VLM Interpretability

Vision-language models (VLMs) show great potential for damage assessment after a disaster, but a recurring deficiency is that they are reluctant to declare a hazard; that is, recall is low even when overall accuracy appears adequate. This study examines that deficiency by using signal detection theory to decompose the decision behavior into perceptual capability and decision-criterion placement. We then propose a novel method for correcting the over-conservative decision policy, inspired by the finding that fear makes humans risk-averse, and ask whether an affective representation associated with fear can be causally manipulated to similarly alter a VLM's decision tendency. Using mechanistic interpretability, we localize a causally implicated affective circuit in the model and use activation steering to manipulate it while observing the effect on downstream prediction. The method is tested on a two-stage SeisMLLM pipeline built on Qwen2.5-VL-7B-Instruct, which flags only 27.0% of genuinely unsafe buildings on the SeisMLLM-1K test split and never issues a false Red, an SDT criterion of c = +1.354, despite adequate evidence quality (d' = 1.521). An affective direction is localized on emotion-rich natural scenes, causally validated by sparse-neuron knockout and distributed steering on held-out emotion data, and then injected into the building task. Fear-direction injection raises Red recall to 75.7% (p<0.001), and subtracting the same direction suppresses Red predictions entirely, whereas norm-matched random and matched happiness directions show no significant effect. The mechanism is a shift in criterion (c=-1.515) while discrimination is not improved (d'=-0.493). These results show that VLM decisions can be adjusted at inference time without retraining and demonstrate how mechanistic interpretability can be used to diagnose and control VLM decision behaviors in engineering applications.

Vision-language models increasingly reason through crops, regions, and tool-produced observations. Yet an observation can influence the answer without benefiting the candidate it supports. We study candidate-bound visual contribution: valid evidence should help, invalidating its supporting relation should remove its additional effect, and valid rebinding should redirect that effect to the newly supported candidate. We introduce CROSS-Bench, a benchmark of 28,000 decision problems, with matched invalidation and rebinding tests on a dedicated evaluation subset. Our RIVET interface preserves evidence identity and uncertainty, composes a candidate-conditioned response, and separately controls its strength. Shared-evidence experiments show that task accuracy and evidence ownership can diverge. Under matched capacity and training, RIVET increases normalized effect transfer from 0.512 to 0.651 where clean evidence has a positive effect. The advantage persists on common evaluation examples and across repeated decision-layer fits. With evidence predicted from raw inputs, RIVET improves CROSS-Bench accuracy by an average of 5.70 pp across four frozen backbones, relative to the same models without auxiliary evidence. These results separate the utility of visual evidence from the candidate-specific destination of its effect.

As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at GitHub repository.

Current vision-language models (VLMs) encode visual information in dense hidden states where object identity, spatial layout, and local attributes are implicitly entangled rather than explicitly disentangled, limiting their ability to isolate and modulate the specific visual evidence required by a given language query. Inspired by sparse population coding and top-down modulation in biological vision, we introduce NeuronEye, a plug-in framework that constructs a sparse, concept-level neuron vocabulary from intermediate VLM representations and selectively activates query-relevant visual concepts during inference. NeuronEye decomposes vision-token states into an overcomplete sparse basis organized by concept-level clusters, uses the language query to activate relevant clusters and localize the patches where selected concepts are expressed, and injects the focused evidence back into vision tokens. A complementary suppression mechanism attenuates dominant perceptual directions to preserve weaker but relevant cues. All operations run in a single forward pass over a frozen VLM backbone. On Qwen2.5-VL-7B, NeuronEye raises CV-Bench overall accuracy by +3.1 with gains of +9.5 on Distance, and improves BLINK Multi-view by +8.3, with similar trends on LLaVA-1.6-7B. These results suggest that sparse neuron vocabularies can serve not only as post-hoc interpretability tools but also as active interfaces for concept-level visual reasoning.

Concept Bottleneck Models (CBMs) built on vision-language models such as CLIP represent a latent space as human-understandable concepts. These representations are unfaithful: related concepts are entangled, so individual scores do not reflect their intended meaning. We propose CHOQOLATE, an interpretable-by-design layer based on 2-additive Choquet integrals, which merges correlated concepts into compact nodes. Across four datasets, CHOQOLATE achieves a favorable accuracy-interpretability trade-off, with weight-sparse and semantically coherent nodes. A closed-form gradient derivation, backed by experiments, explains why Choquet layers drive this organization without explicit supervision. Choquet weights also map directly to Shapley values, which enables test-time intervention. On standard bias-mitigation benchmarks, suppressing spurious concepts after training performs on par with methods that require group annotations or retraining, while needing neither.

There has been significant work on understanding the In-Context Learning capabilities of Large Language Models, especially on the induction circuit. For a few-shot classification task, the induction circuit leverages linear representations of each labeled example in-context in order to classify an unlabeled query. However, few works focus on how those linear representations are built in the first place. Leveraging the expressivity of the vision modality compared to text, we uncover a Shared Discriminative Geometry (SDG) inside Large Vision Language Models (LVLMs). It is a low-dimensional space, shared across all image classification tasks, in which in-context images are compressed into linearly separable representations later used to perform classification. We observe that this is the result of the model performing a dimensionality reduction of vision representations in early layers. In order to explain this phenomenon: (1) We show analytically that linear self-attention can perform a dimensionality reduction by projecting in-context data onto its principal components, with each layer implementing one gradient descent step toward this objective. (2) We provide evidence that trained LVLMs reduce the dimensionality of vision representations in early layers via a similar mechanism.

Large vision-language models (LVLMs) exhibit strong reasoning capabilities, yet the visual and textual evidence supporting the generated responses remains difficult to identify. Faithful token attribution explains an LVLM's response by assigning scores that rank image and prompt tokens by how much the model relies on them, such that removing higher-ranked tokens causes the likelihood of the generated response to drop more rapidly. However, existing token-attribution methods have been developed mainly for text-based language models, and our empirical study reveals two challenges when complex multimodal sources are involved. First, the joint image-text attribution can underrepresent visual evidence relative to text, obscuring the image regions supporting the response. Second, visual evidence may influence the generated response through multiple intermediate reasoning paths, while existing methods trace only a limited subset of these paths, causing important visual contributions to be underestimated. Motivated by these insights, we introduce VTrace, a multimodal token-attribution framework that traces input contributions through intermediate reasoning and calibrates attribution scores across modalities. VTrace constructs pairwise attributions that highlight token-specific contributions and aggregates all forward attribution paths in closed form to account for both direct and indirect contributions. Cross-modal calibration then rescales image and text attribution scores using modality contributions estimated from response-likelihood changes, enabling a unified ranking of input tokens. Evaluations against seven baselines across six visual reasoning benchmarks demonstrate the superior attribution faithfulness. Project page: https://vtrace-attribution.github.io/.

Current explanation methods for contrastive vision-language models such as CLIP mainly identify important regions without showing how to change the input in order to get a target prediction. We introduce Mask-guided Adaptive Counterfactual Explanations (MACE), a targeted visual counterfactual method designed specifically for CLIP zero-shot classification. MACE constructs an editable region from either source attribution or source-target attribution differences and expands the mask only when needed to reach a specified target class. A latent diffusion inpainting model then modifies the selected region, while a frozen CLIP model provides modification guidance and anchors the remaining image content to the original input. We evaluate MACE on ImageNet, Food-101, Oxford Pets, and CUB-200. The source-mask variant achieves the highest target top-1 success rate across all four datasets, while the difference-mask variant produces the smallest pixel level and perceptual changes and the best realism scores. Both variants improve proximity and realism over a Stable Diffusion-only baseline using the same generative backbone. These results show that adaptive mask-guided editing produces effective CLIP counterfactuals. They further reveal a tradeoff between counterfactual validity and source-image preservation.

Vision-Language Models are increasingly used in autonomous-driving systems, yet their ability to recover dynamic physical state from visual input remains insufficiently characterized. We study velocity understanding as a controlled diagnostic across three tasks: surrounding-agent speed, current ego speed, and short-horizon future ego-speed proposal. On nuScenes, we evaluate open-weight general-purpose and PhysicalAI VLMs, together with the driving-oriented Alpamayo-1.5 Vision-Language-Action model, using multiple input and output formulations. We combine verbal evaluation with temporal perturbations, counterfactual ego-speed hints and linear probes of hidden representations. The tasks exhibit distinct failure modes. Surrounding-agent speed is weakly encoded in an agent-specific form, whereas current ego speed is often internally accessible but poorly verbalized: continuous probes achieve 4.7-5.8 km/h MAE compared with 10.2-16.8 km/h MAE for verbal outputs. Multiple frames provide inconsistent verbal gains to single frame inputs, and frame order is rarely exploited. Under non-optimized QLoRA, task-specific adaptation improves both task-relevant latent speed representations and verbal readout, but continuous surrounding-agent speed estimation remains weak, while most future-speed gains survive frame shuffling, indicating limited temporal grounding. Driving specialized Alpamayo-1.5 shows stronger latent representations for surrounding-agent and future ego speed, while current ego-speed decodability is comparable and substantial probe-verbal gaps remain. Thus, driving specialization can strengthen motion representations but does not guarantee stronger encoding across both scene and ego states or reliable readout. The results show that plausible planning outputs do not necessarily imply reliable recovery or temporal grounding of the underlying dynamic state.

Medical Vision-Language Models (VLMs) show significant promise for clinical image understanding, offering accurate diagnosis with interpretable reasoning. However, a critical performance gap exists between their strong vision encoders and the full multimodal model: in dermatology, the MedSigLIP encoder outperforms MedGemma by an average of 10.26 percentage points even when both use zero target-task labels; few-shot linear probing provides further evidence of strong visual representations. This gap motivates an investigation of how visual information is used in end-to-end diagnosis and why plausible-sounding predictions can lack grounding in image evidence. Using dermatology as our primary testbed, we systematically investigate three hypotheses for this phenomenon. We further provide a mechanistic analysis of the model's internal attention patterns, showing that a simple describe-then-decide prompting strategy increases vision attention by 30-40% during generation. Task-specific fine-tuning improves dermatology classification but reduces cross-domain medical question-answering performance in our evaluation. To address these challenges, we combine label-free prompting with low-label encoder-assisted reranking while keeping the VLM frozen. We validate the interventions across five VLM backbones in dermatology and provide supporting representation and attention analyses across additional medical modalities.

High instance-level accuracy can mask inconsistencies in spatial reasoning when objects exchange positions or their roles are reversed in the query. The internal representations supporting relative-position reasoning remain poorly understood. We investigate two complementary components of this process: tracking object locations in the input and representing their query roles. Across three VLMs with visual or textual inputs and their language-model backbones, activation patching reveals a staged progression from early-layer source representations through intermediate-layer query-object representations to late-layer answer states. Targeted interventions further establish causal links along this progression: manipulating source-side representations shifts location information at query-object mentions and ultimately alters relation predictions. Beyond object-location information, we also identify a stable query-side direction associated with the roles of the two objects in the comparison. Steering along directions estimated on synthetic scenes generalizes to natural-image benchmarks, improving accuracy and both forms of paired consistency in most settings without retraining. Our findings reveal complementary components of relational reasoning across visual and textual settings and show how targeted interventions can improve the consistency of models' behavior.

Despite substantial progress across downstream applications, vision-language models (VLMs) remain susceptible to language bias, often prioritizing linguistic cues over visual evidence and consequently producing incorrect predictions. Prior studies have proposed various approaches to understanding and mitigating language bias in VLMs, yet their findings often conflict due to the difficulty of tracing how language bias propagates within black-box VLMs. Building on the word completion task, we trace how language bias propagates through VLM inference by (1) proposing a diagnostic framework that decomposes the inference process into four distinct yet interdependent stages to trace the propagation of language bias; and (2) examining how two key factors underlying language bias, i.e., linguistic priors and cross-modal coverage, evolve across these stages and ultimately give rise to incorrect predictions. The linguistic prior captures the strength of statistical bias induced by the language model component of a VLM and represents the origin of language bias, whereas cross-modal coverage measures the extent to which linguistic cues cover the visual content. By decomposing inference into four stages and characterizing the interplay between linguistic priors and cross-modal coverage across these stages, we propose a systematic framework for tracing the propagation of language bias throughout the inference process; and uncover the underlying mechanism of language bias by revealing the interplay between linguistic priors and cross-modal coverage.

When a VLM answers a visual query, current interpretability tools rely on text rationales, which use a mismatched modality, or on internal read-outs, which originate too early to reflect the final output and require white-box access to the model. We introduce AnswerMap, a training-free, task-agnostic, black-box visual rationale constructed from the output head. The image is cut into K row and K column bands, each shown alone to the frozen model along with the query in the format of a yes/no relevance question. The outer product of the row and column ``yes'' posteriors gives the query-conditioned spatial map. Crucially, by defining a fixed read-out R (e.g., expectation, maximum) on top of AnswerMap, we can derive continuous outputs like location natively. This bypasses the reliance on discrete text tokens for continuous-output tasks and guarantees an image-dependent answer by construction. However, a rationale can be confabulated, so we validate AnswerMap across four models and three query distributions with two tests: (a) agreement with the model's own generated point and (b) deletion of the map's region. The map lands where the model points (AUC 0.85 against 0.38 for attention), and deleting its region flips 53% of correct answers (against 19% for attention's). Beyond establishing faithfulness, we demonstrate the map's task-agnostic utility through three distinct read-outs: its maximum flags hallucinated objects without generation, its expectation localizes correctly when the model's own pointing fails, and its top-mass region, fed back as a crop, fixes half of the model's wrong answers. AnswerMap thus offers a new lens on VLM interpretability and, through its read-outs, a new output interface for visual tasks beyond text tokens.

Vision-language models (VLMs) can answer simple visual questions, but often struggle when one question requires several visual judgments. We study this gap with controlled tasks for feature binding, numerosity, spatial relations, and amodal completion, together with a Composite task that combines them. Matched counterfactual image pairs isolate changes in the visual evidence needed to answer. Across four models, direct answers, hidden-state readouts, and state interventions show that the individual judgments can be made without explicit reasoning and that intervening on the corresponding states can affect the answer. During reasoning, the Composite answer becomes decodable from hidden states and usable from shortened traces, often before the model stops on its own. We train a small detector to predict this readiness and stop reasoning at that point. On MMStar and RealWorldQA, this reduces mean reasoning tokens by 79.1% and 74.5%, while average accuracy rises by 3.13 and 3.30 percentage points, respectively. These findings connect the internal development of answer readiness to a practical rule for allocating reasoning computation.

Recent studies suggest that large language models encode emotion concepts as structured internal representations, but most existing work focuses on text and a single architecture. Therefore, we ask, do emotion concepts generalize across sources, modalities, and architectures in vision--language models (VLMs)? To address this, we construct CMES (Cross-Modal Emotion Stimuli), a multi-source collection of emotion-conditioned stories, real facial expressions, synthetic portraits, and synthetic emotion-evoking scenes. For each stimulus source, we extract a separate set of six Ekman emotion vectors from each of three VLMs. We report four main findings as follows: 1) Image-derived emotion vectors form a low-dimensional geometry similar to that of text-derived vectors. Valence is relatively stable across sources, while arousal varies more. 2) Text- and image-derived emotion vectors have modest cosine similarity but still show held-out cross-modal correspondence. Text-derived vectors can also steer image interpretation. 3) Cross-architecture correspondence remains even when native cosine is near zero. Transformations estimated from generic ImageNet activations recover both correspondence and causal transfer without using the six emotion vectors or their labels. 4) After aligning representations across architectures, we construct a shared emotion subspace that preserves affective geometry and selective steering effects. The corresponding consensus emotion vectors also generalize to a held-out fourth architecture at two model sizes. These results suggest that emotion representations can share relational structure and causal effects across sources, modalities, and architectures, even when individual vector directions differ.

Layer-wise visual-text similarity in Multimodal Large Language Models (MLLMs) is widely interpreted as evidence that the language model progressively integrates visual content into a shared representation space. This reading rests on the assumption that scalar alignment scores reflect content-level cross-modal interaction. To test this assumption, we apply controlled interventions to the visual stream. Across 13 MLLMs from five families spanning 0.5B to 72B parameters, replacing projector-output visual tokens with Gaussian noise sharply reduces task accuracy, yet four standard scalar measures (CKA, SVCCA, MIR, and the leading principal-angle cosine) fail to consistently separate the corrupted stream from the original. We call this failure the alignment illusion and trace it to the shared language-model pathway: anisotropic MLP down-projections pull visual and text tokens toward common output directions, producing weight-induced alignment. Because this component is essentially one-dimensional, we introduce the principal-angle gap (PA gap), defined as the difference between the top two principal-angle cosines, which separates weight-induced similarity from multi-directional visual structure. Under graded visual corruption, the PA gap tracks task accuracy more consistently than the scalar scores we consider; under a structured but irrelevant image, it further exposes regimes in which internal geometry and task accuracy come apart. Internal visual-text alignment in MLLMs is therefore best read as a geometric diagnostic of the visual stream inside the language model rather than a direct proxy for content-level cross-modal interaction, and is most informative when calibrated by controlled task evidence.

Unified vision-language models (VLMs) that tokenize images through a vector-quantized (VQ) codebook routinely hallucinate objects on grounded yes/no benchmarks, yet existing decoding-time fixes treat this as generic miscalibration without an architectural account. Using activation patching across twenty-five models spanning eight LLM families, we identify an early-layer ($L_0$) attention routing circuit shared across VQ-tokenized VLMs and propose a three-gate diagnostic that distinguishes the models carrying it from those that do not. The diagnostic isolates ten positive models (five natural unified-VQ VLMs across three LLM families and five induced variants) and rejects the remaining fifteen. A single-variable architectural swap (LLaVA-1.6 CLIP+MLP $\rightarrow$ VQ+Linear) installs the circuit, while a matched-compute MLP control on identical data does not, isolating vector quantization as the source of the pathological signal; the routing pathway that carries it is one that the backbone already provides. Against tuned VCD and DoLA baselines, tuned DoLA wins on binary calibration, but \textbf{only $L_0$ ablation reduces object hallucination in open-ended generation} (CHAIR$_i$ reduces by $31\,\%$ relatively, whereas tuned DoLA and VCD leave it unchanged or worsen it). These results recast object hallucination in unified VQ VLMs as a property of architecture and pretraining, and yield a targeted intervention that mechanism-agnostic decoding cannot replicate.

Modern Vision-Language Navigation (VLN) models rely mostly on pre-trained large Vision-Language Models (VLMs) to predict navigation actions. While this fusion of language instructions and visual observations allows multimodal reasoning, it obscures how information is routed across modalities or what mechanisms drive navigation decisions. Thus, it remains unclear whether VLN models ground their predictions in relevant semantic cues or can track task progress. In this work, we study the interpretability and steerability of VLN models. We use intervention-based metrics that measure how visual observations, instructions, and visual memory causally influence navigation decisions. Our results show that these navigation policies are sensitive to all input modalities and do not depend on a single one. We further show that these agents encode navigation progress and retain semantic structure from their VLM backbones, enabling concept-level steering through internal activations. Finally, we extract activation vectors for abstract behaviors to transfer them zero-shot to out-of-distribution real-world scenarios, improving performance without additional fine-tuning.

When large vision-language models misclassify harmful memes, the failure may reflect missing internal evidence or an inability to route represented evidence to their outputs. We distinguish these cases in Gemma-3 and Qwen3.5 using sparse autoencoders, role-conditioned probes, causal interventions, and recovery experiments across six harmful content benchmarks, with additional Spanish and Hindi-English code-mixed evaluations. Sparse readouts outperform native prediction on all six primary binary tasks: Qwen averages $0.740$ versus $0.432$ for native macro-F1, residual reconstruction reaches $0.486$, and Gemma improves from $0.532$ to $0.714$. These gains measure how accessible the label is to a supervised readout; they do not show that the model's native generation already applies such a decision rule. Under the evaluated scales, Qwen silent-feature ablation is $24-63$ times more probe-sensitive, whereas routed-feature patching on literal yes/no tasks is $16-140$ times more output-sensitive. Native-only threshold calibration explains much, but not all of the gap: on five tasks with matched probe scores, it recovers $69.8$% of the raw native-to-probe difference, while direct routing adds $0.094$ mean macro-F1 beyond calibrated native scoring. Joint gold-label, probe-KL, and pairwise LoRA supervision improves dedicated FHM prediction, but a gold-only adapter performs better on the shared seven-task mean. A case study of Gemma-3-12B on the Facebook Hateful Memes dataset finds a distributed rank-32 image-prompt interaction, reaching $0.756$ versus $0.685$ native macro-F1. Robustness controls show that the signal is not explained solely by accompanying OCR and depends on paired visual evidence, and that it extends beyond English. In many of the errors we study, the evidence is represented but does not reach the answer; therefore, routing is a common bottleneck in harmful meme classification.

How do VLMs map from pixels to semantics? To understand this general question, we focus on a narrow one: studying how VLMs perform optical character recognition (OCR). Across four models, we identify attention heads causally necessary for OCR, and discover that these are in fact general-purpose heads that output interpretable semantic features across all image tokens. For example, pointing these heads at an image token containing the word "bike" causes Qwen3-VL-8B to output "bike," but pointing them at a bird wing causes the model to output the token "feathers." We collapse these heads' attention weights into a single verbalization lens transformation that reveals interpretable semantic features in hidden states across all layers. When combined with projection to vocabulary space, we can obtain interpretable labels starting from layer 0, showing that image representations are in fact aligned with language in early layers. We find that we can also use the inverse of this transformation to edit non-word concepts, e.g., replacing a tractor with a revolver in a naturalistic image, providing causal evidence that this subspace is useful for more than just OCR. Our results are an example of how the study of specific mechanisms can shed light on broader interpretability problems.

Vision-Language Models (VLMs) can produce Natural Language Explanations (NLEs) that sound plausible yet remain inconsistent with the visual evidence they cite. We present Explanation-Driven Counterfactual Testing (EDCT), an intervention-based protocol that extracts visual concepts cited in a model's explanation, applies verified minimal edits to them, and tests whether the resulting answer and explanation remain consistent with the edited image. Using this protocol, we create EDCT-Bench, a comprehensive benchmark spanning three complementary domains: knowledge-intensive visual question answering (OK-VQA), safety-critical driving (DriveLM), and 3D spatial reasoning (3DSRBench). Across the evaluated VLMs, EDCT reveals substantial faithfulness gaps, with models frequently producing responses inconsistent with verified visual changes. Finally, our fine-tuning study suggests that EDCT-generated counterfactuals provide high-impact training signals.

Vision Transformers (ViTs) are central to most modern vision models, yet obtaining input attributions that are fine-grained, faithful, and stable remains challenging. Layer-wise Relevance Propagation (LRP) has been adapted to transformer attention, but in ViTs it often produces noisy, unfaithful explanations. We show that the missing ingredient is the treatment of residual connections: cancellation effects in residual pathways lead to attribution explosion. Moreover, we find that these cancellations are substantially stronger in ViTs than in language transformers. To address this issue, we introduce Residual-aware Layer-wise Relevance Propagation (ResLRP), a simple extension of LRP whose propagation rules explicitly account for cancellations in residual branches, are exactly conservative, and provably bound relevance explosion. Causal channel-wise interventions confirm that residual cancellation, not a generic regularization effect, drives the instability. ResLRP substantially improves attribution quality across faithfulness and localization, evaluated on ViT architectures spanning supervised, self-supervised, contrastive, hierarchical, and multimodal families, as well as on the ground-truth-controlled FunnyBirds benchmark. The largest gains arise in modern Vision Language Models (VLMs), with +27-29% localization and up to 3.4x faithfulness scores. Beyond benchmarks, ResLRP localizes Sparse Autoencoder (SAE) features in input space, and our residual amplification measure serves as an architecture-level diagnostic predicting where attribution degrades.

Query-conditioned vision--language models enable fine-grained interpretation by revealing how visual evidence changes with textual queries. However, evidence conditioned on complete descriptions does not necessarily resolve into object-specific evidence, nor does an exposed evidence map necessarily identify the evidence that constitutes the model's prediction. Across multiple VLM architectures and independent benchmarks, we find that object-level queries often retain evidence from co-occurring objects and shared context. In this paper, we introduce \textbf{ProtoLIP}, a lightweight prototype-mediated evidence layer that organizes reusable visual prototypes into text-derived semantic families and uses query-dependent family routing to constrain which prototypes may provide evidence. Without spatial annotations or backbone retraining, ProtoLIP improves evidence localization and separation across query granularities, with localization gains transferring to independently pretrained VLMs with well-aligned patch--text representations. Despite using only text-derived weak supervision, ProtoLIP remains competitive with a spatially supervised grounding model while maintaining strong matching and competitive image--text retrieval. Crucially, ProtoLIP constructs its matching score directly from localized prototype evidence, enabling the score to be exactly decomposed into semantic-family and prototype contributions.

Vision-language models retain a substantial amount of pixel-decodable visual content in their visual key-value cache. We show, in our setting, that this retention is task-inert: across our preregistered tests, how much a unit retains never positively tracks whether the computation that answers the question causally relies on it. We measure retention with a learned pixel-inversion decoder and causal use with single-super-patch KV ablation, the teacher-forced drop in gold-answer log-probability, and relate the two within images under a preregistered, sign-calibrated, held-out design. Retention is decoupled from attention and, in a well-powered null, from causal utilization. Utilization is not inert to every proxy: attention weakly but significantly tracks it, the only signal we find that does and the design's positive control. We characterize pixel-decodable retention as an informational axis of the visual KV cache, orthogonal to the functional one. How much task-inert content a cache holds differs by architecture in our model pair: the encoder-free model retains 2.7 times more than the encoder-based one. The engineering consequence is a controlled negative result. At super-patch granularity, deconfounded pixel-decodable retention ranks KV eviction no better than random; at token granularity it acquires only a weak inverse-importance signal at larger budgets, dominated at every budget by attention magnitude. In our setting, pixel-decodable reconstructability is not a competitive KV-compression signal at any granularity we test.

Visual encoders construct a representation of the image input for Vision-Language models. How much conceptual, as opposed to immediately visible, information does this representation contain? We use canonical color as a controlled test case to ask whether vision encoders make canonical-color information linearly accessible, even when color is removed from the input image. We construct a dataset of objects with canonical colors, and probe vision encoders for both color and object identity using color and grayscale images. We find that canonical color remains decodable from grayscale images, and is tied to predicted object identity, indicating a conceptual link. Extending this analysis to full VLMs, we find that VLM post-training can have a surprisingly large effect on color decodability in the vision encoder. Overall, canonical color provides a usefully controllable lens for tracing object-level conceptual semantic information in vision encoders and VLMs.

Vision-language models (VLMs) are increasingly used in settings where some input modalities may be unavailable, yet we know little about whether they can faithfully explain how such missing information affects their own predictions. We introduce an interventional protocol for evaluating self-explanations of modality dynamics: models state what each modality alone would support, whether restoring a missing modality would change their answer, and whether the available evidence is sufficient; we then execute the corresponding intervention and compare these claims with realized behavior. We evaluate ten VLMs spanning open-weight and proprietary models across four tasks covering mixed, redundant, and unique modality regimes. We find a systematic tendency to overstate the sufficiency of available evidence. Models substantially underestimate the effect of restoring missing modalities: executed change exceeds predicted change in 78 of 80 model-task-condition settings, with task-level median executed change rates reaching 70.1% while median predicted rates remain at most 9.6%. Insufficiency claims have low recall, leaving many cases in which behavior changes despite a stated claim of sufficiency. Retrospective self-explanations show the same tendency, over-crediting single-input sufficiency in mixed regimes and interchangeability in redundant ones. Together, these results show that VLMs systematically mischaracterize how their predictions depend on available and missing evidence, motivating executable interventions as a behavioral test of multimodal self-explanations.

Vision-language models (VLMs) are increasingly evaluated on complex image and video understanding tasks, yet conventional metrics primarily assess final-answer quality and reveal little about how different information sources shape the generation process. We propose a causal and temporal evaluation framework that traces the evolving roles of visual input, question text, and generated prefixes during autoregressive decoding. Grounded in a Structural Causal Model, we use interventions and backdoor adjustment to derive three step-indexed causal-drive metrics---Visual Causal Drive (VCD), Question Causal Drive (QCD), and Prefix Causal Drive (PCD)---for characterizing source-specific generation patterns without requiring reference answers. Experiments on Qwen3-VL-8B-Instruct across MAVIS, LLaVA-Video-178K, and MiraData, together with cross-model validation on InternVL2-8B, reveal a consistent transition from stronger early question and visual guidance toward increasing reliance on generated prefixes. Randomized-intervention validation shows that QCD and PCD reduce recovery error over observational PMI baselines by 34.8% and 47.1%, respectively. On VLMBias, the prefix--visual imbalance score achieves 0.767 AUROC and 0.873 AUPRC for distinguishing prior-driven from visually grounded generations. These results show that causal-drive trajectories provide complementary source-level diagnostics for multimodal generation.

Whether attention weights faithfully reflect model reasoning has been actively debated in NLP, yet this question remains largely unexplored for the visual modality in Vision-Language Models (VLMs). We address this gap through causal perturbation analysis on current VLMs, evaluating both the comprehensiveness and sufficiency gap of attention-ranked visual tokens. Our analysis reveals that visual attention faithfulness is heterogeneous, manifesting in three distinct processing modes: Faithful-Sufficient, where top-$k$ attention tokens are both necessary and sufficient for prediction; Faithful-Distributed, where they are necessary but broader visual context remains required; and Non-Focal, where no localized attention region is individually necessary while visual information remains an essential trigger for prediction. Furthermore, human-annotated ground-truth regions satisfy comprehensiveness in only $\sim 60$% of cases compared with model attention rankings, revealing systematic divergence between model visual reliance and human intuition. We demonstrate these patterns across both general VQA on VQAv2 and document tasks on VRDU and ChartQA, showing that visual attention faithfulness varies systematically with processing demands and model architectures rather than being uniformly faithful or unfaithful.

We investigate how vision-language models (VLMs) handle context-memory conflicts; that is, situations in which the model is given information in context that differs from what was stored parametrically during training. We document asymmetric biases: models tend to prefer in-context information about entities which appear in text, but prefer parametric information about entities which appear in images. We relate this asymmetry to the late representational alignment across modalities, showing that the longer processing time associated with resolving visual entities prevents the suppression of the model's usual factual recall mechanism, thus resulting in more parametric answers. Chain-of-thought reasoning does not appear to resolve the gap, but increasing the amount of visual information in the context does show an effect. These results illustrate the complexity of ensuring consistent behavior as models become increasingly multimodal and retrieval-augmented.

Vision-language models (VLMs) can comply with harmful requests delivered through images, even when their LLM backbones would refuse the same content in text. While prior work characterizes these jailbreaks empirically or at the representation level, how visual inputs perturb safety pathways at the neuron level remains uncharted. We close this gap with a causal, neuron-level analysis of safety mechanisms in 10 VLMs. We propose a two-stage detection pipeline with iterative ablation that accounts for self-repair, and introduce two modality-isolated benchmarks, ViSafe-Detect and ViSafe-Eval, which decouple visual and textual safety signals. Our analysis reveals: (i) Text safety in VLMs is localizable: $\sim$88 neurons ($<$0.01%) whose targeted ablation substantially reduces refusal. (ii) Text safety neurons constitute the dominant refusal pathway: ablating them is the only intervention that consistently and substantially reduces refusal across all models. (iii) Visual safety is high-dimensional and diffuse at the single-neuron level: text safety concentrates in $\sim$5 subspace directions while visual safety requires $\geq$50. This gap holds across architectures, explaining why current alignment has not closed the visual safety gap. Project page is at: https://jiaxuan-li.github.io/vlm-safety-neuron/ Warning: this paper may include examples of harmful content.