VLM Interpretability

Echocardiography plays an important role in the screening and diagnosis of cardiovascular diseases. However, automated intelligent analysis of echocardiographic data remains challenging due to complex cardiac dynamics and strong view heterogeneity. In recent years, visual language models (VLM) have opened a new avenue for building ultrasound understanding systems for clinical decision support. Nevertheless, most existing methods formulate this task as a direct mapping from video and question to answer, making them vulnerable to template shortcuts and spurious explanations. To address these issues, we propose EchoTrust, an evidence-driven Actor-Verifier framework for trustworthy reasoning in echocardiography VLM-based agents. EchoTrust produces a structured intermediate representation that is subsequently analyzed by distinct roles, enabling more reliable and interpretable decision-making for high-stakes clinical applications.

Sparse Autoencoders (SAEs) have emerged as a powerful tool for interpreting the internal representations of CLIP vision encoders, yet existing analyses largely focus on the semantic meaning of individual features. We introduce information scope as a complementary dimension of interpretability that characterizes how broadly an SAE feature aggregates visual evidence, ranging from localized, patch-specific cues to global, image-level signals. We observe that some SAE features respond consistently across spatial perturbations, while others shift unpredictably with minor input changes, indicating a fundamental distinction in their underlying scope. To quantify this, we propose the Contextual Dependency Score (CDS), which separates positionally stable local scope features from positionally variant global scope features. Our experiments show that features of different information scopes exert systematically different influences on CLIP's predictions and confidence. These findings establish information scope as a critical new axis for understanding CLIP representations and provide a deeper diagnostic view of SAE-derived features.

Vision-language reward modeling faces a dilemma: generative approaches are interpretable but slow, while discriminative ones are efficient but act as opaque "black boxes." To bridge this gap, we propose VL-MDR (Vision-Language Multi-Dimensional Reward), a framework that dynamically decomposes evaluation into granular, interpretable dimensions. Instead of outputting a monolithic scalar, VL-MDR employs a visual-aware gating mechanism to identify relevant dimensions and adaptively weight them (e.g., Hallucination, Reasoning) for each specific input. To support this, we curate a dataset of 321k vision-language preference pairs annotated across 21 fine-grained dimensions. Extensive experiments show that VL-MDR consistently outperforms existing open-source reward models on benchmarks like VL-RewardBench. Furthermore, we show that VL-MDR-constructed preference pairs effectively enable DPO alignment to mitigate visual hallucinations and improve reliability, providing a scalable solution for VLM alignment.

Large vision-language models (LVLMs) achieve strong performance on visual reasoning tasks but remain highly susceptible to hallucination. Existing detection methods predominantly rely on coarse, whole-image measures of how an object token relates to the input image. This global strategy is limited: hallucinated tokens may exhibit weak but widely scattered correlations across many local regions, which aggregate into deceptively high overall relevance, thus evading the current global hallucination detectors. We begin with a simple yet critical observation: a faithful object token must be strongly grounded in a specific image region. Building on this insight, we introduce a patch-level hallucination detection framework that examines fine-grained token-level interactions across model layers. Our analysis uncovers two characteristic signatures of hallucinated tokens: (i) they yield diffuse, non-localized attention patterns, in contrast to the compact, well-focused attention seen in faithful tokens; and (ii) they fail to exhibit meaningful semantic alignment with any visual region. Guided by these findings, we develop a lightweight and interpretable detection method that leverages patch-level statistical features, combined with hidden-layer representations. Our approach achieves up to 90% accuracy in token-level hallucination detection, demonstrating the superiority of fine-grained structural analysis for detecting hallucinations.

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.

Visual document understanding (VDU) is a challenging task for large vision language models (LVLMs), requiring the integration of visual perception, text recognition, and reasoning over structured layouts. Although recent LVLMs have shown progress on VDU benchmarks, their performance is typically evaluated based on generated responses, which may not necessarily reflect whether the model has actually captured the required information internally. In this paper, we investigate how information required to solve VDU tasks is represented across different layers of LLMs within LVLMs using linear probing. Our study reveals that (1) there is a clear gap between internal representations and generated responses, and (2) information required to solve the task is often encoded more linearly from intermediate layers than from the final layer. Motivated by these findings, we explore fine-tuning strategies that target intermediate layers. Experiments show that fine-tuning intermediate layers improves both linear probing accuracy and response accuracy while narrowing the gap.

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.

2D assembly diagrams are often abstract and hard to follow, creating a need for intelligent assistants that can monitor progress, detect errors, and provide step-by-step guidance. In mixed reality settings, such systems must recognize completed and ongoing steps from the camera feed and align them with the diagram instructions. Vision Language Models (VLMs) show promise for this task, but face a depiction gap because assembly diagrams and video frames share few visual features. To systematically assess this gap, we construct IKEA-Bench, a benchmark of 1,623 questions across 6 task types on 29 IKEA furniture products, and evaluate 19 VLMs (2B-38B) under three alignment strategies. Our key findings: (1) assembly instruction understanding is recoverable via text, but text simultaneously degrades diagram-to-video alignment; (2) architecture family predicts alignment accuracy more strongly than parameter count; (3) video understanding remains a hard bottleneck unaffected by strategy. A three-level mechanistic analysis further reveals that diagrams and video occupy disjoint ViT subspaces, and that adding text shifts models from visual to text-driven reasoning. These results identify visual encoding as the primary target for improving cross-depiction robustness. Project page: https://ryenhails.github.io/IKEA-Bench/

Attention sinks are defined as tokens that attract disproportionate attention. While these have been studied in single modality transformers, their cross-modal impact in Large Vision-Language Models (LVLM) remains largely unexplored: are they redundant artifacts or essential global priors? This paper first categorizes visual sinks into two distinct categories: ViT-emerged sinks (V-sinks), which propagate from the vision encoder, and LLM-emerged sinks (L-sinks), which arise within deep LLM layers. Based on the new definition, our analysis reveals a fundamental performance trade-off: while sinks effectively encode global scene-level priors, their dominance can suppress the fine-grained visual evidence required for local perception. Furthermore, we identify specific functional layers where modulating these sinks most significantly impacts downstream performance. To leverage these insights, we propose Layer-wise Sink Gating (LSG), a lightweight, plug-and-play module that dynamically scales the attention contributions of V-sink and the rest visual tokens. LSG is trained via standard next-token prediction, requiring no task-specific supervision while keeping the LVLM backbone frozen. In most layers, LSG yields improvements on representative multimodal benchmarks, effectively balancing global reasoning and precise local evidence.

Large vision-language models (LVLMs) achieve impressive performance, yet their internal decision-making processes remain opaque, making it difficult to determine if the success stems from true multimodal fusion or from reliance on unimodal priors. To address this attribution gap, we introduce a novel framework using partial information decomposition (PID) to quantitatively measure the "information spectrum" of LVLMs -- decomposing a model's decision-relevant information into redundant, unique, and synergistic components. By adapting a scalable estimator to modern LVLM outputs, our model-agnostic pipeline profiles 26 LVLMs on four datasets across three dimensions -- breadth (cross-model & cross-task), depth (layer-wise information dynamics), and time (learning dynamics across training). Our analysis reveals two key results: (i) two task regimes (synergy-driven vs. knowledge-driven) and (ii) two stable, contrasting family-level strategies (fusion-centric vs. language-centric). We also uncover a consistent three-phase pattern in layer-wise processing and identify visual instruction tuning as the key stage where fusion is learned. Together, these contributions provide a quantitative lens beyond accuracy-only evaluation and offer insights for analyzing and designing the next generation of LVLMs. Code and data are available at https://github.com/RiiShin/pid-lvlm-analysis .

Large-scale vision-language models such as CLIP have achieved remarkable success in zero-shot image recognition, yet their predictions remain largely opaque to human understanding. In contrast, Concept Bottleneck Models provide interpretable intermediate representations by reasoning through human-defined concepts, but they rely on concept supervision and lack the ability to generalize to unseen classes. We introduce EZPC that bridges these two paradigms by explaining CLIP's zero-shot predictions through human-understandable concepts. Our method projects CLIP's joint image-text embeddings into a concept space learned from language descriptions, enabling faithful and transparent explanations without additional supervision. The model learns this projection via a combination of alignment and reconstruction objectives, ensuring that concept activations preserve CLIP's semantic structure while remaining interpretable. Extensive experiments on five benchmark datasets, CIFAR-100, CUB-200-2011, Places365, ImageNet-100, and ImageNet-1k, demonstrate that our approach maintains CLIP's strong zero-shot classification accuracy while providing meaningful concept-level explanations. By grounding open-vocabulary predictions in explicit semantic concepts, our method offers a principled step toward interpretable and trustworthy vision-language models. Code is available at https://github.com/oonat/ezpc.

As Automatic Train Operation (ATO) advances toward GoA4 and beyond, it increasingly depends on efficient, reliable cab-view visual perception and decision-oriented inference to ensure safe operation in complex and dynamic railway environments. However, existing approaches focus primarily on basic perception and often generalize poorly to rare yet safety-critical corner cases. They also lack the high-level reasoning and planning capabilities required for operational decision-making. Although recent Large Multi-modal Models (LMMs) show strong generalization and cognitive capabilities, their use in safety-critical ATO is hindered by high computational cost and hallucination risk. Meanwhile, reliable domain-specific benchmarks for systematically evaluating cognitive capabilities are still lacking. To address these gaps, we introduce RailVQA-bench, the first VQA benchmark for cab-view visual cognition in ATO, comprising 20,000 single-frame and 1,168 video based QA pairs to evaluate cognitive generalization and interpretability in both static and dynamic scenarios. Furthermore, we propose RailVQA-CoM, a collaborative large-small model framework that combines small-model efficiency with large-model cognition via a transparent three-module architecture and adaptive temporal sampling, improving perceptual generalization and enabling more efficient reasoning and planning. Experiments demonstrate that the proposed approach substantially improves performance, enhances interpretability, improves efficiency, and strengthens cross-domain generalization in autonomous driving systems. Code and datasets will be available at https://cybereye-bjtu.github.io/RailVQA.html.

Vision-language models (VLMs) achieve strong multimodal performance, yet how computation is organized across populations of neurons remains poorly understood. In this work, we study VLMs through the lens of neural topology, representing each layer as a within-layer correlation graph derived from neuron-neuron co-activations. This view allows us to ask whether population-level structure is behaviorally meaningful, how it changes across modalities and depth, and whether it identifies causally influential internal components under intervention. We show that correlation topology carries recoverable behavioral signal; moreover, cross-modal structure progressively consolidates with depth around a compact set of recurrent hub neurons, whose targeted perturbation substantially alters model output. Neural topology thus emerges as a meaningful intermediate scale for VLM interpretability: richer than local attribution, more tractable than full circuit recovery, and empirically tied to multimodal behavior. Code is publicly available at https://github.com/he-h/vlm-graph-probing.

Multimodal Large Language Models often suffer from object hallucination. While existing research utilizes attention enhancement and visual retracing, we find these works lack sufficient interpretability regarding attention drift in final model stages. In this paper, we investigate the layer wise evolution of visual features and discover that hallucination stems from deep layer attention regressing toward initial visual noise from early layers. We observe that output reliability depends on acquiring visual anchors at intermediate layers rather than final layers. Based on these insights, we propose CLVA, which stands for Cross-Layer Visual Anchors, a training free method that reinforces critical mid layer features while suppressing regressive noise. This approach effectively pulls deep layer attention back to correct visual regions by utilizing essential anchors captured from attention dynamics. We evaluate our method across diverse architectures and benchmarks, demonstrating outstanding performance without significant increase in computational time and GPU memory.

Sparse autoencoders (SAEs) improve interpretability in multimodal models, but it remains unclear whether SAE features form modular, composable units for reasoning-an assumption underlying many intervention-based steering methods. We test this modularity hypothesis and find it often fails: intervening on a task-selective feature set can modestly improve reasoning accuracy, while intervening on the union of two such sets reliably induces output drift (large unintended changes in predictions) and degrades accuracy, even under norm-matched perturbations. This non modular circuit interference is consistent with shared internal pathways where feature unions amplify activation shifts. We develop a reproducible causal pipeline to localize and test these sparse visual thought circuits in Qwen3-VL-8B. On a controlled synthetic benchmark with seven task types and three difficulty levels, linear probes identify a mid decoder locus for task type information. We train SAEs at this layer, construct task-selective sets via an explicit rule, and perform inference time scaling and ablation while quantifying accuracy and drift. Our findings-validated with bootstrapped subsamples and permutation controls, and replicated across multiple VLM families and five diverse datasets clarify the boundaries of SAE feature composability and provide a rigorous diagnostic framework for more reliable VLM control.

Compressed vision-language models (VLMs) are widely used to reduce memory and compute costs, making them a suitable choice for real-world deployment. However, compressing these models raises concerns about whether internal computations and safety behaviors are preserved. In this work, we use causal circuit analysis and crosscoder-based feature comparisons to examine how pruning and quantization fundamentally change the internals across representative VLMs. We observe that pruning generally keeps circuit structure intact but rotates and attenuates internal features, while quantization modifies the circuits at a higher level yet leaves the surviving features better aligned. Leveraging this insight, we also introduce VLMSafe-420, a novel benchmark that pairs harmful inputs with matched benign counterfactuals across various safety categories. Our findings show that pruning causes a sharp drop in genuine refusal behavior, suggesting that the choice of compression has safety implications.

Vision-language model (VLM) encoders such as CLIP enable strong retrieval and zero-shot classification in a shared image-text embedding space, yet the semantic organization of this space is rarely inspected. We present a post-hoc framework to explain, verify, and align the semantic hierarchies induced by a VLM over a given set of child classes. First, we extract a binary hierarchy by agglomerative clustering of class centroids and name internal nodes by dictionary-based matching to a concept bank. Second, we quantify plausibility by comparing the extracted tree against human ontologies using efficient tree- and edge-level consistency measures, and we evaluate utility via explainable hierarchical tree-traversal inference with uncertainty-aware early stopping (UAES). Third, we propose an ontology-guided post-hoc alignment method that learns a lightweight embedding-space transformation, using UMAP to generate target neighborhoods from a desired hierarchy. Across 13 pretrained VLMs and 4 image datasets, our method finds systematic modality differences: image encoders are more discriminative, while text encoders induce hierarchies that better match human taxonomies. Overall, the results reveal a persistent trade-off between zero-shot accuracy and ontological plausibility and suggest practical routes to improve semantic alignment in shared embedding spaces.

The classic duck-rabbit illusion reveals a fundamental property of perception: when visual evidence is ambiguous, a visual system must resolve competing interpretations. But do humans and machine classifiers resolve the same ambiguity in the same way? To explore this, we use semantically ambiguous images as probes of human-machine perceptual alignment. Our psychophysically-informed framework interpolates between concepts in vision-language embedding space to generate continuous spectra of ambiguous images, enabling direct comparison of human and machine perceptual decision boundaries and sensitivities. We find systematic differences in these boundaries: machine classifiers are shifted relative to both human observers and the midpoint of the synthesis continuum. Guidance scale affects both bias and sensitivity in humans and the machine classifiers tested. Controlled semantic ambiguity in generated images, combined with a psychophysical framework, can thus enable quantitative evaluation of human-machine perceptual alignment.

As vision-language models are deployed at scale, understanding their internal mechanisms becomes increasingly critical. Existing interpretability methods predominantly rely on activations, making them dataset-dependent, vulnerable to data bias, and often restricted to coarse head-level explanations. We introduce SITH (Semantic Inspection of Transformer Heads), a fully data-free, training-free framework that directly analyzes CLIP's vision transformer in weight space. For each attention head, we decompose its value-output matrix into singular vectors and interpret each one via COMP (Coherent Orthogonal Matching Pursuit), a new algorithm that explains them as sparse, semantically coherent combinations of human-interpretable concepts. We show that SITH yields coherent, faithful intra-head explanations, validated through reconstruction fidelity and interpretability experiments. This allows us to use SITH for precise, interpretable weight-space model edits that amplify or suppress specific concepts, improving downstream performance without retraining. Furthermore, we use SITH to study model adaptation, showing how fine-tuning primarily reweights a stable semantic basis rather than learning entirely new features.

As large language models (LLMs) continue to advance, there is increasing interest in their ability to infer human mental states and demonstrate a human-like Theory of Mind (ToM). Most existing ToM evaluations, however, are centered on text-based inputs, while scenarios relying solely on visual information receive far less attention. This leaves a gap, since real-world human-AI interaction typically requires multimodal understanding. In addition, many current methods regard the model as a black box and rarely probe how its internal attention behaves in multiple-choice question answering (QA). The impact of LLM hallucinations on such tasks is also underexplored from an interpretability perspective. To address these issues, we introduce VisionToM, a vision-oriented intervention framework designed to strengthen task-aware reasoning. The core idea is to compute intervention vectors that align visual representations with the correct semantic targets, thereby steering the model's attention through different layers of visual features. This guidance reduces the model's reliance on spurious linguistic priors, leading to more reliable multimodal language model (MLLM) outputs and better QA performance. Experiments on the EgoToM benchmark-an egocentric, real-world video dataset for ToM with three multiple-choice QA settings-demonstrate that our method substantially improves the ToM abilities of MLLMs. Furthermore, results on an additional open-ended generation task show that VisionToM enables MLLMs to produce free-form explanations that more accurately capture agents' mental states, pushing machine-human collaboration toward greater alignment.

Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it ``sees'', effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.

Many multimodal tasks, such as image captioning and visual question answering, require vision-language models (VLMs) to bind objects with their properties and spatial relations. Yet it remains unclear where and how such associations are computed within VLMs. In this work, we show that VLMs rely on two concurrent mechanisms to represent spatial variable binding. In the language model backbone, intermediate layers represent content-independent spatial relations on top of visual tokens corresponding to objects. However, this mechanism plays only a secondary role in shaping model predictions. Instead, the dominant source of spatial information originates in the vision encoder, whose representations encode the layout of objects and are directly exploited by the language model backbone. Notably, this spatial signal is distributed globally across visual tokens, extending beyond object regions into surrounding background areas. We validate the generalization of our findings to complex natural images from the COCO dataset, where globally amplifying the vision-derived spatial representations across all image tokens corrects spatial variable binding failures across models of various sizes. Together, our results clarify how spatial variable binding is computed within VLMs and highlight the central role of vision encoders in enabling it.

Sleep staging is essential for sleep assessment and disorder diagnosis. In recent years, automatic sleep staging systems have achieved accuracy approaching that of human experts, but the black-box nature of their predictions hinders clinical adoption. Existing interpretability methods offer partial insight into model behavior, but their outputs still require expert reinterpretation and do not provide a direct basis for auditing individual predictions. To improve trustworthiness, we propose the task of auditable sleep staging. To solve this task, we present SleepVLM, a vision-language model that casts sleep staging as visual reasoning over rendered polysomnography (PSG) waveform images. For each epoch, SleepVLM outputs a stage together with the applicable American Academy of Sleep Medicine (AASM) rules and an auditable rationale. The model is trained using a two-stage framework: Waveform-Perceptual Pre-training followed by Rule-Grounded Supervised Fine-tuning over a mixture of fine-grained and coarse annotations. Experiments on four datasets show that SleepVLM outperforms state-of-the-art methods on average. An automated AASM-feature audit shows broad coverage of stage-defining evidence in the rationales, and independent experts validate their reasoning quality. To facilitate further research, we construct and release MASS-EX, an expert-annotated dataset for rule-grounded sleep staging with AASM rule annotations and expert-written rationales.

Despite remarkable advances in large Vision-Language Models (VLMs), spatial reasoning remains a persistent challenge. In this work, we investigate how attention heads within VLMs contribute to spatial reasoning by analyzing their functional roles through a mechanistic interpretability lens. We introduce CogVSR, a dataset that decomposes complex spatial reasoning questions into step-by-step subquestions designed to simulate human-like reasoning via a chain-of-thought paradigm, with each subquestion linked to specific cognitive functions such as spatial perception or relational reasoning. Building on CogVSR, we develop a probing framework to identify and characterize attention heads specialized for these functions. Our analysis across diverse VLM families reveals that these functional heads are universally sparse, vary in number and distribution across functions. Notably, spatially specialized heads are fewer than those for other cognitive functions, highlighting their scarcity. We propose methods to activate latent spatial heads, improving spatial understanding. Intervention experiments further demonstrate their critical role in spatial reasoning: removing functional heads leads to performance degradation, while emphasizing them enhances accuracy. This study provides new interpretability driven insights into how VLMs attend to space and paves the way for enhancing complex spatial reasoning in multimodal models.

Standard attribution heatmaps show where a vision-language model (VLM) focuses, but they do not reveal whether the recovered evidence is organized by the queried spatial relation or merely reflects image layout. To address this problem, we introduce CREG (Compass Relational Evidence Graph), a training-free diagnostic framework that converts token-level attribution into a reference-centered compass distribution and measures its directional alignment. CREG provides a shared directional readout across attribution methods and makes comparison with geometric controls explicit. Across three spatial-relation benchmarks, box-only geometry achieves Direction Alignment Error 28.4 to 34.4 degrees lower than the best current model-based attribution method on each dataset, leaving a substantial gap between attribution structure and simple target localization. To examine this gap, we apply a diagnostic battery including target intervention, reference-center randomization, and variance partition. Taken together, the results suggest that the directional structure recoverable from current attribution methods is limited and often mixed with image layout. We further find that higher task accuracy does not reliably coincide with better directional attribution: small-scale LoRA training and newer model generations can improve task accuracy while leaving Direction Alignment Error unchanged or worse. These findings characterize what current attribution methods reveal rather than the model's internal spatial representation. CREG provides a controlled protocol for testing whether improvements in spatial reasoning are accompanied by more directionally organized evidence.

How much scene context a single object carries is a well-studied question in human scene perception, yet how this capacity is organized in vision-language models (VLMs) remains poorly understood, with direct implications for the robustness of these models. We investigate this question through a systematic behavioral and mechanistic analysis of contextual inference from single objects. Presenting VLMs with single objects on masked backgrounds, we probe their ability to infer both fine-grained scene category and coarse superordinate context (indoor vs. outdoor). We found that single objects support above-chance inference at both levels, with performance modulated by the same object properties that predict human scene categorization. Object identity, scene, and superordinate predictions are partially dissociable: accurate inference at one level neither requires nor guarantees accurate inference at the others, and the degree of coupling differs markedly across models. Mechanistically, object representations that remain stable when background context is removed are more predictive of successful contextual inference. Scene and superordinate schemas are grounded in fundamentally different ways: scene identity is encoded in image tokens throughout the network, while superordinate information emerges only late or not at all. Together, these results reveal that the organization of contextual inference in VLMs is more complex than accuracy alone suggests, with behavioral and mechanistic signatures

Counting serves as a simple but powerful test of a Large Vision-Language Model's (LVLM's) reasoning; it forces the model to identify each individual object and then add them all up. In this study, we investigate how LVLMs implement counting using controlled synthetic and real-world benchmarks, combined with mechanistic analyses. Our results show that LVLMs display a human-like counting behavior, with precise performance on small numerosities and noisy estimation for larger quantities. We introduce two novel interpretability methods, Visual Activation Patching and HeadLens, and use them to uncover a structured "counting circuit" that is largely shared across a variety of visual reasoning tasks. Building on these insights, we propose a lightweight intervention strategy that exploits simple and abundantly available synthetic images to fine-tune arbitrary pretrained LVLMs exclusively on counting. Despite the narrow scope of this fine-tuning, the intervention not only enhances counting accuracy on in-distribution synthetic data, but also yields an average improvement of +8.36% on out-of-distribution counting benchmarks and an average gain of +1.54% on complex, general visual reasoning tasks for Qwen2.5-VL. These findings highlight the central, influential role of counting in visual reasoning and suggest a potential pathway for improving overall visual reasoning capabilities through targeted enhancement of counting mechanisms.

Cross-Domain Few-Shot Learning (CDFSL) adapts models trained with large-scale general data (source domain) to downstream target domains with only scarce training data, where the research on vision-language models (e.g., CLIP) is still in the early stages. Typical downstream domains, such as medical diagnosis, require fine-grained visual cues for interpretable recognition, but we find that current fine-tuned CLIP models can hardly focus on these cues, albeit they can roughly focus on important regions in source domains. Although current works have demonstrated CLIP's shortcomings in capturing local subtle patterns, in this paper, we find that the domain gap and scarce training data further exacerbate such shortcomings, much more than that of holistic patterns, which we call the local misalignment problem in CLIP-based CDFSL. To address this problem, due to the lack of supervision in aligning local visual features and text semantics, we turn to self-supervision information. Inspired by the translation task, we propose the CC-CDFSL method with cycle consistency, which translates local visual features into text features and then translates them back into visual features (and vice versa), and constrains the original features close to the translated back features. To reduce the noise imported by richer information in the visual modality, we further propose a Semantic Anchor mechanism, which first augments visual features to provide a larger corpus for the text-to-image mapping, and then shrinks the image features to filter out irrelevant image-to-text mapping. Extensive experiments on various benchmarks, backbones, and fine-tuning methods show we can (1) effectively improve the local vision-language alignment, (2) enhance the interpretability of learned patterns and model decisions by visualizing patches, and (3) achieve state-of-the-art performance.

Multimodal Large Language Models (MLLMs) are increasingly applied to pixel-level vision tasks, yet their intrinsic capacity for spatial understanding remains poorly understood. We investigate segmentation capacity through a layerwise linear probing evaluation across the entire MLLM pipeline: vision encoder, adapter, and LLM. We further conduct an intervention based attention knockout analysis to test whether cross-token attention progressively refines visual representations, and an evaluation of bidirectional attention among image tokens on spatial consistency. Our analysis reveals that the adapter introduces a segmentation representation drop-off, but LLM layers progressively recover through attention-mediated refinement, where correctly classified tokens steer misclassified neighbors toward the correct label. At early image token positions, this recovery is bounded by causal attention, which bidirectional attention among image tokens alleviates. These findings provide a mechanistic account of how MLLMs process visual information for segmentation, informing the design of future segmentation-capable models.

Urban flooding poses an escalating threat to transportation network continuity, yet no operational system currently provides real-time, street-level flood depth information at the centimeter resolution required for dynamic routing, electric vehicle (EV) safety, and autonomous vehicle (AV) operations. This study presents FloodLlama, a fine-tuned open-source vision-language model (VLM) for continuous flood depth estimation from single street-level images, supported by a multimodal sensing pipeline using TikTok data. A synthetic dataset of approximately 190000 images was generated, covering seven vehicle types, four weather conditions, and 41 depth levels (0-40 cm at 1 cm resolution). Progressive curriculum training enabled coarse-to-fine learning, while LLaMA 3.2-11B Vision was fine-tuned using QLoRA. Evaluation across 34797 trials reveals a depth-dependent prompt effect: simple prompts perform better for shallow flooding, whereas chain-of-thought (CoT) reasoning improves performance at greater depths. FloodLlama achieves a mean absolute error (MAE) below 0.97 cm and Acc@5cm above 93.7% for deep flooding, exceeding 96.8% for shallow depths. A five-phase mechanistic interpretability framework identifies layer L23 as the critical depth-encoding transition and enables selective fine-tuning that reduces trainable parameters by 76-80% while maintaining accuracy. The Tier 3 configuration achieves 98.62% accuracy on real-world data and shows strong robustness under visual occlusion. A TikTok-based data pipeline, validated on 676 annotated flood frames from Detroit, demonstrates the feasibility of real-time, crowd-sourced flood sensing. The proposed framework provides a scalable, infrastructure-free solution with direct implications for EV safety, AV deployment, and resilient transportation management.