VLM Interpretability

Despite the rapid evolution of training paradigms, the decoder backbone of large vision--language models (LVLMs) remains fundamentally rooted in the residual-connection Transformer architecture. Therefore, deciphering the distinct roles of internal modules is critical for understanding model mechanics and guiding architectural optimization. While prior statistical approaches have provided valuable attribution-based insights, they often lack a unified theoretical basis. To bridge this gap, we propose a unified framework grounded in information theory and geometry to quantify the geometric and entropic nature of residual updates. Applying this unified framework reveals a fundamental functional decoupling: Attention acts as a subspace-preserving operator focused on reconfiguration, whereas FFNs serve as subspace-expanding operators driving semantic innovation. Strikingly, further experiments demonstrate that replacing learned attention weights with predefined values (e.g., Gaussian noise) yields comparable or even superior performance across a majority of datasets relative to vanilla models. These results expose severe misallocation and redundancy in current mechanisms, suggesting that state-of-the-art LVLMs effectively ``get lost in attention'' rather than efficiently leveraging visual context.

Despite the remarkable success of Multimodal Large Language Models (MLLMs) across diverse tasks, the internal mechanisms governing how they encode and ground distinct visual concepts remain poorly understood. To unravel these mechanisms, we propose a causal framework based on activation steering to actively probe and manipulate internal visual representations. Through systematic intervention across four visual concept categories, our results reveal a divergence in concept encoding: entity knowledge is distinctively localized, whereas abstract concepts are globally distributed across the network. Critically, this divergence uncovers a mechanistic driver of scaling laws: increasing model depth is indispensable for encoding distributed and complex abstract concepts, whereas entities maintain a consistently high degree of localization. Furthermore, reverse steering uncovers that blocking explicit output triggers a surge in latent activations, exposing a compensatory mechanism between perception and generation. Finally, by extending our analysis to visual reasoning, we expose a disconnect between perception and reasoning: although MLLMs successfully recognize geometric relations, they treat them merely as static visual features, failing to trigger the procedural execution necessary for solving problems.

Vision-language models such as CLIP achieve strong visual-textual alignment, but often suffer from overfitting and limited interpretability when adapted through continuous prompt learning. While discrete prompt optimization improves interpretability, it usually depends on large external models, leading to high computational costs and limited scalability. In this paper, we propose Interpretable Prompt Learning (IPL), a hybrid framework that alternates between discrete semantic token selection and continuous prompt optimization. Specifically, IPL formulates semantic token selection as an approximate submodular optimization problem, encouraging tokens that are both human-understandable and semantically diverse. It further adopts an alternating optimization strategy to integrate discrete token selection with continuous prompt tuning, improving interpretability while preserving adaptability to downstream tasks. Our framework is plug-and-play, allowing seamless integration with existing prompt learning methods. Extensive experiments on multiple benchmarks show that IPL consistently improves both interpretability and accuracy across five representative prompt learning methods, providing an effective and scalable extension to existing frameworks.

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

Human attention is the gateway to conscious perception, memory and decision-making. However, its role in modern transformer models remains largely unexplored. As these systems increasingly influence what people see, prefer and buy, the question arises as to whether they encode principles of human interest or merely exploit large-scale correlations. Addressing this issue is crucial for understanding cognition and ensuring the responsible use of AI in communication and marketing. In order to address this issue, the concept of visual interest was examined within the multimodal vision-language-model Qwen3-VL-8B, using a pre-defined Common Interestingness (CI) score derived from large-scale human engagement data on the photo-sharing platform Flickr. Here, we analyzed internal representations across vision and language components using methods from the neurosciences. Our analyses revealed that CI information is linearly decodable from final-layer embeddings, indicating that it is aligned with human-derived measures of visual interestingness. Dimensionality reduction and Generalized Discrimination Value (GDV) analyses demonstrate that CI-related hidden representations emerge in intermediate vision transformer layers and becomes progressively more distinguishable across language model layers. Concept vectors derived using geometric, probe, and Sparse Auto-Encoder based methods converge in higher layers, as confirmed by representational similarity analysis. This indicates a robust and structured encoding of visual interestingness without explicit supervision. Future work will seek to identify shared computational principles linking human brain dynamics and transformer architectures, with the ultimate goal of uncovering the organizing mechanisms that give rise to attention and interest in both biological and artificial systems.

Vision-Language-Action (VLA) policies often fail under distribution shift, suggesting that decisions may depend on spurious visual correlations rather than task-relevant causes. We formulate visual-action attribution as an interventional estimation problem. Accordingly, we introduce the Interventional Significance Score (ISS), an interventional masking procedure for estimating the causal influence of visual regions on action predictions, and the Nuisance Mass Ratio (NMR), a scalar measure of attribution to task-irrelevant features. We analyze the statistical properties of ISS and show that it admits unbiased estimation, and we characterize conditions under which action prediction error provides a valid proxy for causal influence. Experiments across diverse manipulation tasks indicate that NMR predicts generalization behavior and that ISS yields more faithful explanations than existing interpretability methods. These results suggest that interventional attribution provides a simple diagnostic approach for identifying causal misalignment in embodied policies.

Text-to-image person re-identification (TI-ReID) relies on natural-language text descriptions to retrieve top matching individuals from a gallery of reference images. While recent large vision-language models (VLMs) achieve strong retrieval performance, their decisions remain largely uninterpretable. Existing interpretability approaches in TI-ReID rely solely on slot-attention to highlight attended regions, but fail to reliably bind visual regions to semantically meaningful concepts, limiting interpretation to qualitative visualizations over a restricted vocabulary. This paper introduces InterPartAbility, an interpretable TI-ReID method that performs explicit part-wise matching and enables phrase-region grounding. Unlike parameter-heavy slot-attention methods that yield only qualitative interpretability, our open-vocabulary patch-phrase interaction module (PPIM) guides a standard TI-ReID model with concept-level phrases. Concept-based part phrases provide evidence that encourages the model to attend to the corresponding local image regions. InterPartAbility further leverages CLIP ViT self-attention to produce spatially concentrated patch activations aligned with each part-level phrase, yielding grounded explanation maps. Finally, a quantitative interpretability protocol for TI-ReID is introduced that extends current perturbation-based evaluation metrics into the TI-Reid domain. This includes a counterfactual region removal that measures retrieval degradation when top-ranked explanatory regions are removed. Empirical results on three challenging benchmarks show that InterPartAbility can achieve SOTA interpretability performance under these metrics, while sustaining competitive retrieval accuracy.

In recent years, the integration of multimodal machine learning in wellbeing assessment has offered transformative potential for monitoring mental health. However, with the rapid advancement of Vision-Language Models (VLMs), their deployment in clinical settings has raised concerns due to their lack of transparency and potential for bias. While previous research has explored the intersection of fairness and Explainable AI (XAI), its application to VLMs for wellbeing assessment and depression prediction remains under-explored. This work investigates VLM performance across laboratory (AFAR-BSFT) and naturalistic (E-DAIC) datasets, focusing on diagnostic reliability and demographic fairness. Performance varied substantially across environments and architectures; Phi3.5-Vision achieved 80.4% accuracy on E-DAIC, while Qwen2-VL struggled at 33.9%. Additionally, both models demonstrated a tendency to over-predict depression on AFAR-BSFT. Although bias existed across both architectures, Qwen2-VL showed higher gender disparities, while Phi-3.5-Vision exhibited more racial bias. Our XAI intervention framework yielded mixed results; fairness prompting achieved perfect equal opportunity for Qwen2-VL at a severe accuracy cost on E-DAIC. On AFAR-BSFT, explainability-based interventions improved procedural consistency but did not guarantee outcome fairness, sometimes amplifying racial bias. These results highlight a persistent gap between procedural transparency and equitable outcomes. We analyse these findings and consolidate concrete recommendations for addressing them, emphasising that future fairness interventions must jointly optimise predictive accuracy, demographic parity, and cross-domain generalisation.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in cross-modal understanding and generation, yet their deployment is threatened by critical safety vulnerabilities. While prior works have demonstrated the feasibility of backdoors in MLLMs via fine-tuning data poisoning to manipulate inference, the underlying mechanisms of backdoor attacks remain opaque, complicating the understanding and mitigation. To bridge this gap, we propose ProjLens, an interpretability framework designed to demystify MLLMs backdoors. We first establish that normal downstream task alignment--even when restricted to projector fine--tuning--introduces vulnerability to backdoor injection, whose activation mechanism is different from that observed in text-only LLMs. Through extensive experiments across four backdoor variants, we uncover:(1) Low-Rank Structure: Backdoor injection updates appear overall full-rank and lack dedicated ``trigger neurons'', but the backdoor-critical parameters are encoded within a low-rank subspace of the projector;(2) Activation Mechanism: Both clean and poisoned embedding undergoes a semantic shift toward a shared direction aligned with the backdoor target, but the shifting magnitude scales linearly with the input norm, resulting in the distinct backdoor activation on poisoned samples. Our code is available at: https://anonymous.4open.science/r/ProjLens-8FD7

Visual agents operating in the wild must respond to queries precisely when sufficient evidence first appears in a video stream, a critical capability that is overlooked by conventional video LLMs evaluated in offline settings. The shift to an online, streaming paradigm introduces significant challenges: a lack of decision transparency, the difficulty of aligning response timing with visual evidence, and the need to maintain a global, causally consistent understanding under tight computational budgets. To address these issues, we propose a novel framework that decouples reasoning control from memory integration. We introduce \model{}, an instantiation of this framework with two core components. First, the Active Thinking Decision Maker (ATDM) is a transparent reasoning controller that externalizes its decision process using observable progress ($\boldsymbolρ$) and confidence ($\boldsymbol{c}$) metrics. This allows it to precisely time its response $t_r$ to match the first-sufficient-evidence timestamp $t^\star$ while streaming its reasoning to the user. Second, the Hierarchical Progressive Semantic Integration (HPSI) module acts as an efficient memory system. It employs a set of learnable, multi-level aggregation tokens that are propagated across clips to build a rich, global cognitive state without exceeding token budgets. %Our approach sets a new standard on key online video understanding benchmarks, achieving strong performance of 71.6% on StreamingBench and 46.9% on OVOBench, demonstrating a robust solution for evidence-aligned and transparent online video analysis. Extensive experiments demonstrate the effectiveness of ATDM and HPSI, e.g., Thinking-QwenVL improves the accuracy of the previous state-of-the-art from 67.63% to 71.60% on the StreamingBench benchmark.

Recent work has increasingly explored neuron-level interpretation in vision-language models (VLMs) to identify neurons critical to final predictions. However, existing neuron analyses generally focus on single tasks, limiting the comparability of neuron importance across tasks. Moreover, ranking strategies tend to score neurons in isolation, overlooking how task-dependent information pathways shape the write-in effects of feed-forward network (FFN) neurons. This oversight can exacerbate neuron polysemanticity in multi-task settings, introducing noise into the identification and intervention of task-critical neurons. In this study, we propose HONES (Head-Oriented Neuron Explanation & Steering), a gradient-free framework for task-aware neuron attribution and steering in multi-task VLMs. HONES ranks FFN neurons by their causal write-in contributions conditioned on task-relevant attention heads, and further modulates salient neurons via lightweight scaling. Experiments on four diverse multimodal tasks and two popular VLMs show that HONES outperforms existing methods in identifying task-critical neurons and improves model performance after steering. Our source code is released at: https://github.com/petergit1/HONES.

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.

We present an amortized framework for real-time visual attribution streaming in multimodal thinking models. When these models generate code from a screenshot or solve math problems from images, their long reasoning traces should be grounded in visual evidence. However, verifying this reliance is challenging: faithful causal methods require costly repeated backward passes or perturbations, while raw attention maps offer instant access, they lack causal validity. To resolve this, we introduce an amortized approach that learns to estimate the causal effects of semantic regions directly from the rich signals encoded in attention features. Across five diverse benchmarks and four thinking models, our approach achieves faithfulness comparable to exhaustive causal methods while enabling visual attribution streaming, where users observe grounding evidence as the model reasons, not after. Our results demonstrate that real-time, faithful attribution in multimodal thinking models is achievable through lightweight learning, not brute-force computation.

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

Recent advances in vision language models (VLMs) offer reasoning capabilities, yet how these unfold and integrate visual and textual information remains unclear. We analyze reasoning dynamics in 18 VLMs covering instruction-tuned and reasoning-trained models from two different model families. We track confidence over Chain-of-Thought (CoT), measure the corrective effect of reasoning, and evaluate the contribution of intermediate reasoning steps. We find that models are prone to answer inertia, in which early commitments to a prediction are reinforced, rather than revised during reasoning steps. While reasoning-trained models show stronger corrective behavior, their gains depend on modality conditions, from text-dominant to vision-only settings. Using controlled interventions with misleading textual cues, we show that models are consistently influenced by these cues even when visual evidence is sufficient, and assess whether this influence is recoverable from CoT. Although this influence can appear in the CoT, its detectability varies across models and depends on what is being monitored. Reasoning-trained models are more likely to explicitly refer to the cues, but their longer and fluent CoTs can still appear visually grounded while actually following textual cues, obscuring modality reliance. In contrast, instruction-tuned models refer to the cues less explicitly, but their shorter traces reveal inconsistencies with the visual input. Taken together, these findings indicate that CoT provides only a partial view of how different modalities drive VLM decisions, with important implications for the transparency and safety of multimodal systems.

Transparency of neural networks' internal reasoning is at the heart of interpretability research, adding to trust, safety, and understanding of these models. The field of mechanistic interpretability has recently focused on studying task-specific computational graphs, defined by connections (edges) between model components. Such edge-based circuits have been defined in the context of large language models, yet vision-based approaches so far only consider neuron-based circuits. These tell which information is encoded, but not how it is routed through the complex wiring of a neural network. In this work, we investigate whether useful mechanistic circuits can be identified through computational graphs in vision transformers. We propose an effective method for Automatic Visual Circuit Discovery (Vi-CD) that recovers class-specific circuits for classification, identifies circuits underlying typographic attacks in CLIP, and discovers circuits that lend themselves for steering to correct harmful model behavior. Overall, we find that insightful and actionable edge-based circuits can be recovered from vision transformers, adding transparency to the internal computations of these models.

Multimodal Large Language Models (MLLMs) struggle with tasks that require reasoning about 2D object orientation in images, as documented in prior work. Tong et al. and Nichols et al. hypothesize that these failures originate in the visual encoder, since commonly used encoders such as CLIP and SigLIP are trained for image-text semantic alignment rather than geometric reasoning. We design a controlled empirical protocol to test this claim by measuring whether rotations can be recovered from encoder representations. In particular, we examine SigLIP and ViT features from LLaVA OneVision and Qwen2.5-VL-7B-Instruct models, respectively, using full images, and examine CLIP representations in LLaVA 1.5 and 1.6 using rotated foreground patches against natural background images. Our null hypothesis is that orientation information is not preserved in the encoder embeddings and we test this by training linear regressors to predict object orientation from encoded features. Contrary to the hypothesis, we find that orientation information is recoverable from encoder representations: simple linear models accurately predict object orientations from embeddings. This contradicts the assumption that MLLM orientation failures originate in the visual encoder. Having rejected the accepted hypothesis that MLLMs struggle with 2D orientation tasks because of visual encoder limitations, we still don't know why they fail. Although a full explanation is beyond the scope of this paper, we show that although present, orientation information is spread diffusely across tens of thousands of features. This may or may not be while MLLMs fail to exploit the available orientation information.

Understanding the internal activations of Vision Transformers (ViTs) is critical for building interpretable and trustworthy models. While Sparse Autoencoders (SAEs) have been used to extract human-interpretable features, they operate on individual layers and fail to capture the cross-layer computational structure of Transformers, as well as the relative significance of each layer in forming the last-layer representation. Alternatively, we introduce the adoption of Cross-Layer Transcoders (CLTs) as reliable, sparse, and depth-aware proxy models for MLP blocks in ViTs. CLTs use an encoder-decoder scheme to reconstruct each post-MLP activation from learned sparse embeddings of preceding layers, yielding a linear decomposition that transforms the final representation of ViTs from an opaque embedding into an additive, layer-resolved construction that enables faithful attribution and process-level interpretability. We train CLTs on CLIP ViT-B/32 and ViT-B/16 across CIFAR-100, COCO, and ImageNet-100. We show that CLTs achieve high reconstruction fidelity with post-MLP activations while preserving and even improving, in some cases, CLIP zero-shot classification accuracy. In terms of interpretability, we show that the cross-layer contribution scores provide faithful attribution, revealing that the final representation is concentrated in a smaller set of dominant layer-wise terms whose removal degrades performance and whose retention largely preserves it. These results showcase the significance of adopting CLTs as an alternative interpretable proxy of ViTs in the vision domain.

While medical Vision-Language models (VLMs) achieve strong performance on tasks such as tumor or organ segmentation and diagnosis prediction, their opaque latent representations limit clinical trust and the ability to explain predictions. Interpretability of these multimodal representations are therefore essential for the trustworthy clinical deployment of pretrained medical VLMs. However, current interpretability methods, such as gradient- or attention-based visualizations, are often limited to specific tasks such as classification. Moreover, they do not provide concept-level explanations derived from shared pretrained representations that can be reused across downstream tasks. We introduce MedConcept, a framework that uncovers latent medical concepts in a fully unsupervised manner and grounds them in clinically verifiable textual semantics. MedConcept identifies sparse neuron-level concept activations from pretrained VLM representations and translates them into pseudo-report-style summaries, enabling physician-level inspection of internal model reasoning. To address the lack of quantitative evaluation in concept-based interpretability, we introduce a quantitative semantic verification protocol that leverages an independent pretrained medical LLM as a frozen external evaluator to assess concept alignment with radiology reports. We define three concept scores, Aligned, Unaligned, and Uncertain, to quantify semantic support, contradiction, or ambiguity relative to radiology reports and use them exclusively for post hoc evaluation. These scores provide a quantitative baseline for assessing interpretability in medical VLMs. All codes, prompt and data to be released on acceptance. Ke

Explainable AI (XAI) methods reveal which features influence model predictions, yet provide limited means for practitioners to act on these explanations. Activation steering of components identified via XAI offers a path toward actionable explanations, although its practical utility remains understudied. We introduce an interactive workflow combining SAE-based attribution with activation steering for instance-level analysis of concept usage in vision models, implemented as a web-based tool. Based on this workflow, we conduct semi-structured expert interviews (N=8) with debugging tasks on CLIP to investigate how practitioners reason about, trust, and apply activation steering. We find that steering enables a shift from inspection to intervention-based hypothesis testing (8/8 participants), with most grounding trust in observed model responses rather than explanation plausibility alone (6/8). Participants adopted systematic debugging strategies dominated by component suppression (7/8) and highlighted risks including ripple effects and limited generalization of instance-level corrections. Overall, activation steering renders interpretability more actionable while raising important considerations for safe and effective use.

Personalized image aesthetics assessment (PIAA) is an important research problem with practical real-world applications. While methods based on vision-language models (VLMs) are promising candidates for PIAA, it remains unclear whether they internally encode rich, multi-level aesthetic attributes required for effective personalization. In this paper, we first analyze the internal representations of VLMs to examine the presence and distribution of such aesthetic attributes, and then leverage them for lightweight, individual-level personalization without model fine-tuning. Our analysis reveals that VLMs encode diverse aesthetic attributes that propagate into the language decoder layers. Building on these representations, we demonstrate that simple linear models can perform PIAA effectively. We further analyze how aesthetic information is transferred across layers in different VLM architectures and across image domains. Our findings provide insights into how VLMs can be utilized for modeling subjective, individual aesthetic preferences. Our code is available at https://github.com/ynklab/vlm-latent-piaa.

Vision--language models (VLMs) have achieved impressive performance on complex multimodal reasoning tasks, yet they still fail on simple grounding skills such as object counting. Existing evaluations mostly assess only final outputs, offering limited insight into where these failures arise inside the model. In this work, we present an empirical study of VLM counting behavior through both behavioral and mechanistic analysis. We introduce COUNTINGTRICKS, a controlled evaluation suite of simple shape-based counting cases designed to expose vulnerabilities under different patchification layouts and adversarial prompting conditions. Using attention analysis and component-wise probing, we show that count-relevant visual evidence is strongest in the modality projection stage but degrades substantially in later language layers, where models become more susceptible to text priors. Motivated by this finding, we further evaluate Modality Attention Share (MAS), a lightweight intervention that encourages a minimum budget of visual attention during answer generation. Our results suggest that counting failures in VLMs stem not only from visual perception limits, but also from the underuse of visual evidence during language-stage reasoning. Code and dataset will be released at https://github.com/leduy99/-CVPRW26-Modality-Attention-Share.

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.

With the widespread deployment of vision-language large models (VLLMs), their safety alignment faces dual challenges across languages and modalities. Existing methods model multilingual and multimodal safety separately, overlooking coupled risks between low-resource-language instructions and visual contexts, which hinders the detection of cross-lingual and cross-modal harmful intent and the formation of robust safety boundaries. To address this, we propose a neuron-level interpretable safety alignment framework that identifies safety neurons and performs neuron-targeted safety tuning to jointly mitigate multilingual and multimodal risks. Specifically, we compare FFN representations elicited by harmful requests and benign inputs to identify neuron activation strengths associated with safety refusals. Next, we jointly model neuron activations and corresponding down-projection columns to derive neuron-level saliency, separating general multilingual and multimodal neurons from safety neurons responsible for model defense. Finally, neuron-targeted gradient masking restricts parameter updates to the safety subspace spanned by the identified neurons, enabling precise and interpretable safety enhancement. Extensive experiments show that our method enhances multilingual and multimodal safety by tuning only a few safety neurons, while preserving general capabilities.

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.

Accurate and interpretable plant disease diagnosis remains a major challenge for vision-language models (VLMs) in real-world agriculture. We introduce AgriChain, a dataset of approximately 11,000 expert-curated leaf images spanning diverse crops and pathologies, each paired with (i) a disease label, (ii) a calibrated confidence score (High/Medium/Low), and (iii) an expert-verified chain-of-thought (CoT) rationale. Draft explanations were first generated by GPT-4o and then verified by a professional agricultural engineer using standardized descriptors (e.g., lesion color, margin, and distribution). We fine-tune Qwen2.5-VL-3B on AgriChain, resulting in a specialized model termed AgriChain-VL3B, to jointly predict diseases and generate visually grounded reasoning. On a 1,000-image test set, our CoT-supervised model achieves 73.1% top-1 accuracy (macro F1 = 0.466; weighted F1 = 0.655), outperforming strong baselines including Gemini 1.5 Flash, Gemini 2.5 Pro, and GPT-4o Mini. The generated explanations align closely with expert reasoning, consistently referencing key visual cues. These findings demonstrate that expert-verified reasoning supervision significantly enhances both accuracy and interpretability, bridging the gap between generic multimodal models and human expertise, and advancing trustworthy, globally deployable AI for sustainable agriculture. The dataset and code are publicly available at: https://github.com/hazzanabeel12-netizen/agrichain

Large-scale vision-language models (VLMs) exhibit remarkable zero-shot capabilities, yet the internal mechanisms driving their anomaly detection (AD) performance remain poorly understood. Current methods predominantly treat VLMs as black-box feature extractors, assuming that anomaly-specific knowledge must be acquired through external adapters or memory banks. In this paper, we challenge this assumption by arguing that anomaly knowledge is intrinsically embedded within pre-trained models but remains latent and under-activated. We hypothesize that this knowledge is concentrated within a sparse subset of anomaly-sensitive neurons. To validate this, we propose latent anomaly knowledge excavation (LAKE), a training-free framework that identifies and elicits these critical neuronal signals using only a minimal set of normal samples. By isolating these sensitive neurons, LAKE constructs a highly compact normality representation that integrates visual structural deviations with cross-modal semantic activations. Extensive experiments on industrial AD benchmarks demonstrate that LAKE achieves state-of-the-art performance while providing intrinsic, neuron-level interpretability. Ultimately, our work advocates for a paradigm shift: redefining anomaly detection as the targeted activation of latent pre-trained knowledge rather than the acquisition of a downstream task.

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.

While Vision-Language Models (VLMs) have demonstrated significant potential in chemical visual understanding, current models are predominantly optimized for direct visual question-answering tasks. This paradigm often results in "black-box" systems that fail to utilize the inherent capability of Large Language Models (LLMs) to infer underlying reaction mechanisms. In this work, we introduce ChemVLR, a chemical VLM designed to prioritize reasoning within the perception process. Unlike conventional chemical VLMs, ChemVLR analyzes visual inputs in a fine-grained manner by explicitly identifying granular chemical descriptors, such as functional groups, prior to generating answers. This approach ensures the production of explicit and interpretable reasoning paths for complex visual chemical problems. To facilitate this methodology, we implement a cross-modality reverse-engineering strategy, combined with a rigorous filtering pipeline, to curate a large-scale reasoning-and-captioning dataset comprising 760k high-quality samples across molecular and reaction tasks. Furthermore, we adopt a three-stage training framework that systemically builds model perception and reasoning capacity. Experiments demonstrate that ChemVLR achieves state-of-the-art (SOTA) performance, surpassing both leading proprietary models and domain-specific open-source baselines. We also provide comprehensive ablation studies to validate our training strategy and data generation designs. Code and model weights will be available at https://github.com/xxlllz/ChemVLR.

Understanding when Vision-Language Models (VLMs) will behave unexpectedly, whether models can reliably predict their own behavior, and if models adhere to their introspective reasoning are central challenges for trustworthy deployment. To study this, we introduce the Graded Color Attribution (GCA) dataset, a controlled benchmark designed to elicit decision rules and evaluate participant faithfulness to these rules. GCA consists of line drawings that vary pixel-level color coverage across three conditions: world-knowledge recolorings, counterfactual recolorings, and shapes with no color priors. Using GCA, we ask both VLMs and human participants to state a threshold rule: the share of an object's pixels that must be a given color for the object to receive that color label. We then compare these rules with their subsequent color attribution decisions. Our findings reveal that models systematically violate their own introspective rules. For example, GPT-5-mini violates its stated introspection rules in nearly 60% of cases on objects with strong color priors. Human participants remain faithful to their stated rules, with any apparent violations being explained by a well-documented tendency to overestimate color coverage. In contrast, we find that VLMs can accurately estimate color coverage, yet directly contradict their own reasoning in their final responses. Across all models and strategies for eliciting introspective rules, world-knowledge priors systematically degrade faithfulness in ways that do not mirror human cognition. Our findings challenge the view that VLM reasoning failures are difficulty-driven and suggest that VLM introspective self-knowledge is miscalibrated, with direct implications for high-stakes deployment.