at https://github.com/Patchwork53/VLMs-Need-Words-COLM2026.

Existing visual grounding benchmarks primarily evaluate alignment between image regions and literal referring expressions, where models can often succeed by matching a prominent named category. We explore a complementary and more challenging setting of scenario-based visual grounding, where the target must be inferred from roles, intentions, and relational context rather than explicit naming. We introduce Referring Scenario Comprehension (RSC), a benchmark designed for this setting. The queries in this benchmark are paragraph-length texts that describe object roles, user goals, and contextual cues, including deliberate references to distractor objects that often require deep understanding to resolve. Each instance is annotated with interpretable difficulty tags for uniqueness, clutter, size, overlap, and position which expose distinct failure modes and support fine-grained analysis. RSC contains approximately 31k training examples, 4k in-domain test examples, and a 3k out-of-distribution split with unseen object categories. We further propose ScenGround, a curriculum reasoning method serving as a reference point for this setting, combining supervised warm-starting with difficulty-aware reinforcement learning. Experiments show that scenario-based queries expose systematic failures in current models that standard benchmarks do not reveal, and that curriculum training improves performance on challenging slices and transfers to standard benchmarks.