Vision-language models (VLMs) are increasingly evaluated on complex image and video understanding tasks, yet conventional metrics primarily assess final-answer quality and reveal little about how different information sources shape the generation process. We propose a causal and temporal evaluation framework that traces the evolving roles of visual input, question text, and generated prefixes during autoregressive decoding. Grounded in a Structural Causal Model, we use interventions and backdoor adjustment to derive three step-indexed causal-drive metrics---Visual Causal Drive (VCD), Question Causal Drive (QCD), and Prefix Causal Drive (PCD)---for characterizing source-specific generation patterns without requiring reference answers. Experiments on Qwen3-VL-8B-Instruct across MAVIS, LLaVA-Video-178K, and MiraData, together with cross-model validation on InternVL2-8B, reveal a consistent transition from stronger early question and visual guidance toward increasing reliance on generated prefixes. Randomized-intervention validation shows that QCD and PCD reduce recovery error over observational PMI baselines by 34.8\% and 47.1\%, respectively. On VLMBias, the prefix--visual imbalance score achieves 0.767 AUROC and 0.873 AUPRC for distinguishing prior-driven from visually grounded generations. These results show that causal-drive trajectories provide complementary source-level diagnostics for multimodal generation.
Whether attention weights faithfully reflect model reasoning has been actively debated in NLP, yet this question remains largely unexplored for the visual modality in Vision-Language Models (VLMs). We address this gap through causal perturbation analysis on current VLMs, evaluating both the comprehensiveness and sufficiency gap of attention-ranked visual tokens. Our analysis reveals that visual attention faithfulness is heterogeneous, manifesting in three distinct processing modes: Faithful-Sufficient, where top-$k$ attention tokens are both necessary and sufficient for prediction; Faithful-Distributed, where they are necessary but broader visual context remains required; and Non-Focal, where no localized attention region is individually necessary while visual information remains an essential trigger for prediction. Furthermore, human-annotated ground-truth regions satisfy comprehensiveness in only $\sim 60$% of cases compared with model attention rankings, revealing systematic divergence between model visual reliance and human intuition. We demonstrate these patterns across both general VQA on VQAv2 and document tasks on VRDU and ChartQA, showing that visual attention faithfulness varies systematically with processing demands and model architectures rather than being uniformly faithful or unfaithful.
Language models learn about grammatical number primarily from co-occurrence, and show frequency effects as a result---sometimes taken to indicate that they do not learn abstract ``rules'', and are instead dependent on specific lexical items. Testing generalization with text stimuli alone cannot settle this debate, since distributional cues (is/are, this/these) easily give number away. We instead use cross-modal generalization as a tool to investigate abstractions in LMs that can also accept visual inputs (VLMs), restricting the evidence that diagnoses number to an extra-linguistic modality. We teach VLMs pairs of new nouns by adding new embeddings and only updating them during learning, comparing conditions where number is diagnosed by visual cues alone against ones where it is disambiguated by text. Across behavior, representational dynamics, and causal mechanisms, we find non-trivial evidence for cross-modal generalization across both exposure conditions, and that linguistic vs. extra-linguistic cue conditions are treated in similar ways in the internal mechanisms of the model. This suggests that statistical learners like VLMs can generalize beyond surface-level co-occurrence and show genuine abstraction-compatible behavior.
A model's behavior on a task is jointly determined by the input it receives and the prior it brings in, i.e. the distribution over stimuli it implicitly expects. Interpretability research has traditionally studied models by holding inputs fixed and examining model responses either mechanistically, probing how internal structure represents inputs, or behaviorally, measuring how variation in inputs leads to variation in outputs. Neither reconstructs the prior distribution itself, since internal structure shows what a model can represent, not what it expects, and any fixed stimulus set leaves most of the possible input space unseen. In particular, such an input space in real-world settings, such as images seen by VLMs, is extremely high-dimensional and diverse. These priors thus remain a poorly understood component of models that nonetheless influence real-world behavior. We propose a method to sample from models' perceptual prior distributions directly, by steering a generative model to produce stimuli along controllable axes and running Gibbs sampling over that space with the model under study as the judge. We apply this to a variety of categories and target variables (such as trustworthiness in faces and cheapness in art images) and recover both canonical biases and surprising novel priors invisible to direct prompting, warranting further investigation of their downstream effects.
Vision-language models (VLMs) can locate an image region referred to by a text prompt and route the corresponding visual evidence to the output, yet the internal mechanism behind this behavior is not understood. Inspired by retrieval heads in large language models, we ask whether VLMs contain an analogous mechanism for visual retrieval. We answer affirmatively by introducing Visual Retrieval Heads (VRHs), a small subset of attention heads (about 1.7-2.6%) that are causally responsible for grounding text descriptions to image regions. To find them, we recast existing head-scoring methods under a unified design space over query tokens, key aggregation, and cross-sample aggregation. We then show that scoring attention from output prediction tokens with a sum over the ground-truth referent region most reliably identifies causal heads. Across eleven VLMs and five referring-expression benchmarks, masking only the top 20 VRHs reduces grounding accuracy by up to 80 percentage points, while masking the same number of random heads has little effect. Beyond replicating the causal-sparse-universal triad established for text retrieval heads, VRHs exhibit several properties not previously reported: they generalize across visual reference tasks, remaining causal on attribute, spatial, counting, and visual-math benchmarks despite being discovered through bounding-box prediction; they are functionally specific, preserving output format while corrupting localization; and they are architecturally shared, transferring causally across VLMs that share an LLM backbone but differ in vision encoder, projector, and instruction tuning.
Vision-language models such as LLaVA-1.5-7B often hallucinate objects absent from the image when generating captions. We ask whether an interpretability diagnosis of this failure can guide a targeted fix, and we measure what that fix actually changes. We rank attention heads by how much their image attention drops around hallucinated object words, then screen the shortlist by ablating candidate heads and measuring the change in hallucination-token log probability, yielding a 32-head set. We restrict two interventions to these heads: a head-sliced LoRA adapter and an inference-time grounding controller. On 400 held-out COCO images, the combined method lowers CHAIRs (the fraction of captions with a hallucinated object) from 0.370 to 0.230 and CHAIRi (the fraction of hallucinated object mentions) from 0.156 to 0.096 (p < 0.001, paired sign-flip tests). Two controls sharpen attribution. A random-head LoRA control, matched layer-for-layer and trained identically, performs no better than the matched baseline on a separate 200-image control split, supporting the role of head selection rather than LoRA capacity. Under fixed decoding budgets, the CHAIR reduction persists and grows with budget (23% at 64 tokens to 58% at 128), arguing against a pure max-token or truncation artifact, although the method remains shorter and more conservative. The resulting behavior reduces unsupported object mentions while also lowering object recall (0.78 to 0.70). We present a diagnosis-to-intervention pipeline for object hallucination, and, more importantly, a controlled account of what acting on the diagnostic signal actually does: it localizes intervention sites with real, non-random leverage, reported as a behavioral profile rather than a single score.
Prompt learning adapts vision-language models such as CLIP by optimizing continuous prompt vectors, but the learned prompts are difficult to interpret in natural language. We present PromptSpLiCE, a post-hoc method that expresses each class-conditioned text embedding as a sparse combination of concepts from a fixed natural-language dictionary. Using the same dictionary before and after prompt learning allows us to compare changes in their concept profiles. We evaluate PromptSpLiCE on CoOp, a representative prompt-learning method, across 11 image-classification datasets. The concept profiles change substantially: on average, only 1.6 of the initial top-10 concepts remain in the top 10 after learning. Across datasets, profile change is positively associated with accuracy gain. We also derive a local gradient expression that provides geometric intuition for why image-aligned concept directions distinct from the current prompt can have greater loss sensitivity.
Vision-Language Models (VLMs) achieve strong performance in visual reasoning tasks, but it remains unclear whether they understand visual relations, or simply employ shortcuts such as language cues or priors. To investigate this, we use the Qwen3-VL-4B (Bai et al., 2025), a modern VLM, to decode how visual information is encoded across depths. For this, we propose a synthetic dataset of simple geometric shapes for controlled analysis, along with queries crafted to precisely test language cues. Furthermore, the dataset is modified to test causal reliance on visual evidence. Our results show that current VLMs combine genuine visual reasoning with shortcut strategies primarily rooted in language cues.
Video understanding requires intelligent agents to transcend mere recognition of visual facts and comprehend the underlying intents behind human actions (often termed the "dark matter" of social intelligence). To bridge the gap between visual observation and intent reasoning, we introduce a novel task, IntentQA, and contribute a large-scale VideoQA dataset specifically tailored for this purpose. However, recognizing that standard metrics may overestimate capabilities due to dataset biases, we go beyond simple accuracy to rigorously evaluate model robustness. We augment the benchmark by generating five distinct contrast sets via Large Language Models (LLMs) and introducing a "Contrast Performance Decline" metric. We propose the X-CaVIR (eXplainable Context-aware Video Intent Reasoning) framework, which leverages three types of "Cognitive Context" to enhance video analysis: i) Situational Context via a cross-modal Video Query Language (VQL) module, ii) Contrastive Context via a Contrastive Learning module, and iii) Commonsense Context via a Commonsense Reasoning module. Crucially, to overcome the opacity of traditional black-box models, we refine the integration of LLMs within X-CaVIR by employing a transparent pipeline that synergizes video captions with VQA model outputs. This approach not only improves performance by effectively utilizing rich commonsense knowledge but also renders the reasoning process explicitly interpretable. Extensive experiments demonstrate the effectiveness of our components, the superiority of X-CaVIR over state-of-the-art baselines, and its stability against perturbations on the contrast sets.
Vision-language models are known to encode spatial information in their hidden states, yet often fail to use it when answering. However, it remains unclear when and where this encoded information reaches the answer. We address this with direction patching, a class-conditioned causal intervention applied across layers, token positions, and prompt formats. Using spatial-ID directions constructed following prior encoding evidence, we find that causal influence on answer logits emerges only at mid-to-deep depths. Text chain-of-thought suppresses immediate object-word argmax-level transport in most models, while visually grounded prompts keep it open. Positive target-logit gain can remain below the argmax threshold, and transport can re-emerge at the final prefix token or at the answer step in deeper layers. Across the ten VLMs we study, these local effects form descriptive transport patterns. Complementary experiments characterize how these patterns shift across datasets, attributes, and encoding amplitudes. Together, these results reframe the encoding-grounding gap as a problem of conditional transport in VLMs.
Souptik Kumar Majumdar, Fabian Kögel, Andreas Bullingcs.CV cs.AI
Linear probes and activation steering have uncovered that vision-language models (VLMs) internally represent mental states such as agents' beliefs, knowledge, and intentions. However, it is unclear whether and how these representations are used by downstream predictions along these axes. To close this gap, we introduce Cross-Axis Routing Diagnostic (CARD), which steers activations along one axis while measuring the response of a different axis's prediction. Applied to open-weight VLMs on Relay Chain -- a new cooperative grid-world benchmark we propose -- we diagnose a critical routing failure: models fail to incorporate belief representations into their next action prediction, effectively leaving valuable information about their partners unused.
Xiutian Zhao, Luqi Sun, Björn Schuller +1cs.CL eess.AS eess.IV
Modern multimodal foundation models (MFMs) have made rapid progress on tasks requiring integrated perception across speech, vision, and language, including emotion recognition. However, it remains unclear whether they recognize speech and facial emotion through shared affective functional units or modality-specific pathways. We explore emotion-sensitive neurons (ESNs), sparse decoder neurons selectively associated with emotion categories, in three MFMs: Gemma-4-12B-it, MiniCPM-o-4.5, and Qwen2.5-Omni-7B. Using speech emotion recognition and facial expression recognition as complementary probes, we identify acoustic and visual ESNs. Visual ESNs are causally meaningful: deactivating them selectively impairs recognition of the associated facial emotion, whereas steering their activations selectively enhances recognition of that emotion relative to other emotion categories. Acoustic and visual ESNs further show emotion-matched overlap and similar layer-wise distributions, indicating partial structural alignment between affective representations across speech and faces. Finally, cross-modal interventions reveal bidirectional causal transfer: ESNs identified from one modality produce emotion-specific effects when applied to the other. Our findings provide one of the first cross-modality activation-level analyses of affective functional units in MFMs, suggesting that speech and facial emotion recognition partially converge onto sparse decoder-level components that can be localized and manipulated without training.
Vision-language models (VLMs) achieve strong performance on high-level multimodal tasks, yet numerosity perception, a cognitive ability that emerges in human infants before language acquisition, remains poorly understood in current models, as existing counting benchmarks entangle numerosity with correlated visual factors. We introduce a cognitively inspired diagnostic benchmark, NumerosityVLM, comprising 10,800 synthetic images across six controlled conditions. The benchmark orthogonally manipulates object size, spatial arrangement, and numerosity, while progressively ablating texture, shape, and color. Evaluating seven VLMs in a zero-shot setting, multi-factor analysis reveals that model architecture explains the largest proportion of performance variance (partial $ω^{2}=0.325$), far exceeding visual conditions. Layer-wise probing further shows that linearly separable numerosity signals consistently emerge at early stages of the vision encoder, while performance differences across evaluated models are primarily associated with the language model component. Code and data are publicly available at https://github.com/fuy3/NumerosityVLM-Benchmark, and https://huggingface.co/datasets/fuy3/NumerosityVLM.
Fnu Pramono, John Cai, Sourabh Kulkarnics.CV cs.AI cs.CL cs.LG
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.
Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering and controlling multimodal behavior. MMDiff supports three uses: (i) feature isolation, by diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training; (ii) task-specific feature detection, via per-token contrastive firing analysis that isolates causal features; and (iii) feature-level control, by causally removing or steering the discovered feature directions. We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR. MMDiff discovers sparse, causally specific features whose removal selectively degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR, and reduces attack success rate by 24% on multimodal safety attacks, with no impact on VQA performance. Steering these features improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer steering baseline. These results show that multimodal SAEs can serve not only as interpretability tools, but as mechanisms for auditing, steering, and controlling MLLMs behavior toward safer and more capable generations.
Alexander Hackett, Arnaud Denis-Remillard, Axel Cassoucs.CV cs.AI cs.LG
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.
Xiutian Zhao, Philipp Koehn, Björn Schuller +1cs.CL eess.AS
Emotion is central to human communication, and its expression varies across languages. Large audio-language models (LALMs) achieve strong performance on multilingual speech tasks, yet it remains unclear whether they encode emotion through language-specific correlations or language-agnostic representations. We present the first neuron-level interpretability study of this question. We define Multilingual Emotion Neurons (MLENs) as functional units exhibiting stable emotional selectivity and aligned causal effects across languages, and introduce Consistency-Regularized Fusion (CR-Fusion) to identify them. Across four modern LALMs and 12 typologically diverse languages, emotion-sensitive neurons identified independently per language show minimal overlap, and additional monolingual identification data saturates quickly without isolating more transferable units, motivating identification from pooled cross-lingual evidence. Causal interventions demonstrate that MLENs identified by CR-Fusion provide more precise and transferable affective control than monolingual neuron sets in both zero-shot and low-resource settings. Leave-one-out ablations further reveal asymmetric transfer: individual identification languages, including low-resource ones, contribute non-redundant evidence, while several low-resource languages benefit most from the resulting cross-lingual transfer. Together, our findings provide the first causal, neuron-level account of how LALMs encode emotion across languages, and establish multilingual neuron identification as an effective mechanism for understanding cross-lingual affective behavior.
Shalom Kachko, Raz Lapid, Margarita Vald +2cs.AI cs.CL cs.CV
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.
Despite remarkable progress in vision-language generation, Vision-Language Models (VLMs) remain prone to hallucinations, producing content that is inconsistent with or unsupported by the input image. Existing works largely design detection or mitigation methods around one specific hallucination pattern, such as visual-textual imbalance, but real VLM hallucinations arise from a mixture of multiple patterns, so signals bound to a single pattern struggle to remain stable across models and tasks. Under a unified head-level view, we find that hallucination-induced changes manifest as localized deviations from each head's faithful contextual behavior, a phenomenon we term Role-Break. Detailed analysis reveals that these deviations are systematically organized across attention heads, contextual sources, and deviation directions, and that the resulting signal is linearly readable once head identity is preserved. Based on these findings, we build a lightweight linear detector on top of Role-Break that requires no fine-tuning of the VLM, whose feature dimension stays below 5,000 and reaches an average AUROC of 93.23 across six VLMs and four benchmarks. A small-scale intervention experiment further shows that the detected tokens can be directly acted upon in the discriminative setting.
Multimodal Large Language Models (MLLMs) achieve strong performance by integrating visual inputs with the rich priors of pretrained language models. However, they often fail on vision-centric tasks, especially when visual evidence conflicts with pretrained knowledge. We explore these failures separately using two diagnostic paradigms: (1) probing whether visual information is available, via image reconstruction, and (2) measuring multimodal context sensitivity, the extent to which the model follows visual context versus the language prior. To support the second, we introduce the WhatIfVis, a benchmark spanning five coarse-grained dimensions (spatial-temporal, color, count, size, and weight) whose questions admit answers from either the image or the prior. Our analysis yields three findings: (i) Coarse-grained visual evidence is preserved, as these attributes can be reconstructed from the final-layer image tokens of frozen MLLMs. Failures on questions about these attributes therefore point to post-perceptual utilization, rather than to degraded visual encoding during perception. (ii) Even when explicitly instructed to use or ignore visual evidence, vanilla models (without supervised fine-tuning on the WhatIfVis) show unstable visual context sensitivity. Supervised fine-tuning (SFT) improves this controllability and generalizes across domains, and activation patching further localizes the vision-versus-prior trade-off at architecture-specific depths across all six models. (iii) The vision-versus-prior trade-off is controllable along a learned vector. Applying this steering vector, even without any intent instruction, improves controllability over the vanilla model. Together, these results relocate the bottleneck, indicating that for the coarse attributes we study, MLLMs encode the visual evidence but cannot reliably control their reliance on it.
Brian Song, Michael A. Lepori, Ellie Pavlickcs.AI cs.CV
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.
Sana Tonekaboni, Viktoria Schuster, Caroline Uhlercs.LG
Real-world perception and decision making are inherently multimodal, integrating complementary signals across modalities. However, training multimodal models faces two main obstacles. First, collecting large-scale, well-aligned paired multimodal datasets is often impractical, making end-to-end multimodal training difficult. Second, existing multimodal representations frequently entangle information shared across modalities with modality-specific information, hindering interpretability and control. We introduce MultiLoReFT, an efficient and scalable low-rank representation fine-tuning framework for multimodal learning with pretrained unimodal models. MultiLoReFT extends low-rank adaptation to the multimodal setting and learns interpretable projection subspaces that decouple shared and modality-specific information. Across simulated and real-world benchmarks, it produces representations that support multimodal prediction while explicitly revealing how shared and modality-specific information is distributed across modalities.
We propose a language representation for multimodal data in which any observation, whether image, video, or text, is expressed as a bag of atomic propositions, simple statements about the entities, actions, and relations in a scene. A global semantic codebook unifies these into a shared vocabulary of canonical atomic propositions, placing every modality and observation into one interpretable space that spans fine grained facts to high level concepts and composes into richer ones. This brings interpretability with reasoning, cross-modal understanding and retrieval, and compositionality that enables complex multimodal understanding, rich data curation and complex structured retrieval. We demonstrate the framework on autonomous driving and open-world data.
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.
Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces. We ask whether CoT requires continued access to image tokens, or whether it mainly operates over visual information already made available earlier in the forward pass. We introduce Visual Access Sweep, a causal intervention that masks attention from generated-token queries to image-token keys along layer depth and generation time, and define the Visual Access Boundary (VAB) as the minimal access region that preserves task accuracy. Across six model configurations from Qwen2.5-VL and InternVL3, both no-CoT direct answering and CoT prompting exhibit finite VABs. In Qwen2.5-VL-32B and InternVL3 at 14B and 38B scales, when CoT is evaluated against the no-CoT full-access target, its VAB layer differs from the no-CoT boundary by at most two layers, despite substantially longer generations. This suggests that CoT does not primarily improve performance by prolonging direct image-token access throughout the reasoning trace, but by extending language-side computation over image-derived hidden-state information. We further show that CoT gains are constrained by perceptual readout. CoT helps when the queried visual attribute can be reliably read out by the model, but not when that readout is unreliable. A symbolic-attribute oracle shows that CoT can improve counting once ground-truth attributes are supplied as text, while a single-object probe-vs-decode check shows that hard attributes can be linearly recoverable from hidden states yet difficult for the model itself to output. Together, these analyses place the bottleneck at readout rather than counting.
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.
Ahmed Oumar El-Shangiti, Abzal Nurgazy, Hilal AlQuabeh +2cs.CV cs.LG
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.
Yoav Baron, Sara Dorfman, Roni Paiss +2cs.CV cs.AI
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.