VLM Interpretability

VLM: Vision-Language Model

Follow this topic Ask about this topic

Momentum

24 papers in the last four weeks, up 200% on the four weeks before. 0.2% of all new papers.

Jul 13Week of Sep 28

Latest papers 158

Jun 4, 2026cs.CL

Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads

While Multimodal Large Language Models (MLLMs) demonstrate remarkable proficiency on complex vision-language tasks, the mechanisms by which they extract query-relevant visual features from complex, noisy contexts remain opaque. In this paper, we present an in-depth interpretability study that uncovers a profound structural property within MLLMs: functional sparsity in cross-modal retrieval. Leveraging a token-level metric termed Retrieval Attention Mass (RAM), we identify and characterize a highly specialized subset of attention heads, referred to as Context-aware Retrieval (CoRe) heads. Across diverse visual domains and model scales, we observe a clear functional division: CoRe heads act as dedicated information extractors, while most other heads distribute attention over broader contextual regions. Causal interventions further demonstrate the necessity of these specialized heads. Ablating only the top 5% of CoRe heads causes significant degradation in multimodal reasoning performance, whereas ablating lower-ranked heads has minimal effect. Moreover, acceleration experiments validate the utility of CoRe heads, showing that leveraging this localized sparsity significantly accelerates inference while maintaining robust task performance. Our findings reveal a structural principle of functional sparsity within MLLMs, refining the current understanding of mechanistic interpretability and laying a theoretical foundation that can inspire future architecture design and model optimization.

Ruoxi Sun, Quantong Qiu, Juntao Li +3

Soochow University · Peking University

Multimodal Large Language Models VLM Interpretability Attention Head Specialization

Jun 2, 2026cs.CV

Visual Instruction Tuning Aligns Modalities through Abstraction

Visual instruction tuning effectively adapts a pre-trained Large Language Model (LLM) to process image information alongside text. Yet, it remains unclear how visual features are embedded into the layer-wise hierarchy of abstractions of the LLM backbone. Across a diverse set of vision-language architectures, we show that instruction tuning primarily serves as a bridge, embedding visual features directly into the intermediate semantic layers of the LLM, bypassing the early layers devoted to unimodal processing. With probing analyses and causal interventions, we show that these intermediate layers are the semantic core of vision-language processing and play a critical role in the performance on a broad set of multimodal benchmarks. In addition, by comparing the geometry of semantically equivalent visual and textual representations, we find that fine-tuning extends and strengthens the existing abstraction phase, aligning visual features with pre-existing textual ones. Finally, we confirm the functional role of this localized alignment by restricting fine-tuning to intermediate layers alone: this strategy preserves the performance of full fine-tuning on vision-centric benchmarks while reducing training time. Our results suggest that multimodal integration is a localized phenomenon driven by the repurposing of the internal abstraction engine of the LLM.

Luis Palacios, Lorenzo Basile, Diego Doimo +1

Area Science Park, Trieste, Italy

VLM Adaptation VLM Interpretability Vision-Language Alignment Multimodal Fusion Multimodal Instruction Tuning

Jun 2, 2026cs.CV

Beyond Compression: Quantifying Spectral Accessibility in Vision Representations

Vision-language models map visual features into a shared embedding space through learned projection layers, yet it remains unclear how these transformations alter the structure of visual information. This study examines changes in representation through spatial-frequency accessibility, measured by the linear recoverability of band-limited Fourier energy from model representations. To isolate effects beyond dimensionality reduction, we introduce Residual Spectral Loss (RSL), which evaluates changes relative to a dimension-matched random projection baseline. To reduce confounding effects from optimization, the analysis uses pretrained models with all parameters frozen. The experimental results show consistent frequency-dependent changes in accessibility across CLIP and DINOv2 on ImageNet and MS-COCO datasets. Spectral accessibility follows a non-monotonic trajectory across depth, peaking at intermediate layers before decreasing toward the output representation. The final transformation differs across architectures: CLIP's learned projection is spectrally neutral, with changes explained by compression, whereas DINOv2's [CLS] pooling induces a structured loss across the spectrum. These findings identify intermediate layers and pooling mechanisms as primary drivers of spectral transformation in modern vision encoders.

Akayou A. Kitessa, Yijun Zhao

Fordham University New York City, NY

Visual Representation Learning Frequency-Domain Feature Learning VLM Interpretability

Jun 2, 2026cs.AI

Decomposing how prompting steers behavior

Prompting steers large language models (LLMs) and vision-language models (VLMs) without weight updates, but it remains unclear how instruction changes reshape internal representations to produce behavior. We introduce a nested geometric decomposition framework that treats prompting as a transformation of the representational geometry of the content following the prompt. For each prompt pair, we align representations of the same stimuli under two prompts using increasingly expressive stimulus-invariant maps: translation, rigid transformation with uniform scaling, sequential axis scaling, affine transformation, and nonlinear transformation. We then causally test each map by replacing a single layer's prompt-A hidden state for held-out stimuli with its mapped counterpart and measuring recovery of prompt-B representational geometry and behavior. Across three LLMs, three VLMs, and six text or image datasets spanning style, emotion, scene content, and number, prompts consistently reshape representations toward the instructed task structure. Cross-validated variance decomposition shows that much prompt-induced activation change is captured by shape-preserving maps, especially translation and rigid transformation with uniform scaling, while tier profiles reveal model- and task-specific routing strategies across layers. Crucially, although translation and rigid tiers already improve behavioral agreement, affine transformation is the first tier to nearly recover target-prompt task geometry and yields corresponding behavioral gains. This suggests that cross-dimensional linear mixing is a key mechanism by which prompts reorganize representations toward instructed task structure. Our framework decomposes prompt-induced representational change into interpretable geometric components and reveals how models route task-relevant structure to produce prompt-driven behavior.

Fan L. Cheng, Nikolaus Kriegeskorte

Columbia University

LLM Prompting LLM Interpretability VLM Interpretability Mechanistic Interpretability Causal Interventions in Language Models Representation Geometry in Language Models

Jun 1, 2026cs.CL

Encoded but Not Routed: Explaining the Table-Chart Gap in Scientific Claim Verification

Multimodal LLMs are increasingly used to assist scientific peer review, where a core requirement is verifying whether claims in a paper are supported by its evidence. Prior work has shown that models perform substantially better at this task when the evidence is a table than when it is a chart of the same underlying data. This raises the question of whether models fail to extract information from charts, or do they extract it but fail to use it when forming their prediction? We study this question through layer-wise linear probing and attention analysis on three open-weight VLMs over table and chart evidence, representing the same underlying data. We find consistent evidence for the latter. Chart information is encoded in the models' intermediate representations but does not reach the prediction position, a gap that is absent for tables and holds across all conditions tested. Attention analysis further reveals that this disconnect takes two architecturally distinct forms across model families. These findings reframe the table-chart gap as a failure of how encoded visual information is routed at prediction time, rather than a failure of encoding itself.

Sunisth Kumar, Xanh Ho, Tim Schopf +3

The University of Tokyo · NII LLMC · National Institute of Informatics +2

VLM Interpretability Chart Understanding Multimodal Model Evaluation Scientific Claim Verification

Jun 1, 2026cs.CV

Self-Improving Small Object Grounding in LVLMs

Can internal attention patterns in Large Vision Language Models (LVLMs) identify reliable small-object boxes without fine-tuning? In this work, we provide an affirmative answer. Attention structure in LVLMs encodes grounding quality-a lightweight IoU regressor trained solely on attention maps achieves strong IoU prediction (Pearson r > 0.67). This regressor powers the regressor-based variant of our Attention-based Candidate Selection (ACS) framework, called ACS-Learned, which selects the best box from multiple sampled candidates to improve object grounding. By analyzing what the regressor learns, we reveal which transformer layers and heads are most critical and derive ACS-Free: a training-free selector that ranks candidates by attention entropy on these discriminative heads, with no learned component at inference. Experiments on COCO and Objects365 demonstrate up to 19% self-improvement on small object localization, with ACS-Free ranking best among all training-free methods, demonstrating that useful attention structure improves both localization reliability and interpretability in LVLMs.

Tianze Yang, Yucheng Shi, Ruitong Sun +2

University of Georgia

Large Vision-Language Models VLM Interpretability Vision-Language Grounding Visual Grounding

May 31, 2026cs.CV

Beyond Visual Memory: Mechanistic Diagnostics of Latent Visual Reasoning

Recent latent visual reasoning methods achieve substantial gains by inserting continuous latent tokens into multimodal language models. These gains are commonly attributed to the tokens encoding visual evidence; recent analyses, however, reveal a paradox: the tokens are loosely tied to the image and contribute little to the answer. Critically, these analyses treat latent tokens as a single unit, obscuring the true source of the gains. We therefore decompose latent tokens into three testable components: latent slots, boundary markers, and format, and develop a state-of-the-art method as a probe under favorable conditions. Across six method-stage settings and four perception-heavy benchmarks, latent slots fail every prediction of the visual-memory account. Strikingly, retaining only the boundary markers preserves 78 to 100% of the gain in several settings, while the model attends to the image more narrowly at latent positions than at answer positions. The gain therefore comes from boundary markers, format, and this attention pattern, not from latent slots. How each method engages this mechanism depends on its training supervision: at matched accuracy, mechanisms can still differ markedly. Latent visual reasoning thus needs evaluation not only by accuracy but by what the model actually relies on.

Garvin Guo, Yu Chen, Xiang Wang +4

Amap, Alibaba Group · Shanghai Innovation Institute

VLM Evaluation VLM Interpretability Latent Visual Reasoning VLM Reasoning Visually Grounded Reasoning Multimodal Understanding

May 30, 2026cs.CL

MLLM-Microscope: Unlocking Hidden Structure Within Multimodal Large Language Models

This work presents MLLM-Microscope, a novel system designed for analyzing the hidden representations within Multimodal Large Language Models (MLLMs). Our system evaluates the linearity, intrinsic dimension, and anisotropy of multimodal token embeddings across transformer layers. Utilizing the ScienceQA dataset, we evaluate two state-of-the-art MLLMs, LLaVA-NeXT and OmniFusion. We find that both the main and residual streams for tokens of both modalities exhibit highly linear behaviors across transformer layers. However, LLaVA-NeXT's image tokens reveal a slight decline in linearity, whereas OmniFusion's remain consistent. Image token dimensions in OmniFusion remain consistently higher across layers compared to LLaVA-NeXT. Also, the OmniFusion's anisotropy is observed to stay consistently low throughout the layers. These findings suggest that the inner workings of MLLMs highly depend on the nature of modality fusion performed before passing the token sequence into LLM. This and other new potential insights obtainable from our system are surely capable of enhancing our understanding of the inner workings of MLLMs, informing future model design and optimization.

Ravil Mussabayev, Rustam Mussabayev

Satbayev University

Representation Geometry Multimodal Large Language Models VLM Interpretability Multimodal Embedding Multimodal Foundation Models Multimodal Model Interpretability+1

May 30, 2026cs.CV

Through the PRISM: Principle-Aware, Interpretable, and Multi-Scale Evaluation of Visual Designs

Effective visual communication stems from the harmony of multiple design principles, such as readability, contrast, alignment, overlap, and coherence, which collectively govern clarity and intent of the communicator. While human designers reason holistically over these principles, machine agents typically condense them into a single heuristic score, offering limited interpretability and diagnostic precision. To address this gap, we introduce PRISM (PRinciple-aware, Interpretable, and Structure-guided Design Modifications), a benchmark that systematically perturbs professional layouts from the Crello dataset along measurable design principles. The benchmark comprises 100K perturbed training samples and 10K perturbed validation designs, each isolating a specific principle violation for controlled analysis of multimodal reasoning about design quality. We show that models like Qwen-2.5-VL and GPT-4o-mini are largely insensitive to targeted principle degradations, whereas GPT-4o exhibits global awareness without fine-grained disentanglement. Building on these insights, we propose a multi-scale evaluation framework that integrates lightweight scorers for quantitative assessment, instruction-tuned vision-language models for localised feedback, and prompt-based methods for global reasoning. Our framework provides interpretable explanations of design failures. Using these localised insights, we show targeted refinements that improve layout quality. Together, PRISM and our framework lay the foundation for interpretable design-literate multimodal reasoning systems.

Mona Gandhi, KJ Joseph, Srinivasan Parthasarathy +1

Ohio State University · Adobe Research

VLM Evaluation VLM Interpretability Multimodal Reasoning

May 29, 2026cs.CV

Vision-Language Models Suppress Female Representations Under Ambiguous Input

Alignment teaches vision-language models (VLMs) to avoid expressing demographic biases, and when gender is clearly visible they largely succeed. Far less is known about ambiguous inputs (a worker in full gear, a figure seen from behind), cases common in practice yet rarely studied. We find that minimal prompting pressure exposes occupation-gender defaults when prompting ambiguous input images, with models collapsing to male even for strongly female-stereotyped occupations. But do these outputs reflect what models actually encode internally? We introduce LALS (Latent Association Leaning Score), a zero-shot metric that projects visual-token activations into the model's text-embedding space to measure concept associations per token and layer. Across 15 occupations, over 800 gender-ambiguous images, and four VLMs, internal representations and outputs often become systematically decoupled: models often encode a female association internally yet output male. Layer-wise analysis reveals an asymmetric filter: male signal amplifies end-to-end while female signal peaks mid-network and is suppressed before generation. A color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.

Arnau Marin-Llobet, Simon Henniger, Mahzarin R. Banaji

School of Engineering and Applied Sciences · Department of Psychology

VLM Evaluation VLM Interpretability Implicit Bias

May 29, 2026cs.CV

The Regularizing Power of Language-Training Deepfake Detectors

Recently, thanks to the advent of Multimodal-LLMs, deepfake detectors are striving not only to be generalizable but also interpretable. We propose that these two challenges can effectively be tackled jointly, since describable artifacts typically generalize better, opening the possibility to use language as a regularization mechanism. Since deepfake detection generally suffers from overfitting to low-level domain-specific artifacts, our intuition is that an LLM that has been pretrained on language would prefer high-level artifacts that can be described better. This way, we can use high-level features where possible, while training the model to use low-level features where necessary. We utilize a dual-encoder architecture, pairing a frozen specialist detector with a LoRA-tuned MLLM encoder, and a two-stage training curriculum: first, a binary alignment phase demonstrates that the intrinsic capability of MLLMs can effectively combine features to mitigate overfitting to dataset-specific artifacts. To further bolster generalization and achieve interpretability, we employ a reinforcement learning stage that encourages the model to generate descriptive reasoning before classifying, using only binary labels. By rewarding this "explain-then-classify" behavior, we explicitly incentivize the model to prioritize high-level, robust features. Crucially, this process yields both interpretable descriptions and a further boost in cross-dataset performance, even when reasoning chains are omitted at inference. Extensive experiments on benchmark datasets validate our approach, outperforming state-of-the-art methods by a large margin.

Benedikt Hopf, Zongwei Wu, Radu Timofte

Computer Vision Lab, CAIDAS, University of Würzburg, Germany

VLM Robustness Explainable Deepfake Detection Multimodal Large Language Models VLM Interpretability Deepfake Detection

May 28, 2026cs.CV

Why Far Looks Up: Probing Spatial Representation in Vision-Language Models

Vision-language models (VLMs) achieve strong performance on spatial reasoning benchmarks, yet it remains unclear whether this reflects structured 3D understanding or reliance on statistical shortcuts in natural images. We introduce a representation-level analysis framework that constructs minimal contrastive pairs to measure how spatial axes are organized and disentangled within VLM embeddings. Our analysis across multiple model families reveals a consistent vertical-distance entanglement: models conflate vertical image position with distance, mirroring the perspective bias of natural photographs. This bias produces a significant accuracy gap between perspective-consistent and counter-heuristic examples, and intensifies under data scaling even as overall benchmark accuracy improves. We further show that models with similar benchmark scores can exhibit different internal representations, and that these differences predict accuracy and robustness across diverse spatial reasoning benchmarks. To isolate this bias from evaluation-set skew, we introduce SpatialTunnel, a synthetic benchmark designed to expose spatial shortcut biases by removing common correlations present in natural images. Experiments confirm that the entanglement is model-intrinsic, and that models with well-separated spatial axes exhibit greater robustness, suggesting that well-structured spatial representations lead to more reliable spatial reasoning across diverse benchmarks. Code and benchmark are available on the project page: https://cheolhong0916.github.io/whyfarlooksup.github.io/.

Cheolhong Min, Jaeyun Jung, Daeun Lee +5

Seoul National University · The Ohio State University · NVIDIA

VLM Evaluation Spatial Reasoning Benchmarks Representation Learning Visual Spatial Reasoning VLM Interpretability

May 27, 2026cs.CV

VITAL: Visual-Semantic Dual Supervision for Enhanced and Interpretable Latent Reasoning in Medical MLLMs

Latent reasoning enables reasoning over continuous hidden states rather than explicit tokens, avoiding the language bottleneck and inference overhead of chain-of-thought for medical VQA. However, existing methods suffer from modality collapse, insufficient visual supervision, and train-inference mismatch. Moreover, their opaque latent states offer no interpretability, which is critical in clinical applications. We propose VITAL, a latent-space reasoning framework for medical MLLMs with visual-semantic dual supervision: an auxiliary text decoder reconstructs reasoning chains from latent states, while a visual projector regresses ROI features from a frozen, independent medical vision encoder. Both modules are discarded at inference with zero overhead, yet can be re-attached post-hoc for dual interpretability, providing textual and visual explanations of the reasoning process without sacrificing efficiency. We construct a 61K dataset spanning 9 imaging modalities, exceeding prior medical visual latent reasoning datasets by an order of magnitude. Experiments on 7 benchmarks show that VITAL consistently and substantially outperforms the backbone, all latent reasoning baselines, and medical MLLMs trained on far larger data, achieving state-of-the-art results competitive with trillion-parameter proprietary models.

Qiaoru Li, Shaotian Liang, Jintao Chen +4

3Tencent · 1Zhejiang University · 2Shanghai AI Laboratory +1

Medical VQA VLM Interpretability Latent Visual Reasoning Medical VLMs Explainable Medical Image Analysis

May 27, 2026cs.AI

Explaining is Harder Than Predicting Alone: Evaluating Concept-based Explanations of MLLMs as ICL Visual Classifiers

In-context learning (ICL) enables multimodal large language models (MLLMs) to classify images from a few labelled examples. Yet, how these models use the provided context remains opaque. While Chain-of-Thought prompting is widely used, recent work argues that it may not reflect true internal computation. In this paper, we systematically evaluate the concept-based explainability of frozen MLLMs under few-shot ICL using five conditions of increasing formal rigour, ranging from baseline classification to Description Logics (DL) axiom generation. Evaluating four state-of-the-art MLLMs via an independent LLM-as-a-judge pipeline, we demonstrate that explaining is genuinely harder than predicting alone. Surprisingly, forcing models to generate formally structured, concept-based explanations degrades predictive accuracy monotonically (from 93.8% to 90.1%), contradicting the assumption that explicit reasoning universally aids performance. However, when models successfully articulate class-discriminative visual features, explanation quality strongly correlates with correct predictions. Our findings suggest that while MLLMs excel at visual classification, they lack the specific instruction-tuning required for formal, machine-verifiable explainability.

Carmen Quiles-Ramírez, Leticia L. Rodríguez, Nicolás Martorell +1

Universidad de Granada, Spain · Universidad de Buenos Aires, Argentina

VLM Evaluation Multimodal Large Language Models VLM Interpretability Concept-Based Explanations In-Context Learning Explainable Image Classification+1

May 26, 2026cs.CL

Real Images, Worse Judgments: Evaluating Vision-Language Models on Concreteness and Imagery

Visual inputs are often assumed to improve language understanding in multimodal models. We examine this assumption by asking whether vision-language models (VLMs) can distinguish useful visual evidence from incidental image context in lexical judgments. We use human concreteness and imagery ratings because they span words with varying expected visual relevance, from abstract and low-imagery words to concrete and high-imagery words. We find that real-image contexts do not yield consistent gains and often hurt alignment with human ratings, most sharply when visual evidence is least relevant. Through probing and canonical correlation analysis, complemented by an attribution case study, we find that real-image contexts are associated with representational shifts and greater sensitivity to spurious visual cues, coinciding with weaker recoverability of the targeted lexical properties. We further show that instructing models to focus solely on textual content at inference time can reduce this degradation, with the clearest gains on these vulnerable subsets. Our findings suggest that current instruction-tuned VLMs need better calibration of when visual context should inform lexical judgments.

Yifan Jiang, Ruoxi Ning, Sheng Yao +1

University of Waterloo · Vector Institute

VLM Evaluation Vision-Language Models VLM Interpretability

May 26, 2026cs.CV

Can Retrieval Heads See Images? Multimodal Retrieval Heads in Long-Context Vision-Language Models

Large vision-language models increasingly rely on long-context modeling to reason over documents, hour-level videos, and long-horizon agent trajectories, requiring them to locate relevant evidence across interleaved text and images. Prior work has studied this behavior using retrieval heads in large language models, but its copy-based criterion does not directly apply when evidence appears in images. We introduce a multimodal retrieval head detection method that scores attention from question tokens to textual or visual evidence. With this method, we show that multimodal retrieval heads are sparse, intrinsic, and causally important: only 4.4-10.2% of attention heads account for 50% of the positive retrieval-score mass, and masking the top-5% selected heads reduces performance on two tasks from MMLongBench by 42.5 and 62.3 percentage points, while random-head masking is far less damaging. Further analysis shows that these heads are partly shared across modalities yet remain dynamic within each modality, with image retrieval heads changing more than text retrieval heads as context length and haystack modality change. Without further training, we find that these heads can also be used directly to rank visually rich documents: on MMDocIR, Qwen3-VL-8B selected-head scoring improves Recall@1 by 7.7/7.4 macro/micro points for page retrieval and 6.3/6.8 points for layout retrieval over the strongest reported baseline.

Aaron Branson Cigres Li, Zhaowei Wang, Yu Zhao +9

1HKUST · University of Edinburgh · 4CUHK +2

Vision-Language Models VLM Interpretability Vision-Language Retrieval Visual Document Retrieval Multimodal Foundation Models Long-Context VLMs+1

May 25, 2026cs.CV

Binding Visual Features Point by Point

Despite success on standard benchmarks, vision language models display persistent failures on tasks involving processing of multi-object scenes, including many tasks that are relatively easy for humans. Recent work has found that these failures may stem from a basic inability to accurately bind object features in-context, a challenge that is referred to as the "binding problem" in cognitive science and neuroscience. The human visual system is thought to solve this binding problem via serial processing, attending to individual objects one at a time so as to avoid interference from other objects. Recent work has proposed "pointing" -- the use of explicit spatial coordinates to refer to objects -- as an analogous solution for vision language models, and found that it improves performance on challenging multi-object tasks. However, it is unclear

\textit{why}

(i.e., on a mechanistic or representational level) this approach improves performance, and how directly this relates to serial processing in human vision. Here, we investigate this question. We find that learning to point-via-text induces an internal visual search routine, and we characterize the mechanisms that support this procedure. We also find that pointing behavior can be generalized to new tasks via fine-tuning, and that doing so eliminates binding errors and enables compositional generalization. These results provide a proof-of-principle that serial processing can solve the binding problem for vision language models just as it does for biological vision.

Udith Haputhanthri, Declan Campbell, Rim Assouel +2

Princeton University · Mila – Quebec AI Institute · Université de Montréal

Vision-Language Models VLM Interpretability Vision-Language Grounding Visual Search

May 24, 2026cs.CV

Interpretability Transfer from Language to Vision via Sparse Autoencoders

Recent advances in language model interpretability using sparse autoencoders (SAEs) have yet to effectively translate to the visual domain, mainly due to the difficulty and ambiguity of labeling visual concepts. In this paper, we introduce Visual Interpretability via SAE Transfer Alignment (VISTA), a framework that transfers interpretability from language to vision in a LLaVA-style vision-language model by constraining a visual projector to map visual tokens into an LLM's pre-existing, labeled textual SAE space. This approach enables visual interpretability without training dedicated vision SAEs. By regularizing the projector using the LLM's SAE reconstruction loss, VISTA achieves a threefold increase in the matching rate, which measures how accurately the most activating textual concepts in the SAE space correspond to semantic elements in the image. Using this framework, we further analyze spatial localization properties of different vision encoders and show that DINOv2 features have stronger localization abilities than other encoders. Leveraging this precision, we validate VISTA's cross-modal alignment through fine-grained, localized concept interventions, where specific objects are removed or replaced in the model's perception while preserving the surrounding scene. This results in improvements of 35% in object removal and 47% in object replacement tasks over vision-only baselines, providing causal evidence that visual tokens inhabit the text SAE manifold. These contributions are validated across multiple LLM architectures.

Alexey Kravets, Da Li, Chuan Li +2

VLM Interpretability Sparse Autoencoders Vision-Language Alignment

May 21, 2026cs.CV

Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?

Benchmark accuracy is often implicitly assumed to reflect grounded visual understanding in vision-language models (VLMs), yet it remains unclear to what extent such scores truly reflect reliance on visual evidence. Motivated by a surprising observation that removing a substantial fraction of image tokens only degrades model performance very slightly on a widely used hallucination benchmark, we systematically investigate this mismatch in a set of open-source VLMs. Our analysis spans multiple levels of granularity, spanning global visual degradation, localized occlusion, question reformulation, answer-space expansion, and decision-level analyses beyond standard accuracy. We further complement these behavioral results with a layer-wise analysis of vision-token geometry. Throughout the experiments, we find that although VLMs do incorporate visual input, their predictions are less sensitive to the loss of fine-grained visual evidence that standard accuracy should have suggested. Even when the final prediction remains unchanged, the model's internal support for the correct answer may already be weakened. We further complement a representation-level analysis, which shows increasing similarity among visual tokens in deeper layers, providing a possible explanation for our findings. Together, these results suggest that current benchmarks are not sufficient to reliably evaluate fine-grained visual grounding in VLMs.

Zixuan Lan, Luzhe Sun, Matthew R. Walter +1

University of Chicago · Toyota Technological Institute at Chicago · Stony Brook University

VLM Evaluation VLM Interpretability Vision-Language Grounding Visually Grounded Reasoning Fine-Grained Visual Perception

May 21, 2026cs.LG

Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models

Generative Vision-Language Models (VLMs) perform well on multimodal reasoning, but how visual inputs are transformed to text remains poorly understood. Existing interpretability work on VLMs uses Sparse Autoencoders (SAEs), which decompose static residual representations and miss the functional updates that drive cross-modal interaction. We adopt a function-centric framework based on Transcoders, sparse approximations of MLP sublayers that act as a causal proxy for layer-wise computation. Applied to Gemma 3-4B-IT, the framework decomposes the model into interpretable computational pathways linking image patches to directions in token generation. Transcoder attributions produce stronger and more stable effects on visually grounded tokens under patch ablation than SAE attributions, and align better with semantically relevant image regions. A False Visual Grounding counterfactual analysis confirms that the recovered pathways are specific to vision-language interaction.Finally, we perform a structural analysis of hallucinated generations, by extracting graph-based indicators from circuit traces produced by the transcoders. A logistic classifier over these mechanistic graph features predicts hallucinations at AUC

0.68

. These results show that function-centric circuit decomposition yields interpretable and predictive accounts of multimodal computation in VLMs.

Dimitrios Damianos, Leon Voukoutis, Georgios Skyrianos +2

Institute of Language and Speech Processing, Athena Research Center Athens, Greece

Object Hallucination in VLMs VLM Interpretability Mechanistic Interpretability Vision-Language Grounding Multimodal Model Interpretability

May 21, 2026cs.CV

Conceptualizing Embeddings: Sparse Disentanglement for Vision-Language Models

Vision-language models learn powerful multimodal embeddings, yet their internal semantics remain opaque. While sparse autoencoders (SAEs) can extract interpretable features, they rely on expanding the representation dimension, which compromises the original geometry and introduces redundancy. We introduce CEDAR (Conceptual Embedding Disentanglement via Adaptive Rotation), a post-hoc method that reveals the compositional structure of pretrained embeddings without increasing dimensionality. By learning an invertible transformation with a top-

k

sparsity bottleneck, CEDAR concentrates semantic information into axis-aligned disentangled coordinates. In CLIP-like architecture, individual coordinates can be interpreted with textual concepts, while for generative models such as BLIP, they can be decoded into natural language descriptions. Experiments demonstrate that CEDAR achieves a competitive reconstruction-sparsity trade-off while producing explanations that are more interpretable and better aligned with human perception. Our results suggest that the apparent entanglement in vision-language representations can be resolved through a suitable change of basis, eliminating the need for overcomplete expansions.

Piotr Kubaty, Patryk Marszałek, Łukasz Struski +3

Faculty of Mathematics and Computer Science, Jagiellonian University, Kraków, Poland · Doctoral School of Exact and Natural Sciences, Jagiellonian University, Kraków, Poland

Disentangled Representation Learning Multimodal Disentangled Representation Learning VLM Interpretability Sparse Autoencoders Multimodal Embedding

May 21, 2026cs.CV

SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation

While large language models provide strong compositional reasoning, existing reasoning segmentation pipelines fail to transparently connect this reasoning to visual perception. Current methods, such as latent query alignment, are end-to-end yet opaque "black boxes". Conversely, textual localization readout is merely readable, not truly interpretable, often functioning as an unconstrained post-hoc step. To bridge this interpretability gap, we propose SegCompass, an end-to-end model that leverages a Sparse Autoencoder (SAE) to forge an explicit, interpretable, and differentiable alignment pathway. Given an image-instruction pair, SegCompass first generates a chain-of-thought (CoT) trace. The core of our method is an SAE that maps both the CoT and visual tokens into a shared, high-dimensional sparse concept space. A query codebook selects salient concepts from this space, which are then spatially grounded by a slot mapper into a multi-slot heatmap that guides the final mask decoder. The entire model is trained jointly, unifying reinforcement learning for the reasoning path with standard segmentation supervision. This SAE-driven interface provides a "white-box" connection that is significantly more traceable than latent queries and more coherent than textual readouts. Extensive experiments on five challenging benchmarks demonstrate that SegCompass matches or surpasses state-of-the-art performance. Crucially, our visual and quantitative analyses show a strong correlation between the quality of the learned sparse concepts and final mask accuracy, confirming that SegCompass achieves superior results through its enhanced and inspectable alignment. Code is available at https://github.com/ZhenyuLU-Heliodore/SegCompass.

Zhenyu Lu, Liupeng Li, Jinpeng Wang +4

Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Peng Cheng Laboratory · University of Chinese Academy of Sciences +3

Image Segmentation Text-Guided Image Segmentation VLM Interpretability Visually Grounded Reasoning Vision-Language Alignment Reasoning Segmentation

May 21, 2026cs.AI

Measuring Cross-Modal Synergy: A Benchmark for VLM Explainability

Vision-Language Models (VLMs) map complex visual inputs to semantic spaces, but interpreting the cross-modal reasoning of VLMs currently relies on post-hoc explainers evaluated via unimodal perturbation metrics. We expose a limitation in this paradigm: because multimodal datasets contain language priors and modality biases, VLMs frequently exhibit cross-modal redundancy, allowing them to answer visual queries using text alone. Consequently, unimodal metrics penalize faithful explainers, triggering an evaluation collapse where visual and textual rankings fundamentally contradict each other. %(Kendall's

τ= -0.06

). To resolve this, we introduce Synergistic Faithfulness (

\mathcal{F}_{syn}

), a scalable metric rooted in the Shapley Interaction Index that strictly isolates the joint Harsanyi dividend between modalities, serving as a highly accurate surrogate (

ρ= 0.92

) while achieving a

24\times

computational speedup. Evaluating 8 distinct XAI methods across 3 VLM architectures and 3 benchmark datasets, reveals that explainers proposed for VLMs heavily over-index on visual salience and significantly underperform adapted attention-based methods in capturing true cross-modal synergy. By decoupling visual plausibility from cross-modal faithfulness, this work provides a rigorous evaluation framework required to safely audit VLM reasoning in high-stakes deployments.

Joël Roman Ky, Salah Ghamizi, Maxime Cordy

University of Luxembourg, Luxembourg · Luxembourg Institute of Health (LIH), Luxembourg

VLM Interpretability Shapley Value Attribution Multimodal Model Interpretability Explainability Evaluation Multimodal Understanding

May 21, 2026cs.CV

Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow

Large Vision-Language Models (LVLMs) represent a significant leap towards empathetic agents, demonstrating remarkable capabilities in emotion understanding. However, the internal mechanisms governing how LVLMs translate abstract visual stimuli into coherent emotional narratives remain largely unexplored, primarily due to the scarcity of visual counterfactuals and the diffuse nature of emotional expression. In this paper, we bridge this gap by introducing a steering-vector-based causal attribution framework tailored for descriptive emotional reasoning. To this end, we construct a specialized dataset to demystify the emotional circuits underlying the three-stage ``Adapt-Aggregate-Execute'' mechanism. Crucially, we discover a functional decoupling: visual emotional cues are aggregated in middle layers via sentiment-specific attention heads, but are subsequently translated into narrative generation in deep layers through emotion-general pathways. Guided by these insights, we regulate the emotional information routing to strengthen attention flow and amplify the semantic activation to consolidate expression. Extensive experiments on the comprehensive MER-UniBench demonstrate that our methods significantly improve performance via inference-time intervention, effectively mitigating emotional hallucinations and corroborating the causal fidelity of the discovered circuits.

Chengsheng Zhang, Chenghao Sun, Zhining Xie +1

MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China · AIPD, Tencent

VLM Hallucination VLM Interpretability Multimodal Emotion Recognition VLM Reasoning

May 20, 2026cs.CV

Ablate-to-Validate: Are Vision-Language Models Really Using Continuous Thought Tokens?

Vision-language models (VLMs) are increasingly augmented with continuous or latent non-textual tokens intended to support "visual thinking." Despite improved task accuracy, this alone does not show that models actually use these tokens for reasoning -- gains may arise from confounds such as added context length, special-token anchoring, or training-time regularization. We formalize a diagnostic principle, Ablate-to-Validate, for testing whether latent-token content is genuinely utilized, and instantiate it as the Token Replacement Test (TRT), a standardized suite of content-replacement ablations. TRT holds the prompt, image, token budget, and decoding fixed while replacing intermediate tokens with zero, random, first-repeat, or oracle alternatives, isolating whether performance depends on token content or merely on token presence. As a controlled testbed, we study relative depth reasoning with LLaVA-13B and Qwen2.5-VL-3B, training models to predict and consume continuous or discrete depth spans across multiple frozen encoders (SigLIP2, CLIP, DINOv2) and token budgets. We additionally apply TRT to three off-the-shelf visual-thinking systems (Mirage, Mull-Tokens, CoVT) on BLINK, VSP, and CV-Bench. Across all settings, accuracy gains are a misleading proxy for latent-token reasoning: VLMs retain most improvement even when token content is corrupted or replaced, revealing a persistent gap between having a latent channel and using it as an information bottleneck. We recommend TRT as a standard diagnostic alongside accuracy for any method introducing continuous thought tokens.

Tianyi Zhang, Mahtab Bigverdi, Ranjay Krishna

University of Washington

VLM Evaluation VLM Interpretability Latent Visual Reasoning VLM Reasoning Faithfulness in VLMs

May 19, 2026cs.CV

Do Vision-Language Models Understand 3D Scenes or Just Catalogue Objects?

Vision-language models reliably name objects in a scene, but do they represent the 3D layout those objects inhabit? We introduce a 3,034-sample human-curated benchmark targeting three components of spatial understanding: depth-ordered occlusion (probed via three independent counterfactual operationalisations), optical-geometry inference over visible reflections, and volumetric rearrangement planning. Six frontier and open-weight VLMs, scored by trained annotators on 18,204 responses with no LLM-as-judge, reveal a sharp dissociation: models that plan rearrangements over visible layouts at 53--97% accuracy and rarely violate collision constraints fall to 6--45% on occlusion and below 7% on reflections. An embodied-reasoning model reproduces the same profile. White-box analysis on Qwen3-VL-8B-Thinking localises the failure to the visual-token merger: spatial information recoverable throughout the vision encoder becomes inaccessible after token compression and only stabilises again when clean post-merger activations are patched into the language decoder.

Animesh Maheshwari, Divyansh Sahu, Nishit Verma

Deccan AI

VLM Evaluation Spatial Reasoning Benchmarks Occlusion Reasoning 3D Spatial Reasoning VLM Interpretability Visually Grounded Reasoning

May 19, 2026cs.CV

Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models

Large Vision Language Models (LVLMs) show promise in medical applications, but their inability to faithfully ground responses in visual evidence raises serious concerns about clinical trustworthiness. While visual attribution methods are widely used to explain LVLM predictions, whether these explanations actually reflect the visual evidence underlying the model's decision is largely unverified, since ground-truth annotations for internal model reasoning are typically unavailable. We address this question for chest X-ray (CXR) reasoning by developing a causal evaluation framework that retains only CXR-VQA samples for which the expert-annotated region is verified, via counterfactual editing, to be causally responsible for the model's prediction. Using this framework across 11 attribution methods, six open-source LVLMs, and two output modes (direct answer and step-by-step reasoning), we find that existing attribution methods often fail to identify the evidence used by LVLMs. To address this failure, we propose MedFocus, a concept-based attribution method that localizes clinically meaningful anatomical regions via unbalanced optimal transport and measures their causal effect on model outputs through targeted interventions. MedFocus produces spatial, concept-level, and token-level attributions and substantially outperforms prior methods, taking a step toward more trustworthy attribution for medical LVLMs. Our data and code are available at https://github.com/gzxiong/medfocus/.

Guangzhi Xiong, Qiao Jin, Sanchit Sinha +2

University of Virginia · National Institutes of Health

Medical VQA VLM Interpretability Concept-Based Explanations Medical VLMs Explainable Medical Image Analysis Causal Interpretability

May 19, 2026cs.CV

Mechanisms of Object Localization in Vision-Language Models

Visually-grounded language models (VLMs) are highly effective in linking visual and textual information, yet they often struggle with basic classification and localization tasks. While classification mechanisms have been studied more extensively, the processes that support object localization remain poorly understood. In this work, we investigate two representative families, LLaVA-1.5 and InternVL-3.5, using a suite of mechanistic interpretability tools, including token ablations, attention knockout, and causal mediation analysis. We find that localization is driven by a containerization mechanism in which object-aligned tokens define the spatial extent of the object, while the semantic arrangement of tokens within those boundaries is largely irrelevant to the predicted box. Only a very small set of attention heads mediates the causal effect for both classification and localization, concentrating in early-mid layers for LLaVA and mid-late layers for InternVL. The two tasks share some early processing but ultimately depend on largely distinct specialized heads. Overall, we provide the first layer- and head-level account of localization in VLMs, revealing narrow computational pathways that can guide future model design and grounding objectives.

Timothy Schaumlöffel, Martina G. Vilas, Gemma Roig

Goethe University Frankfurt, Germany · The Hessian Center for AI, Germany

Vision-Language Models VLM Interpretability Vision-Language Grounding

May 19, 2026cs.AI

Causal Evidence for Attention Head Imbalance in Modality Conflict Hallucination

Modality-conflict hallucination occurs when multimodal large language models (MLLMs) prioritize erroneous textual premises over contradictory visual evidence. To understand why visual evidence fails to prevail during generation, we take a mechanistic perspective and examine which internal components drive or resist this failure. We perform head-level causal analysis using path patching across five open-source MLLMs and identify two groups of attention heads with opposing causal roles: hallucination-driving heads and hallucination-resisting heads. We find a consistent asymmetry: driving effects are more broadly distributed and carry greater aggregate weight, whereas resisting effects concentrate in a small number of high-importance heads. Ablation experiments further confirm that these groups exert opposing effects during generation: distributed driving influence and localized resistance together form an imbalanced routing structure that biases generation toward the erroneous premise. Motivated by this finding, we propose MACI (Modality-conflict-Aware Causal Intervention), a conditional intervention that suppresses causally identified hallucination-driving heads only when conflict is detected. Across five MLLMs, MACI achieves the largest hallucination reduction among compared inference-time baselines on the MMMC benchmark with a favorable hallucination-accuracy trade-off, and transfers zero-shot to the SCI-SemanticConflict test.

Jinrui Jiang, Zhangtai Wu, Zhen Wu +1

National Key Laboratory for Novel Software Technology, Nanjing University, China · School of Artificial Intelligence, Nanjing University, China

Hallucination in Language Models Multimodal Large Language Models VLM Interpretability Multimodal Hallucination Causal Interventions in Language Models Hallucination Mitigation+1

May 17, 2026cs.CV

Employing Vision-Language Models for Face Image Quality Assessment

Face Image Quality Assessment (FIQA) is a crucial control step in biometric pipelines. It ensures only reliable samples are processed to maintain system accuracy. State-of-the-art FIQA methods achieve high utility but typically operate as "black boxes." They produce scalar scores without human-interpretable justifications. This lack of transparency limits their effectiveness in human-in-the-loop scenarios, such as automated border control, where actionable feedback is essential. In this paper, we investigate the potential of off-the-shelf Vision-Language Models (VLMs) to bridge this gap by performing FIQA in a zero-shot setting. We present a comprehensive evaluation framework for assessing VLM performance. This involves benchmarking traditional FIQA methods through error-versus-reject curves. Additionally, using a diverse set of datasets, ranging from surveillance-oriented to synthetically generated, we analyzed their interpretability, consistency, and robustness to prompt changes. Our results show biometric utility performance depends significantly on architecture, not merely on parameter count. Most VLMs' outputs align with those of traditional methods. We also find that VLM ranking performance and the generated scores may vary across prompts. Our synthetic ablation study shows that while increasing the parameter count can improve internal consistency, it yields worse degradation-detection performance than smaller models. These findings suggest that zero-shot FIQA score estimation using VLMs is promising and could effectively complement conventional FIQA pipelines as an interpretability module. The codes are available at https://github.com/ThEnded32/VLM4FIQA.git.

Erdi Sarıtaş, Eren Onaran, Vitomir Štruc +1

Department of Computer Engineering, Istanbul Technical University, Istanbul, Türkiye · Faculty of Electrical Engineering, University of Ljubljana, Ljubljana, Slovenia · Division of Engineering, NYU Abu Dhabi, Abu Dhabi, UAE

VLM Evaluation VLM Robustness VLM Interpretability Image Quality Assessment