Modality Gap in VLMs

Class-Incremental Learning (CIL) requires models to recognize new classes over time without forgetting previously learned ones. With the rise of vision-language pre-training, CLIP has become a strong foundation for CIL. A common design in CLIP-based CIL is to construct textual classifier weights by encoding class-name templates with the CLIP text encoder, and then classify visual features by image-text cosine similarity. This design is appealing: since CLIP aligns images and text in a shared embedding space, textual weights appear to provide an off-the-shelf classifier for incremental classes. However, we show that this seemingly natural design is not always beneficial, as a modality gap can still separate the two modalities and make textual classifier weights deviate from visual class distributions. Empirically, under identical task-wise CIL training, initializing the cosine classifier with visual class centers yields lower loss and better incremental accuracy than using CLIP textual features. Motivated by these observations, we propose VIS, a visual-only method for CLIP-based CIL that removes the deployed textual branch and constructs the incremental classifier entirely in the visual space. To obtain stronger task-adaptive visual representations, VIS uses only base-session data to enhance CLIP's final visual representation with informative visual-layer features. Built on the enhanced visual representation, VIS employs a simple kernelized incremental least-squares SVM, whose classifier weights are solved in closed form from additive sufficient statistics. When new classes arrive, VIS accumulates their sufficient statistics and recomputes the classifier weights for all seen classes, enabling efficient incremental updates while preserving historical class knowledge. Extensive experiments show that VIS achieves state-of-the-art performance without a textual branch.

Visual distinctions are often finer than those reflected in linguistic conceptualization. Vision-language models exhibit a similar asymmetry: a distinction can remain discriminable in frozen image geometry while being weakly addressable through the native text interface. We study this gap by separating visual discriminability from linguistic addressability in text-to-image retrieval. Using FactorAtlas, a fully crossed testbed of 23,040 images spanning shape, hue, pattern, and nuisance variation, we compare both readouts on held-out images of the same distinctions. We then derive image-side contrasts that separate each value from its alternatives for matched visual grounding, and test whether this reduces the native-text access gap across factors and models. Direction-specific and visual-absence controls tie these gains to the relevant visual contrast; the gains persist after global alignment and extend to compositional retrieval and natural images. Together, these results show that visual discriminability and linguistic addressability need not coincide, and that matched visual grounding can probe and reduce the resulting access gap.

Benefiting from transferable visual-textual alignment, CLIP has been widely adopted for class-incremental learning (CIL). However, existing learners either repeatedly update components shared across tasks, leading to knowledge overwriting, or overly isolate new-task updates, hindering the reuse of CLIP's transferable knowledge and limiting plasticity. Moreover, the text-based or bimodal classifier designs still fail to effectively integrate complementary information from the visual and textual modalities. To address these challenges, we introduce DuLBE, which couples dual-mode low-rank learning with a bridge-prototype ensemble classifier for exemplar-free CIL. DuLBE allocates two visual low-rank update modes according to the gradient demand and uses gradient routing to coordinate them: a compact and rewritable shared mode is selected from historically occupied visual directions to reuse transferable knowledge, while residual modes provide low-interference channels for task-specific variations. Building on the resulting stable inter-modal structure, we further construct geodesic bridges between visual prototypes and text embeddings on the unit hypersphere, and ensemble reliable bridge prototypes to compensate for the modality-gap limitations of textual decision boundaries. Extensive experiments under multiple settings show that DuLBE achieves state-of-the-art CIL performance while retaining the high parameter efficiency of low-rank tuning.

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

Large language models are vulnerable to prompt injection attacks, where third-party adversarial content can hijack the model's behavior. In this paper, we study the role played by the adversarial data's input modality, and identify a systematic asymmetry: multimodal LLMs are more likely to follow adversarial instruction when they appear as text than when the same instruction is delivered through a non-textual channel (e.g., as an image). We hypothesize that this modality gap arises from text-centric instruction tuning, which teaches models to obey textual instructions while treating other modalities mainly as content to parse or describe. We then demonstrate how this gap can be turned into a training-free defense, by rendering all untrusted payloads as typographic images (or audio) before they reach the model. Across ten models and two prompt injection benchmarks (DirectInject and AgentDojo) we show that our defense Pictionary consistently reduces attack success rates even against the strongest adaptive attacks and human red teamers, while largely preserving benign utility. We further show that benign fine-tuning on image-rendered instructions erodes the modality gap, tracing it to the text-centric instruction-tuning distribution.

Contrastive vision-language models learn shared embedding spaces by aligning matched image-text pairs, yet their representations remain separated by a modality gap. Prior work reports divergent effects of modifying this gap: reducing it can improve zero-shot classification and cross-modal alignment, whereas removing gap-related structure can degrade image-text retrieval. In this paper, we provide a unified geometric explanation for these task-dependent effects. Across CLIP and SigLIP encoders, we find that a single dominant direction captures 94.4-99.9% of the squared norm of the image-text mean separation, revealing that the mean-separation component is approximately rank-one. A decomposition of the similarity score then identifies three task-specific roles. In zero-shot classification, query-side fixed gap-offset subtraction is exactly equivalent to an additive class bias. In standard cross-modal retrieval, projecting out the gap direction and renormalising residuals discards candidate-specific norm information, inducing a multiplicative ranking distortion; a geometry-derived exponent tracks the grid-search optimum (Spearman rho = 0.93) and restores performance in some settings, although the gains transfer unevenly. In mixed-modal retrieval, the gap direction sorts candidates by modality; its removal can improve cross-modal ranking, unlike random or non-gap controls. Residual semantic structure after removal defines the limits of the rank-one account. Together, these results explain why gap modification can improve, degrade, or restore performance across downstream settings. By clarifying when and why gap modification changes model behavior, this account provides a principled basis for selecting gap interventions in similarity-based vision-language systems across evaluated downstream tasks.

Large-scale vision-language models (VLM) such as CLIP enable strong open-vocabulary reasoning, yet deploying these capabilities on resource-constrained edge devices remains challenging. EdgeVL addresses this problem by distilling CLIP representations into lightweight multi-modal encoders and applying quantization-aware training (QAT) for efficient Open-Vocabulary Classification (OVC) on edge hardware. However, its two-stage optimization applies different objectives for distillation and QAT, and contrastive learning is performed within the quantized student space, which can result in inconsistent optimization and reduced training efficiency. Moreover, identical supervision across RGB and non-RGB modalities may lead to modality imbalance. We propose a unified framework for quantized semantic distillation tailored to edge deployment. By jointly optimizing distillation and quantization within a unified teacher-anchored framework, our method ensures consistent training under quantization, suppressing hard negatives and enlarging decision margins. Additionally, we design a lightweight cross-attention adapter that enhances non-RGB representations through RGB-guided semantic transfer, narrowing the modality gap. Extensive experiments demonstrate consistent improvements on non-RGB modalities while maintaining deployment efficiency.

In-context learning (ICL) enables models to infer tasks from demonstrations, but existing benchmarks generally lack matched text and image versions needed to compare ICL performance across modalities. We introduce TwinICL, a procedurally generated benchmark providing such pairs for controlled comparison. Across six open-weight models and 38 tasks, multimodal ICL consistently underperforms text-only ICL, with gaps varying by task family. To test whether this gap can be recovered, we target visual access, task framing, and reasoning through three interventions. Their combination recovers strong multimodal ICL performance on a diagnostic subset, despite limited or inconsistent individual effects. To distinguish difficulties in executing tasks from those in inferring them, we evaluate models with explicit task instructions, revealing a modality gap even when the task is known. We then examine how adding demonstration inputs and outputs reshapes this gap, highlighting demonstrations' dual role as additional context to process and evidence about the task. The dataset is available at https://github.com/lab-flair/TwinICL.

Vision-language models such as CLIP embed images and text in a shared space, where modality-specific distributions often remain separated. Existing accounts connect this modality gap to initialization, contrastive dynamics, and information imbalance, while its distributional and pairwise contributions to retrieval remain unresolved. We introduce UOT-Gap, a training-free variational diagnostic that models frozen image and text embeddings with unbalanced entropic optimal transport (UOT). The UOT optimum separates transport, coupling complexity, and marginal mass variation; a complementary pair-aware residual compares observed image-caption pairs with the UOT soft matching. On Flickr8K and COCO-1K with frozen CLIP, OpenCLIP, and SigLIP encoders, caption degradation reduces Flickr8K Recall@1 from 0.559 to 0.003. Across six dataset-model conditions, the pair-aware residual tracks retrieval degradation with mean absolute Spearman 0.973, compared with 0.392 for the mean gap. The association remains stable across five random COCO-1K subsets at $0.954\pm0.026$, with a minimum of 0.943. UOT barycentric updates reduce the transport objective while degrading retrieval, distinguishing geometric objective descent from task improvement. These results establish UOT-Gap as a diagnostic for caption quality, modality alignment, and retrieval robustness.

Vision-language models are evaluated by aggregate accuracy on multimodal benchmarks, a practice that implicitly assumes the model uses its visual input. We show this assumption fails on 40%--97% of samples across six VLMs and three perceptual benchmarks: blurring the question-relevant visual region leaves the next-token distribution nearly unchanged. We name this phenomenon the Visual Insensitivity Gap and quantify it with a per-sample Visual Sensitivity Index (VSI). The gap is a property of samples, not of models: VSI ranks correlate across models (grand-mean Spearman rho=+0.40, permutation p<10^-3), so the same samples are flagged insensitive by VLMs sharing no architectural detail beyond a contrastively pretrained vision tower. The mechanism is concrete: on the insensitive samples, a linear probe on each model's own vision tower distinguishes perturbed from clean images at 0.72--0.79 accuracy, yet the model's argmax token changes on only 2%--11% of the same samples, an encoder--LLM gap above 0.65 on every model. Mapping VSI's diagnostic utility cell by cell surfaces a strong regime (multi-choice reasoning on capable VLMs: AUROC=0.85--0.87) and a weak regime (well-calibrated factuality, where softmax confidence already leads). VSI is not a universal best abstention signal; it is a sample-intrinsic indicator of vision-ignoring failure, best used as a conditional ensemble component.

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

Human language is highly polysemous. Many common words (e.g., 'bank' or 'palm') carry several distinct meanings that shape what humans communicate and imagine. Large language models (LLMs) have been shown to understand this multiplicity of meaning, but much less is known about how polysemy surfaces in other modalities such as images. We study this across 17 text-to-image and 15 text-generation models by giving each a polysemous word with no context to fix its meaning and measuring which senses are produced over many samples. We find a clear multimodal gap, where within every model family, generated images settle on far fewer senses than generated sentences (normalized entropy 0.10 vs. 0.25), and both are far less varied than what people imagine for the same words (normalized entropy 0.47). However, when we instead ask a model to list how often it would generate outputs corresponding to each possible meaning of a word, it predicts distributions that are more diverse than the actual space of outputs. These results reveal a multimodal gap in how foundation models express meaning, and how their understanding may not transfer faithfully nor equally across modalities.

Multimodal Large Language Models (MLLMs) inherit rich relational priors from their language backbones, yet often fail when asked to apply these relationships in visual contexts. We trace this failure to a structural blind spot: projection-based alignment trains each visual token to carry the right semantics, but never asks whether the relationships between concepts survive the crossing from language to vision. To address this, we propose MIRROR (Mapping Inter-concept Relations from language to visual Representation via Optimal-transport-based Regularization), a geometric regularization framework that transfers relational priors from language to vision by exploiting the rich relational structure encoded in language representations. Specifically, we derive a surrogate loss from the proposed Semi-Inverse Gromov-Wasserstein (SI-GW) problem, an inverse geometric problem that aligns visual representations with language-derived relational priors. We show that this formulation admits a unique closed-form solution that prescribes the ideal visual relational structure implied by language geometry and cross-modal coupling. The structure of the formulation also enables efficient computation, making it applicable to long token sequences. Applying SI-GW inside decoder-only Transformers requires careful design. We introduce targeted strategies at the layer, head, and token levels to ensure stable extraction without additional parameters or inference cost. MIRROR improves relational consistency while preserving performance on general vision-language tasks.

When vision contradicts text, multimodal large language models (MLLMs) consistently favor text, even when images provide clear evidence otherwise. This bias poses risks for applications requiring visual grounding, yet its cause remains unclear. In this paper, we uncover a surprising finding: models often get it right initially, forming correct vision-based predictions in their intermediate layers, before changing their minds and favoring text in the final output. We call this "late-layer textual override". The visual information is encoded, it simply does not survive to the output. More intriguingly, we find that how predictions change reveals whether they're correct: 85% of failures shift toward text, while 89% of successes shift toward vision. This directional signature enables a simple but powerful intervention: when we detect a confident visual prediction being suppressed, we restore it. We propose CALRD (Conflict-Aware Layer Reference Decoding), a training-free method that recovers overridden predictions at inference time. Experiments across five MLLMs of varying architectures demonstrate up to 9.4% absolute improvements on conflict benchmarks while largely preserving standard performance, without training or external knowledge. It recovers what the model already knew but failed to preserve.

Vision-Language Models (VLMs) struggle when applied to medical image-text data, yet the tools available to diagnose this failure remain limited. Existing representation alignment metrics are symmetric, collapsing both modalities into a single score and hiding which modality drives cross-modal degradation. We introduce the Spectral Alignment Score (SAS), an asymmetric metric that projects both modalities onto the principal eigenbasis of an anchor modality and computes eigenvalue-weighted per-eigenmode correlations, resulting in directional scores whose difference quantifies modality information imbalance. We embed SAS within a benchmarking framework evaluating 15 VLMs across natural and medical image-text datasets alongside 6 alignment metrics and bidirectional retrieval. Our experiments show that medical images retain richer structural information than their paired clinical reports, a directional asymmetry invisible to all competing metrics, and that SAS achieves the strongest zero-label correlation with retrieval performance in the medical domain, positioning it as a practical diagnostic tool for clinical deployment. Code is available at this URL: https://github.com/iamalegambetti/medical-vlms-assessment.

Vision-Language models (VLMs), such as CLIP, achieve powerful zero-shot classification. However, their predictions remain sensitive to spurious correlations, where contextual cues dominate over semantic content. Earlier solutions typically rely on fine-tuning or prompt engineering, which either undermine the advantages of pre-trained models or are prone to hallucination. In this work, we propose Density-Aware Translation (DAT) that refines image-text similarity scores using a local geometric density term derived from group reference sets. Our approach is motivated by the phenomenon that CLIP embeddings exhibit a modality gap and lie on an anisotropic shell in the feature space: common patterns cluster near the mean, while rare patterns are pushed outward. This geometry creates uneven alignment, where spurious correlations are amplified while semantically meaningful but rare cues are marginalised. To address this, we employ a relative measure to rescale similarities based on embedding density, suppressing overconfident scores in diffuse regions while preserving dense, semantically consistent matches. Experimental results on benchmark datasets demonstrate consistent improvements in worst-group and average accuracy, highlighting density-aware translation as a simple and effective calibration mechanism for reliable zero-shot classification using multimodal models.

Vision-Language Models (VLMs) have achieved substantial progress across a wide range of understanding and reasoning tasks, driven by large-scale image-text training aimed at multimodal fusion. Ideally, replacing a textual question with its rendered-image counterpart should leave model performance essentially unaffected. In practice, however, such modality substitution induces dramatic performance degradation. We attribute this "carrier sensitivity" issue to an inherent bias in current training corpora. Across prevalent datasets such as image captioning, VQA, OCR, and web-sourced interleaved data, text and images are typically organized into distinct and asymmetric roles, with text serving as linguistic queries and images as visual references. Such data bias leads VLMs to exhibit distinct preferences for information acquisition across different modalities. Consequently, VLMs fail to align representations of semantically equivalent content across textual and visual carriers, making model reasoning fragile under modality substitution. To address this, we propose Local Modality Substitution (LoMo), a lightweight, architecture-agnostic data curation paradigm designed to provide supervision for cross-modal representational invariance between semantically equivalent text and image carriers. LoMo achieves this by reformulating single-modality prompts into seamlessly interleaved multimodal sequences. It dynamically selects target text spans and recasts them as rendered images, thereby preserving the same semantics across "text, visual, text" carriers. Extensive experiments across 13 diverse multimodal benchmarks demonstrate that LoMo significantly improves overall multimodal reasoning and yields deeper cross-modal fusion. Specifically, it delivers consistent gains across foundational models, improving over standard SFT by 2.67 points on LLaVA-OneVision-1.5-8B and 2.82 points on Qwen3.5-9B.

Out-of-distribution (OOD) detection has emerged as a popular technique to enhance the reliability of machine learning models by identifying unexpected inputs from unknown classes. Recent progress in pre-trained vision-language models (VLMs) has enabled zero-shot OOD detection without access to in-distribution (ID) training data; in this setting, existing methods commonly treat text embeddings of class names as class prototypes. In this paper, we challenge the widely adopted text-as-prototype paradigm by theoretically showing that off-the-shelf textual prototypes are generally misaligned with the optimal visual prototypes, yielding an intrinsic modality gap that cannot be eliminated by prompt engineering alone. To mitigate this gap under the post-hoc constraint, this paper presents an online pseudo-supervised framework that directly learns class prototypes in the visual feature space using unlabeled test-time data streams and soft predictions from the pre-trained VLMs. We provide theoretical guarantees for the convergence of the online optimization procedure. Extensive experiments empirically demonstrate that our method achieves a new state of the art across a variety of OOD detection setups.

Vision-language models (VLMs) are increasingly proposed for clinical decision support, yet their reliability in real-world scenarios that require integrating both visual and textual context from medical records remains poorly characterized. This paper identifies three failure modes: (1) modality over-reliance on text over images, (2) spurious reliance on irrelevant clinical history, and (3) prompt sensitivity across semantically equivalent inputs. We evaluate a diverse set of general-domain and medically-tuned open and closed VLMs on chest x-ray tasks using MIMIC-CXR. By systematically manipulating image-text alignment, clinical history, and prompt formulations, we found that VLM decisions are dominated by the text modality, even when visual evidence is available. Moreover, we observed that VLMs are heavily influenced by irrelevant reports, while minor prompt changes can reverse correct image-based predictions. Our findings underscore the need for explicit safeguards and stress-testing before considering the use of these models in clinical practice.

Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input. We investigate the root causes of these failure modes with a mechanistic analysis focusing on the decoder-based VLMs. We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence. While prior work either aggressively closes this gap or suppresses hallucinations through expensive black-box decoding strategies, none addresses the underlying geometric cause. We provide the first quantitative characterization of this over-alignment, demonstrating that linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace. Building on this insight, we propose two complementary remedies: a training-free inference strategy and a bias-aware fine-tuning paradigm, both of which explicitly project out this subspace from visual representations. Our methods significantly reduce hallucinations across POPE, CHAIR, and AMBER benchmarks, and improve CLAIR scores on long-form captioning tasks, with the training-free variant adding no computational overhead over the base model.

The visual modality of vision-language models (VLMs) is an underexplored attack surface for bypassing safety alignment. We introduce four jailbreak attacks exploiting the vision component: (1) encoding harmful instructions as visual symbol sequences with a decoding legend, (2) replacing harmful objects with benign substitutes (e.g., bomb -> banana) then prompting for harmful actions using the substitute term, (3) replacing harmful text in images (e.g., on book covers) with benign words while visual context preserves the original meaning, and (4) visual analogy puzzles whose solution requires inferring a prohibited concept. Evaluating across six frontier VLMs, our visual attacks bypass safety alignment and expose a cross-modality alignment gap: text-based safety training does not automatically generalize to harmful intent conveyed visually. For example, our visual cipher achieves 40.9% attack success on Claude-Haiku-4.5 versus 10.7% for an equivalent textual cipher. To further our insight into the attack mechanism, we present preliminary interpretability and mitigation results. These findings highlight that robust VLM alignment requires treating vision as a first-class target for safety post-training.

The rapid proliferation of Vision-Language Models (VLMs) is often framed as enabling unified multimodal knowledge discovery but rests on an under-examined assumption: that current VLMs faithfully synthesise multimodal data. We argue they often do not, and this gap reflects a trustworthiness problem in the dominant Vision Encoder-Projector-LLM paradigm. Rather than extracting grounded knowledge from visual inputs, state-of-the-art models frequently exhibit functional blindness, i.e., exploiting strong language priors to bypass severe visual representation bottlenecks. In this work, we challenge the conventional methodology of multimodal evaluation, which relies on data ablation or new dataset creation and therefore conflates dataset biases with architectural incapacity. We propose an information-theoretic departure: the Modality Translation Protocol, designed to quantify what we call the Expense of Seeing. By translating semantic payloads rather than ablating them, we formulate three novel metrics -- the Toll (ToS), Curse (CoS), and Fallacy (FoS) of Seeing -- culminating in the Semantic Sufficiency Criterion (SSC). Furthermore, we hypothesise a Divergence Law of Multimodal Scaling: as the underlying language engines scale to unprecedented reasoning capabilities, the penalty of the visual knowledge bottleneck may increase rather than diminish. We argue the community should move beyond "multimodal gain" as a primary evaluation target. By elevating the SSC from a passive diagnostic constraint to an active architectural blueprint, we provide a foundation for guiding the next generation of AI systems toward genuine multimodal reasoning.

Reasoning in vision-language models (VLMs) has recently attracted significant attention due to its broad applicability across diverse downstream tasks. However, it remains unclear whether the superior performance of VLMs stems from genuine vision-grounded reasoning or relies predominantly on the reasoning capabilities of their textual backbones. To systematically measure this, we introduce CrossMath, a novel multimodal reasoning benchmark designed for controlled cross-modal comparisons. Specifically, we construct each problem in text-only, image-only, and image+text formats guaranteeing identical task-relevant information, verified by human annotators. This rigorous alignment effectively isolates modality-specific reasoning differences while eliminating confounding factors such as information mismatch. Extensive evaluation of state-of-the-art VLMs reveals a consistent phenomenon: a substantial performance gap between textual and visual reasoning. Notably, VLMs excel with text-only inputs, whereas incorporating visual data (image+text) frequently degrades performance compared to the text-only baseline. These findings indicate that current VLMs conduct reasoning primarily in the textual space, with limited genuine reliance on visual evidence. To mitigate this limitation, we curate a CrossMath training set for VLM fine-tuning. Empirical evaluations demonstrate that fine-tuning on this training set significantly boosts reasoning performance across all individual and joint modalities, while yielding robust gains on two general visual reasoning tasks. Source code is available at https://github.com/xuyige/CrossMath.

While humans develop core visual skills long before acquiring language, contemporary Multimodal LLMs (MLLMs) still rely heavily on linguistic priors to compensate for their fragile visual understanding. We uncovered a crucial fact: state-of-the-art MLLMs consistently fail on basic visual tasks that humans, even 3-year-olds, can solve effortlessly. To systematically investigate this gap, we introduce BabyVision, a benchmark designed to assess core visual abilities independent of linguistic knowledge for MLLMs. BabyVision spans a wide range of tasks, with 388 items divided into 22 subclasses across four key categories. Empirical results and human evaluation reveal that leading MLLMs perform significantly below human baselines. Gemini3-Pro-Preview scores 49.7, lagging behind 6-year-old humans and falling well behind the average adult score of 94.1. These results show despite excelling in knowledge-heavy evaluations, current MLLMs still lack fundamental visual primitives. Progress in BabyVision represents a step toward human-level visual perception and reasoning capabilities. We also explore solving visual reasoning with generation models by proposing BabyVision-Gen and automatic evaluation toolkit. Our code and benchmark data are released at https://github.com/UniPat-AI/BabyVision for reproduction.