VLM Interpretability

We study audio-visual conflict as a compositional generalization test for AV-LLMs: the model must combine synchronized but semantically incompatible audio and video evidence and decide whether the pair matches. On VideoLLaMA 2-7B-AV, three alignment configurations remain nearchance on the scored exact-string Yes/No subset of AVHBench, even though their output priors shift substantially. Similarly, off-the-shelf InternVideo2 experienced a 32.3% accuracy decrease specifically under cross-modal conflict, accompanied by a 17.3% instruction-following failure. We call this failure mode prior dominance: late-layer commitment to an internally preferred answer pattern that is weakly grounded in the conflicting inputs. To explain this behavior, we conduct a mechanistic interpretability analysis and find that commitment remains concentrated at 25.5 $\pm$ 1 layers. We show that stronger temporal alignment changes answer bias, but do not improve compositional conflict resolution. Code and data to reproduce our mechanistic audit and behavioral evaluations are available at https://github.com/AdarshSudheer09/AVHBench-dmai.

When visual evidence is occluded or chaotic, models should abstain. In this paper, we show that Vision-Language Models (VLMs) can internally distinguish when abstention is required, but fail to express it anyway. We introduce TRAPSBench, a procedurally generated video benchmark of 1,404 matched physics pairs in which a single targeted change renders the outcome undeterminable from the visual evidence. Furthermore, we introduce Penalized Epistemic Calibration Score (PECS), a new robust metric that requires models to both answer correctly when the outcome is knowable, and abstain when the outcome is not. Across 16 VLMs spanning five families, spontaneous restraint is poor: the best PECS is 0.292. The bottleneck is expression, not perception: linear probes decode answerability from hidden states at up to 0.91 AUROC across physics domains; steering a single-layer void direction causally induces or suppresses abstention. Our results replicate across three open-weight families (Qwen, Gemma, LLaVA). The failure is also more pronounced in visual than textual uncertainty: models detect textual impossibility about 4x more readily than missing visual evidence. Closing this representation--output gap likely requires output-stage interventions.

Attribute hallucination---where vision-language models (VLMs) correctly identify an object but mischaracterize its properties---is prevalent yet mechanistically poorly understood. The dominant explanation, language-prior dominance, has motivated prior-suppression methods, but this explanation has not been directly tested at the attribute level. We present VISOR (Visual-Operational Remediation), a unified framework that couples null-image-based diagnosis with routed remediation. Its VSNR diagnostic decomposes each prediction into a visual logit signal and a language-prior signal. Across 10,791 negative-ground-truth samples from three VLM families and three attribute types, the visual signal strongly predicts false positives, whereas the language-prior signal is near chance. VISOR uses this diagnosis to separate two failure modes: low-margin but directionally correct visual signals in color/state attributes, and low-SNR or misaligned visual signals in material attributes. The same diagnosis routes each query to the appropriate operator: calibration for threshold-placement errors, abstention for training-free low-SNR handling, or targeted visual adaptation for material failures that prior suppression cannot correct. Across Qwen, InternVL, and LLaVA, VISOR reduces attribute false positives without relying on the prior-dominance assumption.

Vision-language models can describe an image with remarkable accuracy, yet a more fundamental question remains unanswered: what visual information actually drives their answers? In this work, we investigate this question through causal tracing, and we observe that highly causal vision tokens often lie outside the target region. Extending the analysis to larger vision-language models reveals a similar pattern across models and corruption settings, suggesting that strong multimodal performance does not necessarily imply spatially localized causal representations. We further investigate: can these models preserve visual structure when appearance cues are removed? and find that visual cues are exploited to understand visual structures. Together, our experiments expose a gap between seeing, using, and reasoning over visual structure, and provide a causal framework for studying how visual information is transformed, preserved, and ultimately used by modern vision-language models.

How much of a vision-language model's (VLM) spatial understanding remains after the action post-training process of building a vision-language-action model (VLA)? We probe depth perception, a primitive of spatiogeometric understanding, from every decoder layer of a weight-matched open-source base VLM/VLA pair: Molmo2-ER and MolmoAct2-LIBERO. First, the VLA decodes depth worse at every layer, a persistent gap we call the floor. Second, the degradation is not uniform: while the base VLM's depth decodability improves through its final layers, the VLA's collapses, an additional late-layer drop we call the cliff. We causally localize the cliff to late-layer MLP interference: ablating the late-layer MLP writes recovers the majority of the terminal decodability cliff, while matched attention ablations and the same intervention in the weight-matched base VLM produce no comparable recovery. A module-level decomposition explains this dissociation: the base VLM carries depth most accessibly in accumulated MLP writes, whereas action post-training collapses depth decodability in the late accumulated writes.

Vision-language models (VLMs) process an image as a sequence of visual tokens, which creates a substantial computational bottleneck during inference. Recent visual token pruning methods address this issue by removing seemingly redundant tokens, yet it remains unclear how these pruning decisions relate to the functional roles of visual tokens. In this work, we analyze visual token pruning through the lens of token roles identified by EmbedLens. We first show that representative pruning methods exhibit distinct token-role biases, but these biases do not directly correlate with downstream performance. To better understand this behavior, we refine the token-role assignment procedure and evaluate role-protected pruning variants. Our results show that preserving non-alive tokens can sometimes maintain or improve performance, suggesting that tokens with weak direct semantic alignment may still affect model behavior under pruning. Our code is publicly available at https://github.com/jaykim9870/Not_All_Redundant_Tokens_Are_Alike.

In human visual perception, uppercase lettering serves as a natural salience cue that captures attention within lowercase text. In this paper, we present a systematic empirical characterization study revealing that Large Language Models (LLMs) exhibit an analogous property: letter casing modulates internal attention allocation. Through analysis across 13 models, nine LLMs and four Vision-Language Models (VLMs), with diverse tokenization schemes, we show that formatting target information in alternating or uppercase against a lowercase context concentrates attention on those textual spans. In text this effect is universal, holding across every evaluated non-reasoning model. We frame it as a previously under-explored latent property of pretrained transformers rather than a prescriptive method. Our investigation reveals a central attention-performance divergence: while this "casing effect" robustly shifts attention, its impact on downstream accuracy is non-trivial, increased concentration does not inherently improve task accuracy and, in high-entropy contexts like alternating case, can degrade it. We further identify a boundary condition: the deliberative "thinking" phase in reasoning models acts as a semantic buffer that mitigates typographic sensitivity in text. Extending the study to VLMs, we find the effect transfers partially: the same prompt-side casing reorganizes cross-modal attention along two coupled axes, predominantly a macroscopic disengagement from the image toward the text prompt, and secondarily a concentration of the residual visual attention on the target region. By isolating casing as a zero-shot mechanism for attention steering that requires no model access or fine-tuning, we provide a new foundational understanding of how pretraining internalizes typographic emphasis.

Urban flooding poses an escalating threat to transportation infrastructure, yet no operational system provides real-time, street-level flood-depth estimates at centimeter resolution. This paper presents three vision-language models fine-tuned for continuous flood-depth estimation from street-level imagery: FloodLlama-Dense, a fully fine-tuned QLoRA baseline, and FloodLlama-MI5 and FloodLlama-MI6, interpretability-guided sparse variants that fine-tune only the top five and six causally relevant cross-attention layers identified through mechanistic interpretability analysis, respectively. Training uses an approximately 610,000-image subset of a 2.81-million-image synthetic corpus generated in Unreal Engine 5. The dataset combines single-vehicle subsets with 5 cm depth increments and mixed-vehicle subsets with 1 cm depth increments, spanning seven vehicle types, four weather conditions, and flood depths from 0 to 40 cm. FloodLlama-Dense achieves an MAE of 0.40 cm, an RMSE of 1.97 cm, and an Acc@5cm of 97.59%. Mechanistic interpretability analysis combining linear probing, logit lens, centered kernel alignment (CKA), and cross-attention entropy reveals a two-stage adaptation pattern: layers L13-L22 restructure visual representations, while depth first becomes linearly decodable at layer L23. FloodLlama-MI5 and FloodLlama-MI6 leverage this insight by fine-tuning only five or six of the eight cross-attention layers, achieving an 86-88% reduction in trainable parameters (6.55-7.86 million versus 54.4 million) with minimal accuracy loss. On a real-world benchmark, FloodLlama-MI6 achieves 98.62% accuracy, compared with 86.61% for the published STURM-FloodDepth baseline.

Vision-language models (VLMs) process image patches and text tokens in a shared residual stream, but the local geometry through which the two modalities interact remains poorly understood. Most interpretability methods identify global linear directions, which may miss representations that are globally high-dimensional but locally low-dimensional. We introduce LENS (Local Explanation of Neighborhood Subspaces), a method that decomposes VLM activations into local low-rank Gaussian neighborhoods using a Mixture of Factor Analyzers. Applied to LLaVA-1.5-7B and Qwen3-VL-8B, LENS reveals distinct depth-dependent fusion trajectories consistent with each model's fusion mechanism: LLaVA progressively mixes modalities at later layers, whereas Qwen3-VL mixes them early, partially re-segregates them, and recombines them near the output. An automated multimodal labeling pipeline assigns concise semantic descriptions to these neighborhoods. Interpolating activations toward neighborhood centroids causally redirects generation within and across modalities and outperforms difference-in-means and VL-SAE in most evaluated conditions; in one LLaVA vision-to-vision setting, MFA achieves 5.7 times the VL-SAE score. Human evaluation finds MFA steering competitive with prompting and substantially stronger than the other intervention baselines. Finally, the MFA coefficient space improves Qwen3-VL image-to-rendered-text retrieval at the deepest evaluated layer from 14.9% to 48.6% R@1. Ablations show that the reported fusion trajectories are stable across component counts, local ranks, and modality-purity thresholds. These results support local geometric neighborhoods as useful interpretable and causal units for analyzing cross-modal representations in the evaluated VLMs.

Tool-integrated vision-language agents have made remarkable progress on compositional and multi-step visual reasoning. Yet their outputs frequently exhibit unfaithfulness: the stated reasoning path diverges from the computation that actually produced the answer, undermining reliability in safety-critical applications. We present DiffuseAgent-MI, a self-evolving agent whose perceptual grounding is governed by a KL-minimal energy model over feature units, providing a distributional view of visual mechanistic interpretability. The agent learns an energy landscape that softly constrains generated samples to lie near the native prior conditioned on the chosen interpretable unit, closing the gap between the explanation and the internal representation. A verifier then supplies trajectory-level faithfulness rewards, and a repair branch re-conditions the energy when the verifier flags an unfaithful step. On GeoQA, SciVis, VQA-v2 and an in-house multimodal reasoning set, DiffuseAgent-MI improves accuracy by up to 5.1 points over prior self-evolving agents while more than doubling mutual-information faithfulness and human-interpretability agreement. Our analysis shows the energy term and the verifier are complementary: the former guarantees distributional faithfulness, the latter trajectory-level faithfulness, and only their combination closes both gaps.

As Multimodal Large Language Models (MLLMs) are increasingly deployed in decision-critical pipelines such as robotics, embodied AI, and safety monitoring, the opacity of their spatial judgments limits operator trust and auditability. MLLMs demonstrate strong reasoning but often struggle with fine-grained spatial understanding and object hallucination. Prior work, ByDeWay, introduced Layered-Depth-Based Prompting (LDP), a training-free framework that mitigates hallucinations by structuring prompts using monocular depth estimation. However, coarse depth layering falls short in resolving object-to-object spatial relationships within the same geometric plane, such as projective ("left of", "above") and topological ("inside", "touching") relations. We propose ByDeWay-V2, which integrates explicit spatial relational context alongside depth cues, expressed as human-readable predicates that serve as auditable evidence for downstream decision support. Using an open-vocabulary object detector (YOLO-World-L), our framework computes pairwise geometric relations between detected objects and injects them as structured spatial predicates into the MLLM prompt, bridging 3D scene depth and 2D spatial semantics without any training. We evaluate ByDeWay-V2 on the Visual Spatial Reasoning (VSR) and BLINK benchmarks across multiple MLLMs, with hallucination grounding assessed via POPE. On the BLINK spatial subset, ByDeWay-V2 achieves a 46 percent relative F1 improvement over LDP for Qwen2.5-VL, and recovers BLIP-Base's spatial reasoning on VSR from near-random performance to a competitive F1 of 0.53. Our lightest configuration operates under a strict 40-token context budget on CPU, showing the framework's suitability for resource-constrained, real-time decision-support settings.

Reliable visual document understanding requires a model to attribute each answer to the evidence regions that support it. Recent benchmarks and systems express this step through a coordinate interface: the model outputs the coordinates of bounding boxes that mark the evidence regions in the document. Under this interface, vision-language models often fail to identify the right regions even when the answer is correct, a failure known as Attribution Hallucination. We present a study that investigates whether this failure is partially limited by what the model can express through coordinates. On a verified bilingual CiteVQA subset, we compare the coordinate interface with a language interface in which the model outputs only text, quoting its evidence verbatim, and a multimodal retriever returns the location of each quote as a page region proposed by a layout parser (tables and figures are quoted through their captions or notes); the comparison is repeated over six open vision-language models. Compared with the coordinate interface, evidence recall rises from at most 8 points to between 26 and 47 and the hallucination rate roughly halves, with little change in answer quality. Building on this comparison, we use the same quote-and-retrieve pipeline as a training scaffold: because region-level evidence labels are expensive to collect for long documents, we introduce a GRPO recipe whose reward is a judge's reading of the gold answer and crops of the retrieved regions, training the model to quote better evidence without any region labels and raising an 8B backbone's strict attributed accuracy from 22.4 to 33.8. These findings indicate a practical path to improve attribution"without a coordinate interface and without costly region-level supervision.

Vision-Language Models (VLMs) are demonstrating significant capabilities in medical tasks like radiology analysis, yet providing faithful and interpretable explanations remains a key consideration for their responsible deployment in clinical settings. However, existing explanation methods, such as the widely used FIxLIP framework, often struggle with the fine-grained nature of modern tokenizers. The tokenization problem fragments clinical concepts---splitting terms like "saddle embolus" into scattered, meaningless subwords---which leads to noisy, semantically incoherent cross-modal attributions. Such fragmentation also results in a combinatorial explosion of interaction possibilities, obscuring the model's true reasoning. To address this, we introduce ParseFIxLIP, an extension that incorporates the Tree-Gram Parsing into the Banzhaf interaction game used by FIxLIP. This semantically informed strategy utilizes dependency parsing trees to define explanation players by grouping related text tokens into semantically coherent units. Our smart_depth grouping strategy, merging tokens according to spaCy token dependency tree, successfully mitigates concept fragmentation, yielding substantially more interpretable cross-modal interactions by unifying complex medical concepts. Quantitatively, while baselines struggled with the high dimensionality of long captions, our parsing approach maintained statistical robustness and semantic parsimony. Qualitative analysis on BiomedCLIP, validated on medical imagery (ROCOv2) and general examples, confirms that the approach accurately captures the synergistic influence of grouped words on model predictions. In conclusion, our work offers intuitive and clinically relevant insights into VLM decision-making, fulfilling the critical need for coherent explanations in the medical domain.

Large vision-language models are becoming increasingly dominant in 3D medical image interpretation, but we rarely know <i>which</i> internal units encode clinical findings or <i>where</i> that information lives in the representation. We first study this on a 3D chest vision-language model (Pillar-0) by probing its frozen vision embeddings. We show that (i) each radiological finding is encoded by a <i>sparse</i> set of ~10 vision-encoder channels that match full-feature classification performance and far exceed a zero-shot text prompting; (ii) turning off the channels tied to one finding, that finding's score collapses while unrelated labels stay stable; and (iii) the same sparse probe <i>replicates</i> on an architecturally unrelated 3D abdominal VLM (Merlin) suggesting a general property of frozen medical encoders. Our training-free concept channel probe (CCP) method, paired with a corpus-derived report template, outperforms published CT-CHAT on clinical efficacy and NLG metrics (F1 0.549 vs. 0.184; BLEU 0.483 vs. 0.373) at 22x lower latency. Our results provide a clear, reproducible characterization of how frozen medical encoders represent findings, demonstrating direct applicability across models.

Goal-directed visual processing is a hallmark of human visual intelligence, resulting in representations that support downstream tasks such as categorization or search. Though vision-language models (VLMs) are often faced with these same tasks, their ability to recode visual representations when presented with goal-directed language remains poorly characterized. Indeed, prior work largely treats visual representations in VLMs as static repositories of visual information that are manipulated by language representations. In the present work, we provide evidence for two concrete instances of language-induced recoding of visual representations. First, we identify an abstract reference representation that denotes which objects are goal-relevant under a natural language prompt. We extract contrastive steering vectors corresponding to this reference representation and demonstrate that they are causally implicated in model predictions. These reference representations are abstract in that they generalize to different objects, different task contexts, and even from synthetic to naturalistic images. Second, we demonstrate language-induced attribute modulation: later layers selectively amplify goal-relevant attributes in visual representations of objects. We demonstrate this phenomenon across a range of different prompts. Finally, we provide a causal intervention that demonstrates that attribute modulation mediates a VLM's response distribution. Together, our results support a more dynamic account of cross-modality processing in VLMs -- rather than vision tokens serving as static repositories of information, they are modulated to support queries articulated in language.

A popular route to interpretable zero-shot classification asks a large language model (LLM) to describe each class name and prompts CLIP with the resulting descriptors. We show that these descriptors carry little visual evidence of their own: removing the class name from the prompt collapses ImageNet accuracy from 59.5% to 15.5%. The diagnosis is that the descriptors are conditioned on the label rather than on the images, so they describe the concept in general and mislead exactly when the data shifts; an LLM insists that strawberries are red, but every strawberry in ImageNet-Sketch is a colorless line drawing. We therefore select attributes from the target image collection instead: we score a large attribute pool against the images in CLIP's joint embedding space and keep the top-scoring attributes per class. Selected this way, class-name-free attribute prompts reach 23.8% on ImageNet (against 15.5% for LLM descriptors), the gain holds on four shifted ImageNet variants, and reselecting from the LLM's own pool isolates the selection mechanism as the cause. With one image per class, the selected attributes outperform the prompt-tuning method CoOp by 3 points while fitting in under a minute instead of 14 hours, with no learned soft prompt to obscure the decision. Because the attribute set is chosen by the data, it doubles as a readable summary of a dataset, which we use to describe distribution shift in words. Our code and results are available on our project page: https://ggare-cmu.github.io/AttributeSelect/

Attention and saliency heatmaps are widely used to explain medical Vision-Language Model (VLM) outputs on chest X-rays, yet whether they truly highlight the image evidence driving predictions has not been causally tested. We audit faithfulness via overlap with radiologist bounding boxes on PadChest (n=637), attribution mass within radiologist masks on CheXlocalize (n=643), and 16x16 patch-occlusion maps that record which regions, when hidden, change the answer. We study three MedGemma-4B variants, cross-family probes on LLaVA-RAD and Qwen3-VL-8B-Instruct, and the specialist CheXagent-2-3b, with two CXR-trained classifiers (DenseNet121, ResNet50) as positive controls. A heatmap is faithful only if the model uses the image and attention concentrates on regions whose occlusion alters the prediction. No evaluated VLM meets both criteria. MedGemma and Qwen3-VL use the image, but attention anti-correlates with patch-occlusion importance (rho < 0 with 95% bootstrap CIs below zero). LLaVA-RAD's attention correlates positively, but the model is almost text-only (99.1% text-only agreement, near-zero causal mass), so correlation ties two near-zero signals. Attention also misses annotated anatomy: overlap with true regions never beats shifted or random controls, and no method places more than 22% of its mass inside radiologist masks. The two CXR classifiers pass all metrics, indicating the failure is specific to VLM heatmaps, not the evaluation. These heatmaps are visually reassuring but not faithful; clinical explanations require controlled localization metrics and causal perturbation, not visual inspection alone.

Understanding how vision-language models (VLMs) interpret data visualizations remains an open problem, and is increasingly important as these models are used for analytical tasks where reliable reasoning is essential. We introduce a lightweight, diagnostic saliency map method tailored for text generation over images using transformer models, the current state-of-the-art models in visualization interpretation. Our approach aggregates the language model's attention over the visual tokens across all heads and layers, then maps this attention back onto the vision encoder's patch grid to localise it over the image, producing a direct correspondence between each generated answer token and the image regions it attended to. This yields fast, gradient-free saliency maps that expose how VLMs allocate focus across visual elements during answer generation, enabling inspection of whether model attention aligns with semantically relevant components. We evaluate our approach using a deletion metric which validates the causal faithfulness of our saliency maps to the model's behavior.

Compositional visual question answering requires Vision-Language Models (VLMs) to execute multiple reasoning operations like object selection, spatial relation resolution, and attribute verification. Despite strong aggregate performance, the mechanistic basis of VLM failures on this task remains underexplored. To address this gap, we analyze vision-operation misalignment in VLMs by examining how failures relate to specific reasoning operations and the internal computational pathways through which they arise and propagate. We introduce an Operation-centric mechanistic framework that decomposes VLM failures by both the reasoning operation where they originate and the internal computational pathway through which they propagate. Our analysis reveals four mechanistically distinct failure modes: grounding failure, reasoning failure, attribute extraction failure, and language prior dominance failure. Each characterized by a unique relationship between visual grounding strength and answer correctness. Through three complementary causal interventions applied across all transformer layers, we further demonstrate a pathway dissociation: grounding failures route exclusively through the feedforward network, reasoning failures route through late-layer attention, and attribute extraction failures localize to the answer-position feedforward computation. This dissociation demonstrates that different failure types require fundamentally different corrective strategies, providing a principled foundation for targeted improvements to VLM reliability in multimedia reasoning.

When do vision-language models need direct access to the image while generating an answer? We study image dependence during answer generation by examining how the visual information needed for the current question becomes available in context. We intervene on direct image access while retaining previously computed states. Across real-image and synthetic tasks, we show that, depending on the generation process, direct access can continue to support accuracy after question processing. Supplying the required attributes as text in the context weakens this dependence. On synthetic tasks, we also examine how dependence changes as the model itself states the required attributes. Before attribute expression, severing access reduces accuracy, and replacing image-side states shifts answers toward the counterfactual content. After sufficient expression, both interventions have smaller effects. Even with an identical generated prefix, dependence differs according to whether image access was available during question processing. Thus, both the visible text and the preceding image access matter. Several of these patterns hold across model families, including Qwen2.5-VL-32B and InternVL3-14B. These findings offer a view of image dependence in terms of the information needed for the current question and the history of image access, beyond generation position alone. This perspective provides a basis for deciding when to reduce visual access during an answer and which visual information to retain for subsequent questions.

A score on a temporal video question answering benchmark is meant to measure that a model has temporal understanding, but it conflates two questions. 1. The task question: is the question even temporal, does it need several frames and their order? and 2. The channel question, when it does, does the model recover the order from the pixels, or read it off the positional encoding (RoPE)? Most of a temporal score answers neither, a single frame and answer priors often carry it. The field's validity checks, frame-shuffle sensitivity and the accuracy gained from the full video, speak only to the task question. We contribute a label-free screen for the channel question, the reversal-drop: the accuracy lost when the visual sequence is reversed while RoPE remains forward. It can be applied to compatible temporal benchmarks without new annotations. Paired reverse labels, or tasks whose labels transform deterministically under reversal, distinguish models that follow reversed content from those merely disrupted by the conflict. Molmo2 answers the forward event reading order off positions, while Qwen3-VL answers the reversed event it actually sees, reading visual order (comparatively). We call them position-dominant and visual-sequence-dominant. The split holds across two benchmarks and several temporal tasks at two scales, and activation patching shows it is a real internal property, not an artifact of the conflict. The distinction matters, the two channels fail on opposite inputs so two models with similar score are not interchangable, i.e. an aggregate score does not reflect potential failure modes.

Vision-language models increasingly succeed on multimodal reasoning benchmarks, yet their visual evidence often becomes unstable once it enters the language stack, weakening evidence-grounded reasoning. To understand this fragility, we examine the internal dynamics of VLMs through a mechanistic lens and uncover a stable three-stage redistribution of multimodal attention focus across depth: an early question-conditioned organization, a critical middle visual-dominant relay, and a late return to answer formation. We operationalize the middle phase as the Visual Relay Window (VRW), and show that its geometry varies with task demand, is causally tied to grounded generation, and distinguishes unsupported answers from stronger reasoning trajectories. Guided by this internal rhythm, we propose TRACE, a task-adaptive inference-time control framework with lightweight trained modules. It reshapes relay allocation during prefill and preserves assembled visual support after handoff during decoding. Across four open-weight VLM backbones and seven benchmarks, TRACE delivers large gains on grounding-sensitive settings, improving them by 4.33 points on average and by up to 6.6 points, while also improving reasoning-heavy tasks. These results show that explicitly controlling multimodal focus across depth offers a unified and effective mechanism for strengthening evidence-grounded multimodal reasoning.

Large Vision Language Models (LVLMs) require strong reasoning over both visual and textual input. Recent work suggests that cognitive elements, especially diverse representations and metacognition, correlate with better performance. Many of the needed perceptual functions are already provided by specialized domain-specific computer vision models, which act as the perceptual subsystem for detecting objects, localizing them, inferring states, recovering spatial layout, and reading text. The key challenge is to integrate these multi-encoder experts into a trustworthy, interpretable, and coherent representation that improves verifiability and reduces hallucinations. This is difficult because vision-language questions span different cognitive levels, yet most LVLM pipelines apply the same perception-reasoning routing regardless of the demand of each query. We propose an evidence-driven multimodal reasoning framework that utilizes a Bloom-inspired taxonomy as a hierarchical reasoning protocol. The two-stage cognitive verbalization first produces a Literal Evidence Summary by decomposing expert outputs into short, atomic evidence statements. It then performs Bloom Verbalization to turn these evidence items into a staged reasoning trace, and a lightweight Reasoning Trace Module quantitatively analyzes the trace to make evidence usage and reasoning progression explicit. Through this integration, we observed several improvements in perception and reasoning abilities. Moreover, the trace module provides quantitative evidence that different queries induce different cognitive entry levels and evidence-use trajectories that enable fine-grained analysis.

Despite strong performance on many multimodal tasks, vision-language models (VLMs) still struggle with basic object counting. We investigate whether this reflects missing internal knowledge or a gap between internal representations and verbalized outputs. Training simple probes on activations from four VLMs across five counting datasets reveals that nonlinear probes can reliably detect counting errors, suggesting that VLMs often encode the correct count even when they output the wrong answer. SVCCA analysis shows that probes trained on ground-truth counts and probes trained on model outputs occupy a partially shared activation subspace but read out along misaligned directions. We further validate our findings using a causal steering intervention, proving that strengthening the direction of count-identified probes does improve model counting performance. Motivated by this result, we propose a detector-guided self-correction method that selectively re-prompts the model only when an internal error detector predicts failure. This simple inference-time intervention improves counting accuracy by up to 15.6 absolute percentage points, without any parameter updates. Our results establish activation-based error probing as both a practical tool for improving VLM counting and a mechanistic lens on the gap between internal knowledge and model outputs.

Sparse autoencoders (SAEs) have emerged as a promising technique for mechanistic interpretability by learning a set of sparse latent features in large models, each of which encodes a distinct concept. However, in vision-language models (VLMs), vanilla SAEs struggle to learn modality-consistent concepts, with concepts often exhibiting fragmented coverage (i.e., disjoint regions) in the visual modality. To address this challenge, we propose a Structured Sparse AutoEncoder ($S^2AE$) that enforces concept consistency from both semantic and spatial perspectives in the visual modality. Specifically, we group image patches based on Transformer attention similarity and spatial proximity, and introduce a structured sparsity regularization when training the vanilla SAE. The regularization consists of exclusive sparsity for inter-group concept disentanglement and group sparsity for intra-group concept consistency, which drives the latent neurons by SAEs to specialize in distinct, semantically grounded concepts. Evaluated on the \texttt{Qwen2.5-VL-7B-Instruct} model, the method achieves 6.06% average improvement in semantic alignment (mIoU) and 60.81 in representational efficiency (lower l0 norm) while maintaining near-perfect reconstruction fidelity with an Explained Variance above 99%. Cross-modal analysis further demonstrates that $S^2AE$ enhances neuronal monosemanticity by this visual structural prior, achieving a 3.08% average gain in semantic consistency and a 2.37% average gain in monosemanticity scores for both modalities of multimodal features, thereby fostering more coherent and disentangled representations.

Faithful visual attribution identifies which image regions support a model prediction. Search-based perturbation methods lead the insertion--deletion faithfulness frontier by masking regions and measuring score changes, but they usually output a complete ordering of all regions. Many applications, especially MLLM attribution and repair, only need a compact top-$k$ evidence mask. We study this mask-first attribution problem. An exactly $k$-region mask is combinatorial: useful evidence can depend on interactions among fine regions. Coarse grouping can stabilize early search but aggregates redundant content, whereas one-step scoring can miss high-value combinations. We introduce two forward-only methods. \textsc{CoPAIR} uses a PhaseWin--Greedy gap diagnosis to construct coarse singleton/pair candidates that warm-start full-ordering search. \textsc{TRACE} directly searches fixed-cardinality fine-region masks with cross-entropy sampling, elite retention, and distribution updates, with a finite-budget recovery analysis. The resulting evidence set can be returned as a compact attribution mask or used to initialize Greedy or PhaseWin when a complete ranking is required. Across ImageNet classification with CLIP ViT-L/14, CLIP RN101, and ResNet-101, our initialized search methods establish a new state-of-the-art frontier for faithful full-ordering attribution under inclusive forward-call accounting. On POPE and RePOPE with Qwen2.5-VL-3B-Instruct and LLaVA-v1.5-7B, \textsc{TRACE}+Greedy gives the strongest search-based MLLM attribution results. Direct \textsc{TRACE} masks further achieve single-point RePOPE repair rates of $94.44\%$ and $96.00\%$, showing that compact evidence masks can be actionable attribution outputs, not merely prefixes of full rankings.

Vision-Language Models (VLMs) are increasingly utilized as the conditioning backbone for diffusion-based image editing due to their remarkable multimodal reasoning capabilities. While standalone VLMs demonstrate strong localization capabilities, editing pipelines frequently struggle to maintain this accuracy, particularly in complex, multi-entity scenes. In this work, we investigate this performance gap, hypothesizing that it stems from treating the VLM as a condition encoder. In this role, the model is restricted to a single forward pass, preventing the autoregressive generation process for which it was optimized, thereby failing to fully expose its capabilities. To investigate whether this spatial understanding persists when the VLM is used as a condition encoder, we introduce Analysis-by-Proxy. In this framework, we train a lightweight, interpretable proxy model on the VLM's intermediate representations using an auxiliary localization task. By analyzing the VLM through this proxy, we uncover the specific VLM representations that encode localization information. Our findings expose a fundamental mismatch between how spatial knowledge is represented within a VLM condition encoder and how it is extracted by current editing pipelines. We reveal that under single-pass constraints, the localization signal does not reliably propagate to the predefined layer configurations commonly used for conditioning. Instead, this crucial signal remains hidden within intermediate representations, at locations that vary depending on the input prompt. Using our introduced Analysis-by-Proxy framework, we reveal the fundamental failures of existing condition extraction strategies in editing pipelines, opening the door to more principled design of conditioning architectures.

A frozen vision-language-action model (VLA) receives recent observations at every decision step, yet prior work has focused on adding memory rather than asking how existing history is represented and used. We study this temporal axis using layer-resolved linear probing and causal interchange interventions across three VLAs from two architecture families. We find a three-part dissociation. First, past-frame content remains linearly decodable throughout the network. Second, information unique to history beyond the current frame is nearly absent, indicating that stored history is largely a redundant copy of the present. Third, history is causally deployed only when the current frame is heavily degraded, while the action readout progressively loses dependence on history through the network. Although all models encode history similarly, their deployment strategies differ: under the same occlusion, one architecture increasingly relies on history as a fallback, whereas the other relies on it less. We further introduce a training-free temporal deployment audit that distinguishes these regimes. In the fallback regime, re-injecting history neither repairs occlusion nor disambiguates actions, confirming the redundancy of the stored representation. In the other regime, the same intervention reliably steers the predicted action toward the donor history. These results show that steerability depends on how history is deployed rather than whether it is encoded. VLAs do not forget the past; they largely fail to represent it as information distinct from the present. Our findings suggest that future memory augmentation should inject information unique to the past rather than simply more history.

Vision-language models (VLMs) increasingly read news and web content as images, where the publisher's identity is visually present. We show that VLMs carry a strong source-credibility prior keyed on outlet identity, and study it along three axes. (i) Cross-model benchmark. We introduce CueTrust, a cross-model diagnostic that measures which surface source cue overrides an article's content evidence via a Source-Override Index (SOI). Across seven VLMs and five cues, the vulnerability profile is model- and scale-dependent, and the override is outlet-identity-specific and encoding-invariant, firing from the masthead name, the logo image, or the bare domain, but not from a named author, in-text authority, or page layout (clean negative controls). (ii) Mechanistic account. For the brand cue, we give a full mechanistic account: swapping only the masthead moves credibility across an approximately 11 log-odds range that tracks professional ratings (rho = 0.88 with Media Bias/Fact Check). The prior is dual-coded (name and logo), strengthens with scale, is causally formed at layers 19-21, carried by interpretable seed-stable sparse-autoencoder features, and recurs at the same relative locus in a second model family. It overrides content (about 1.8x) as a signal-magnitude effect within a shared pathway, not a privileged route. Steering the localized direction selectively reduces the override (41% reduction) and generalizes to held-out outlets, confirming the prior is causally used, not merely decodable. Deployed VLMs may thus defer to source identity over the evidence in front of them, a reliability failure we can measure across models, localize, and causally probe. We release the stimulus suite and CueTrust.

We study how visual information is routed in vision-language models (VLMs). Using causal patching on controlled synthetic and natural datasets, we find that models rely on two distinct pathways to solve visual tasks: A direct pathway, where visual information is retained in image token representations and read out by the final token at later layers, and a text-mediated pathway, where visual information is first transferred to the query tokens and then read out by the final token. Across three visual tasks, we show that pathway selection is task-dependent, and that data distribution and prompt design can also modulate which pathway is used to solve the image-based query. Moreover, using attention knockouts and corrupted-input patching, we find that these pathways are flexible, under certain interventions, models can rely on the text-mediated pathway as a fallback when the usual pathway is ablated. This behavior unifies findings in prior work and shows that ablation-based interventions can reveal what models could do rather than what they normally do. Together, our results provide a mechanistic characterization of visual information flow in VLMs and highlight the flexibility of their internal mechanisms under intervention.