Contrastive Learning for VLMs

Adversarial attacks on vision-language models optimize an image toward a text target, then cite the attacked model's similarity score as evidence of success. We ask whether that score - victim-space target alignment (VTS) - predicts recovery of the target by an independent model. We first build a measurement instrument: supervised judges outside the attacked geometry, real-target blend controls, shuffled-target negatives, and a reference level derived from a 50% target-image blend. Two preregistered studies then compare six contrastive encoders under a matched attack at three perturbation budgets. Robustly trained encoders (FARE, TeCoA, PMG, TRADES) transfer substantially more independent evidence than vanilla CLIP or SigLIP; all eight contrasts reject at the bootstrap floor. However, no cell reaches the blend-derived reference level. The three best cells fall within its replication band, leaving practical recovery undecided. Within robust encoders, per-sample alignment gain correlates with evidence gain ($ρ= 0.24-0.51$); within vanilla CLIP the correlation is consistent with zero. Across encoders we find no monotone alignment-evidence relation. VTS is therefore informative only within a fixed robust encoder, and we provide a reporting protocol in its place.

Medical Vision-Language Models (Med-VLMs) have demonstrated strong capabilities in clinical tasks. However, they often struggle to understand anatomical structures and spatial positioning, which are crucial for medical reasoning. To address this, we propose a localization-aware enhancement to the Med-VLM pipeline, introducing improvements at three levels: data,architecture, and alignment. First, we introduce localization lens, a set of expert-validated representations that provide richer anatomical and positional context. However, as these representations increase input complexity, we integrate pixel shuffle within the model architecture to filter and refine representations, enhancing spatial information processing while preserving anatomical continuity. Lastly, to effectively align the localization lens representations with textual features, we incorporate decoupled contrastive loss (DCL) alongside the standard loss function. This ensures better feature discrimination and robustness, particularly in data limited medical settings. Through extensive evaluations on medical visual question answering (Med-VQA) datasets, we show that our methodology improves localization-driven performance across different Med-VLM architectures. Our analysis of localization-based questions further reveals that improvements in anatomy and spatial reasoning directly enhance the overall accuracy of Med-VQA upto 6.2%. The proposed approach is model-agnostic and can be seamlessly integrated into existing Med-VLM pipelines. The dataset, code, and trained models will be made publicly available at https://github.com/CVLABLUMS/localizationlens.

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

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.

Ensuring effective transfer learning for vision-language models without compromising their generalization performance is crucial. However, many existing methods overlook data characteristics and simply reuse the training strategies adopted during pre-training. Specifically, they treat same-class samples as distinct instances and transform images independently of their paired text prompts, which makes model learning more difficult. We address these limitations through transformation-aware prompt conditioning and a re-calibrated contrastive loss. Fixed text descriptors identify the transformations applied to paired images, providing transformation-level consistency without altering class semantics. This design aligns the image and text branches at the transformation level, enabling richer representations while preserving the models' ability to generalize. In addition, our loss function mitigates positive-gradient dilution in soft-target cross-entropy when each anchor has multiple valid positives. During transfer, our approach treats same-class samples as positives rather than distinct instances, enabling the model to learn domain-specific features more effectively. Experiments across distribution shift, transfer learning, and few-shot settings demonstrate consistent improvements over existing approaches. Source code for our method is available at https://github.com/SoongE/ReCalCon.

Large Vision-Language Models (LVLMs) are prone to hallucinations: they fluently describe objects, attributes, and scenes that are not in the image. We connect part of this failure to a measurable property of their representations, feature instability, where mild semantics-preserving perturbations of the input cause large changes in the learned embeddings; hallucination rates rise together with this variability. Existing stability-motivated remedies are explicit, in the sense that they intervene at inference time through latent steering or constrained decoding, and pay for it on every query. We propose implicit stabilization instead: perturbation-invariance is built into the model weights during fine-tuning, and nothing extra runs at deployment. Our framework, INFUSE, first stabilizes visual and textual representations around perturbation-averaged and ground-truth anchors, then aligns the stabilized representations across modalities with bidirectional contrastive objectives. We prove that the anchor's root-mean-square deviation from the perturbation-mean representation shrinks at rate $1/\sqrt{K}$ in the number of views, and that under a Lipschitz decoder, this bounds how much any perturbation can change the model's hallucination behavior. On LLaVA-1.5, LLaVA-1.6, and Qwen3-VL-8B-Instruct, INFUSE reduces AMBER CHAIR by 46-63% relative to each base model, improves ObjHal, MMHal, HallusionBench, and POPE, and preserves VQA-v2 and TextVQA, all with no inference-time overhead.

Pretrained vision-language models such as CLIP excel at zero-shot recognition but often fail at compositionality, particularly attribute-object and relational structures. Recent studies mitigate this issue by augmenting training with synthetic hard negatives generated by a cascade of large language models and text-to-image models, which incurs substantial pipeline overhead. We instead propose MLLMCLIP, a heterogeneous distillation framework that transfers multimodal knowledge directly from a generative Multimodal Large Language Model (MLLM) teacher into a discriminative CLIP student, bypassing synthetic data entirely. To bridge the architectural mismatch between the two paradigms, we introduce an attention-based per-layer token selection and a CKA-based distillation loss. Compared to prior CLIP-enhancement methods, MLLMCLIP achieves state-of-the-art compositional accuracy while delivering consistent gains on standard zero-shot classification and image-text retrieval, showing that feature-level distillation strengthens both compositional and general vision-language representation capability.

Vision-language alignment powers open-vocabulary recognition, retrieval, and LVLM grounding, yet natural captions are often underspecified, making similarity brittle and overly confident under paraphrase and omitted details. We aim to learn representations whose matching is stable across caption views and whose confidence reflects how strongly text constrains an image. We propose Text as Partial Constraint (TPC), a core-residual alignment framework that treats multi-view captions as incomplete supervision. It distills a consensus semantic core as the alignment target, learns a single-view core predictor for standard inference with one query, and explicitly discourages vision-language similarity from depending on the orthogonal unsaid residual. An uncertainty-aware contrastive objective further softens alignment when caption views disagree, reducing overconfident updates under weak language constraints. Across zero-shot recognition and adversarial robustness, TPC achieves 81.42/64.05 Top-1 clean/robust accuracy on ImageNet and 76.19/52.03 on an Avg-14 transfer suite, while improving LVLM transfer with 85.16 POPE F1 and 59.57 OKVQA accuracy under an LLaVA-1.5-7B stack. These results suggest that modeling text as a partial constraint is a practical and principled route to more reliable vision-language representations under underspecified language supervision.

Recent research has shown that contrastive vision-language models such as CLIP often lack fine-grained understanding of visual content. While a growing body of work has sought to address this limitation, we identify a distinct failure mode in the CLIP family, which we term center bias, that persists even in recent model variants. Specifically, CLIP tends to disproportionately focus on the central region of an image, overlooking important objects located near the boundaries. This limitation is fundamental as failure to recognize relevant objects makes it difficult to perform any sophisticated tasks that depend on those objects. To understand the underlying causes of the limitation, we conduct analyses from both representation and attention perspectives. Using interpretability methods, i.e., embedding decomposition and attention map analysis, we find that relevant concepts especially those associated with off-center objects vanish from the model's embedding in the final representation due to information loss during the aggregation of visual embeddings, particularly the reliance on pooling mechanisms. Finally, we show that this bias can be alleviated with training-free strategies such as visual prompting and attention redistribution by redirecting models' attention to off-center regions.

Recent vision-language models (VLMs) typically rely on a single vision encoder trained with contrastive image-text objectives, such as CLIP-style pretraining. While contrastive encoders are effective for cross-modal alignment and retrieval, self-supervised visual encoders often capture richer dense semantics and exhibit stronger robustness on recognition and understanding tasks. In this work, we investigate how to scale the fusion of these complementary visual representations for vision-language modeling. We propose CoME-VL: Complementary Multi-Encoder Vision-Language, a modular fusion framework that integrates a contrastively trained vision encoder with a self-supervised DINO encoder. Our approach performs representation-level fusion by (i) entropy-guided multi-layer aggregation with orthogonality-constrained projections to reduce redundancy, and (ii) RoPE-enhanced cross-attention to align heterogeneous token grids and produce compact fused visual tokens. The fused tokens can be injected into a decoder-only LLM with minimal changes to standard VLM pipelines. Extensive experiments across diverse vision-language benchmarks demonstrate that CoME-VL consistently outperforms single-encoder baselines. In particular, we observe an average improvement of 4.9% on visual understanding tasks and 5.4% on grounding tasks. Our method achieves state-of-the-art performance on RefCOCO for detection while improving over the baseline by a large margin. Finally, we conduct ablation studies on layer merging, non-redundant feature mixing, and fusion capacity to evaluate how complementary contrastive and self-supervised signals affect VLM performance.

Contrastive vision-language (V&L) models remain a popular choice for various applications. However, several limitations have emerged, most notably the limited ability of V&L models to learn compositional representations. Prior methods often addressed this limitation by generating custom training data to obtain hard negative samples. Hard negatives have been shown to improve performance on compositionality tasks, but are often specific to a single benchmark, do not generalize, and can cause substantial degradation of basic V&L capabilities such as zero-shot or retrieval performance, rendering them impractical. In this work we follow a different approach. We identify two root causes that limit compositionality performance of V&Ls: 1) Long training captions do not require a compositional representation; and 2) The final global pooling in the text and image encoders lead to a complete loss of the necessary information to learn binding in the first place. As a remedy, we propose two simple solutions: 1) We obtain short concept centric caption parts using standard NLP software and align those with the image; and 2) We introduce a parameter-free cross-modal attention-pooling to obtain concept centric visual embeddings from the image encoder. With these two changes and simple auxiliary contrastive losses, we obtain SOTA performance on standard compositionality benchmarks, while maintaining or improving strong zero-shot and retrieval capabilities. This is achieved without increasing inference cost. We release the code for this work at https://github.com/saic-fi/concept_centric_clip.

Due to the impressive zero-shot capabilities, pre-trained vision-language models (e.g., CLIP), have attracted widespread attention and adoption across various domains. Nonetheless, CLIP has been observed to be susceptible to adversarial examples. Through experimental analysis, we have observed a phenomenon wherein adversarial perturbations induce shifts in text-guided attention. Building upon this observation, we propose a simple yet effective strategy: Text-Guided Attention for Zero-Shot Robustness (TGA-ZSR). This framework incorporates two components: Local Attention Refinement Module and Global Attention Constraint Module. Our goal is to maintain the generalization of the CLIP model and enhance its adversarial robustness. Additionally, the Global Attention Constraint Module acquires text-guided attention from both the target and original models using clean examples. Its objective is to maintain model performance on clean samples while enhancing overall robustness. However, we observe that the method occasionally focuses on irrelevant or spurious features, which can lead to suboptimal performance and undermine its robustness in certain scenarios. To overcome this limitation, we further propose a novel approach called Complementary Text-Guided Attention (Comp-TGA). This method integrates two types of foreground attention: attention guided by the class prompt and reversed attention driven by the non-class prompt. These complementary attention mechanisms allow the model to capture a more comprehensive and accurate representation of the foreground. The experiments validate that TGA-ZSR and Comp-TGA yield 9.58% and 11.95% improvements respectively, in zero-shot robust accuracy over the current state-of-the-art techniques across 16 datasets.

Reliable crop disease detection requires models that perform consistently across diverse acquisition conditions, yet existing evaluations often focus on single architectural families or lab-generated datasets. This work presents a systematic empirical comparison of three model paradigms for fine-grained crop disease classification: Convolutional Neural Networks (CNNs), contrastive Vision-Language Models (VLMs), and generative VLMs. To enable controlled analysis of domain effects, we introduce AgriPath-LF16, a benchmark of 111k images spanning 16 crops and 41 diseases with explicit separation between laboratory and field imagery, alongside a balanced 30k subset for standardised training and evaluation. We train and evaluate all models under unified protocols across full, lab-only, and field-only training regimes using macro-F1 and Parse Success Rate (PSR) to account for generative reliability (i.e., output parsability measured via PSR). The results reveal distinct performance profiles: CNNs achieve the highest accuracy on in-domain imagery but exhibit pronounced degradation under domain shift; contrastive VLMs provide a robust and parameter-efficient alternative with competitive cross-domain performance; generative VLMs demonstrate the strongest resilience to distributional variation, albeit with additional failure modes stemming from free-text generation. These findings highlight that architectural choice should be guided by deployment context rather than aggregate performance alone.

When a text description is extended with an additional detail, image-text similarity should drop if that detail is wrong. We show that CLIP-style dual encoders often violate this intuition: appending a plausible but incorrect object or relation to an otherwise correct description can increase the similarity score. We call such cases half-truths. On COCO, CLIP prefers the correct shorter description only 40.6% of the time, and performance drops to 32.9% when the added detail is a relation. We trace this vulnerability to weak supervision on caption parts: contrastive training aligns full sentences but does not explicitly enforce that individual entities and relations are grounded. We propose CS-CLIP (Component-Supervised CLIP), which decomposes captions into entity and relation units, constructs a minimally edited foil for each unit, and fine-tunes the model to score the correct unit above its foil while preserving standard dual-encoder inference. CS-CLIP raises half-truth accuracy to 69.3% and improves average performance on established compositional benchmarks by 5.7 points, suggesting that reducing half-truth errors aligns with broader gains in compositional understanding. Code is publicly available at: https://github.com/kargibora/CS-CLIP

Contrastive Language-Image Pre-training (CLIP) relies on Vision Transformers whose attention mechanism is susceptible to spurious correlations, and scales quadratically with resolution. To address these limitations, We present CLIMP, the first fully Mamba-based contrastive vision-language model that replaces both the vision and text encoders with Mamba. The new architecture encodes sequential structure in both vision and language, with VMamba capturing visual spatial inductive biases, reducing reliance on spurious correlations and producing an embedding space favorable for cross-modal retrieval and out-of-distribution robustness-surpassing OpenAI's CLIP-ViT-B by 7.5% on ImageNet-O. CLIMP naturally supports variable input resolutions without positional encoding interpolation or specialized training, achieving up to 6.6% higher retrieval accuracy at 16x training resolution while using 5x less memory and 1.8x fewer FLOPs. The autoregressive text encoder further overcomes CLIP's fixed context limitation, enabling dense captioning retrieval. Our findings suggest that Mamba exhibits advantageous properties for vision-language learning, making it a compelling alternative to Transformer-based CLIP.The code and models are publicly available at https://github.com/NimrodShabtay/CLIMP}

Contrastive vision-language models continue to be the dominant approach for image-text retrieval. Contrastive Language-Image Pre-training (CLIP) trains two neural networks to align their image and text embeddings in a shared latent space. As a challenging case-study for neurosymbolic AI, recent results evaluating CLIP on negated or paraphrased text have shown mixed performance as these are difficult to define formally for text data. Negation produces the opposite meaning using various possible but small lexical changes. Paraphrasing may use very different textual expressions to denote essentially the same thing. As a result, learning of paraphrasing and negation together poses a significant challenge because of the above mismatch between changes in syntax and intended meaning expected to be captured by distances in embedding space. This paper proposes a new CLIP contrastive loss function capable of balancing the requirements of having both paraphrasing and negation. It applies training triplets consisting of original, paraphrased and negated text generated by multiple large language models to the evaluation of CLIP models. The approach, called SemCLIP, aims to learn semantically-relevant and simple embeddings, placing paraphrased captions nearer to the original image embeddings while at the same time pushing negated captions farther away. Empirically, SemCLIP is shown to be capable of preserving roughly the same performance as CLIP augmented with either negation or paraphrasing. Although direct comparisons are difficult to make because the problem of learning with both negation and paraphrasing is different, an expected benefit of SemCLIP should be robustness when applied zero-shot to downstream image classification tasks. Our experiments confirm such robustness as measured by difference in accuracy (mean-accuracy delta) between original and negated captions on five downstream datasets.

Monitoring cattle behaviour provides valuable insights into animal health, productivity and welfare. However, robust behaviour recognition in real-world farm environments remains challenging because well-annotated livestock video datasets are scarce and agricultural surveillance footage differs substantially from large-scale pretraining data. To address these challenges, we propose Cattle-CLIP, a domain-adaptive vision-language framework based on Contrastive Language-Image Pretraining (CLIP), which reformulates cattle behaviour recognition as a cross-modal semantic alignment task. Cattle-CLIP incorporates a temporal integration module that extends image-level contrastive pretraining to video-based behaviour understanding and promotes consistent semantic alignment across time. To reduce domain discrepancy, customised data augmentation and behaviour-oriented prompts are employed. We also construct CattleBehaviours6, a curated dataset containing 1,905 annotated clips from six indoor behaviours: feeding, drinking, standing-self-grooming, standing-ruminating, lying-self-grooming and lying-ruminating. The dataset provides standardised ethogram definitions for model training, evaluation and future livestock behaviour research. Cattle-CLIP was evaluated under supervised and few-shot learning settings, with particular emphasis on data-constrained behaviour recognition. In supervised experiments, the model achieved 96.1% accuracy across six behaviours and showed high precision and recall for feeding, drinking and standing-ruminating. It also maintained promising few-shot performance with limited labelled data, demonstrating the potential of multimodal learning for cattle behaviour recognition in data-constrained farm environments.

Adapting vision-language models in low-resource languages remains challenging, as they often produce fluent text about the wrong objects. This stems from scarce paired data, translation pivots that break alignment, and English-centric pretraining that ignores target-language semantics. We address this with a compute-aware Bengali captioning pipeline trained on LaBSE-verified EN-BN pairs and 110k bilingual-prompted synthetic images. A frozen MaxViT yields stable visual patches, a Bengali-native mBART-50 decodes, and a lightweight bridge links the modalities. Our core novelty is a tri-loss objective: Patch-Alignment Loss (PAL) aligns real and synthetic patch descriptors using decoder cross-attention, InfoNCE enforces global real-synthetic separation, and Sinkhorn-based OT ensures balanced fine-grained patch correspondence. This PAL+InfoNCE+OT synergy tightens real-synthetic patch alignment, reduces spurious matches, and drives strong gains on Flickr30k-1k (BLEU-4 12.29, METEOR 27.98, BERTScore-F1 71.20) and MSCOCO-1k (BLEU-4 12.00, METEOR 28.14, BERTScore-F1 75.40), outperforming strong CE baselines and narrowing the real-synthetic centroid gap by 41%.