Vision language model

Vision-language models (VLMs) are artificial intelligence models that learn relationships between visual data and natural language. The term covers models that compare images with text, models that combine both modalities to make a prediction, and generative models that produce text from images and prompts. A VLM may accept a photograph, diagram, document page, or sequence of video frames; its language input or output can be a label, query, caption, dialogue turn, or other text.

The category is broader than an image-enabled large language model. Dual-encoder VLMs such as CLIP encode images and text separately and compare their vectors, while generative VLMs connect a visual encoder to an autoregressive language model. "Large vision-language model," "large multimodal model," and "multimodal large language model" are used inconsistently across papers. In this article, VLM means a model whose training objective or architecture explicitly connects vision and language; multimodal model is the broader category that can also include audio, speech, sensor data, or other modalities.

Scope and terminology

The boundary of the term depends on the research community and task. Earlier literature often described "vision-and-language" systems built for image captioning, visual question answering, or referring expressions. More recent papers use VLM for both representation models and generative assistants. It is therefore useful to identify a model by its interface and objective instead of relying on the acronym alone.

Scroll sideways for more →

A model does not need to receive both modalities in every individual call. A dual encoder, for example, can index image embeddings before any text query arrives. What makes it a VLM is the learned cross-modal relationship between its visual and linguistic representations.

A vision-language-action model is related but has a different output contract: it maps observations and language to actions for an embodied system. RT-2, for example, represented robot actions as tokens and co-fine-tuned on robotic trajectories and vision-language tasks.^[31] Action-producing systems are not treated here as a third architecture family of ordinary VLMs because their control objectives, data, and safety requirements are distinct.

Historical development

Modern VLMs emerged from several lines of work rather than one architecture. Neural image captioning systems joined a convolutional image encoder to a recurrent language decoder. Show and Tell, published in 2015, optimized the likelihood of a caption conditioned on an image and evaluated the approach on several caption datasets.^[1] The VQA dataset, also introduced in 2015, formalized open-ended natural-language questions about images and collected multiple human answers for each question.^[2]

Transformer pretraining then made cross-modal representations reusable across tasks. ViLBERT used separate visual and text streams connected by co-attentional transformer layers, while LXMERT used object, language, and cross-modality encoders with several pretraining tasks.^[3][4] These models commonly represented an image through object regions produced by a detector, rather than through a dense grid of image patches.

In 2021, CLIP and ALIGN showed that paired web images and text could train transferable dual encoders at much larger scale. CLIP used 400 million image-text pairs and a symmetric contrastive objective; in the authors' evaluation, its zero-shot classifier matched the original ResNet-50 on ImageNet without training on ImageNet labels.^[5] ALIGN used a dual encoder and more than one billion noisy image-alt-text pairs, with minimal data cleaning in the reported setup.^[6] Both results were specific to the papers' models, datasets, prompts, and evaluation protocols. They do not imply that any contrastive VLM will match a supervised classifier.

Several subsequent systems combined representation learning with generation. BLIP trained with image-text contrastive, image-text matching, and language-modeling objectives, and used synthetic captions plus filtering to improve noisy web data.^[7] Flamingo connected frozen vision and language components with a Perceiver Resampler and gated cross-attention, allowing interleaved image-text sequences and few-shot prompting.^[8] CoCa trained an image-text encoder-decoder jointly with contrastive and captioning losses.^[9]

Two influential 2023 designs reduced the amount of cross-modal machinery that had to be trained. BLIP-2 used a Querying Transformer, or Q-Former, between frozen image and language models.^[11] LLaVA projected visual features into a language model and instruction-tuned the combined system on generated image-dialogue examples.^[12] These papers established reusable design patterns, but later VLMs vary in whether their encoders, connectors, and language models are frozen, fine-tuned, or trained together.

Architecture

Dual encoders

A dual encoder contains an image encoder (f_I) and a text encoder (f_T). Each maps its input into an embedding space. Retrieval or classification can then use a similarity function such as normalized dot product:

$$s(I,T)=\frac{f_I(I)^\top f_T(T)} {\lVert f_I(I)\rVert\,\lVert f_T(T)\rVert}.$$

During contrastive training, matched image-text pairs are encouraged to receive higher similarity than mismatched pairs. CLIP normalized similarities across the other examples in a batch.^[5] SigLIP instead used a pairwise sigmoid loss, removing the need for a batch-global softmax normalization.^[10]

Dual encoders are efficient when many images or captions must be searched because each side can be encoded once. Their fixed-size embeddings also impose an information bottleneck: fine relationships among objects, attributes, and word order may not be preserved well enough for every downstream task. A similarity score is not a generated explanation, and it should not be interpreted as a calibrated probability unless the system has been evaluated and calibrated for that use.

Fusion encoders

Fusion models let visual and language features interact before producing an answer. ViLBERT's two streams exchange information through co-attention.^[3] LXMERT first builds modality-specific representations and then applies cross-modality layers.^[4] Other fusion designs concatenate visual and text tokens inside a single transformer.

This interaction can support fine-grained tasks such as determining which object a phrase refers to or whether a proposed caption matches an image. The cost is that every candidate image-text pair generally needs a joint forward pass. A common retrieval system therefore uses a fast dual encoder to obtain candidates and a fusion model to rerank them.

Generative VLMs

A generative VLM usually has three functional parts:

1. A visual encoder converts pixels into region, patch, or latent features.
2. A connector maps or selects those features for the language model.
3. A language decoder predicts output tokens conditioned on the prompt and visual representation.

The separation is conceptual rather than mandatory. Some systems train the components end to end, while others reuse frozen models. A visual encoder is often based on a Vision Transformer, but detector features, convolutional networks, or specialized document and video encoders are also possible.

Connectors control the number and form of visual tokens presented to the decoder. A linear layer or multilayer perceptron can project every selected patch feature into the language model's token space, as in LLaVA's basic design.^[12] BLIP-2's Q-Former uses learned query vectors to extract a fixed-size representation from the image encoder.^[11] Flamingo resamples visual features and inserts gated cross-attention layers into a frozen language model.^[8] These choices trade off information retention, context length, training cost, and the ability to handle multiple images.

Image resolution does not translate directly into language-model resolution. A high-resolution page may be resized, divided into tiles, or compressed into a limited token budget. Small text, crowded charts, and spatially close objects can therefore be lost before the decoder receives them. Reporting the visual preprocessing and tokenization scheme is essential when comparing systems.

Hybrid objectives

Architectural labels are not exclusive. CoCa uses an image encoder and a cascaded text decoder so that contrastive loss can train unimodal embeddings while captioning loss trains cross-modal generation.^[9] BLIP combines contrastive, matching, and generation objectives in one pretraining framework.^[7] A deployed product can also compose separate VLMs, OCR models, retrieval systems, and language models. In that case, observed behavior belongs to the complete pipeline, not necessarily to one checkpoint.

Training data and objectives

VLM training depends on paired or interleaved data that connects visual content with language. Sources include human-written captions, question-answer pairs, document transcriptions, alt text, surrounding web text, synthetic descriptions, and dialogue. Each source encodes different assumptions. Alt text may identify the purpose of an image rather than describe all visible content; a question-answer dataset covers only the questions its collectors asked.

Scroll sideways for more →

Web scale is not a substitute for data design. DataComp provided a 12.8-billion-pair candidate pool, standardized training code, four compute scales, and evaluation on 38 downstream test sets so that filtering strategies could be compared under controlled conditions.^[13] Its experiments showed that changing the dataset while holding the training recipe and compute fixed could materially change downstream results. This is evidence about the tested contrastive setup, not a universal ranking of data filters.

Synthetic text can expand task coverage but also import the generator's errors. BLIP generated captions and filtered image-text pairs with a learned matching model.^[7] LLaVA used a language model to generate instruction-following conversations from image descriptions and object information.^[12] For either approach, the provenance of the original image, the generation prompt, the filtering rule, and the retained synthetic output are part of the training specification.

Instruction tuning changes a model's interface and response style as well as its task performance. It can teach a decoder to answer questions, follow requested formats, or sustain dialogue. It does not by itself prove that the visual encoder supplied the evidence used in an answer. Image ablations, counterfactual inputs, and text-only controls can help test whether a response is visually grounded.

Capabilities and tasks

Retrieval and zero-shot classification

Dual encoders support text-to-image and image-to-text retrieval by ranking embedding similarities. They also support zero-shot classification by comparing an image with textual descriptions of candidate classes. CLIP showed that prompt wording and prompt ensembling affected classification results, so a zero-shot score is partly a property of the text templates used in evaluation.^[5]

Retrieval quality depends on the candidate collection and relevance definition. A benchmark that treats one caption as the only correct result can penalize other accurate descriptions, while a web search system may need to handle duplicates, multilingual queries, and safety constraints not represented in a research dataset.

Captioning and visual question answering

Image captioning produces a description of visual content. Caption metrics can measure overlap with reference captions, but a fluent caption can omit important details or introduce an object that is not present. Human review or grounded factuality measures are needed when those errors matter.^[23]

Visual question answering produces an answer to a question about an image. VQA v2 paired related images with questions that have different answers to reduce reliance on language priors.^[14] The task still ranges from short-answer recognition to questions that require reading, external knowledge, or multi-step reasoning. Scores from different subsets or answer-normalization rules are not interchangeable.

Text, documents, charts, and diagrams

TextVQA contains 45,336 questions on 28,408 natural images and was designed to require reading scene text.^[15] DocVQA introduced 50,000 questions over more than 12,000 document images, including questions that depend on document structure.^[16] ChartQA combines questions about chart content with visual and logical reasoning.^[17]

These tasks involve more than generic object recognition. A system may need to preserve small glyphs during resizing, infer reading order, associate labels with graphical marks, and copy an exact string into the answer. Results are sensitive to whether external OCR is allowed and whether a model sees the original document, a rendered page, cropped regions, or extracted text.

Multi-image and video input

Some VLMs accept several images or sampled video frames in one context. Flamingo was explicitly trained for arbitrarily interleaved visual and textual data and evaluated on image and video tasks.^[8] A model that accepts a video file does not necessarily process every frame. Frame sampling, temporal order, audio availability, and context limits should be reported before attributing a result to video understanding.

Grounding and structured output

Grounding tasks connect words to regions, boxes, points, or masks. They test a more localized relationship than whole-image captioning or retrieval. Generative models can serialize coordinates as text, but a syntactically valid coordinate is not necessarily spatially accurate. Evaluation should specify coordinate normalization, image resizing, tolerance, and how invalid outputs are handled.

Evaluation

No single benchmark measures "vision-language understanding" as a whole. Evaluation should start with the model's intended interface, then separate perception, text recognition, cross-modal association, reasoning, generation quality, calibration, and safety.

Scroll sideways for more →

MMMU-Pro illustrates why benchmark construction matters. It was built from MMMU by filtering questions that text-only models could answer, increasing the number of answer options, and adding a setting in which the question and options appear inside an image.^[30] The reported score drops are evidence about the evaluated models and those transformations. They are not proof that every retained question requires the same visual skills.

Benchmark artifacts can also produce false confidence. Winoground controls vocabulary by pairing two images with two captions that use the same words in different orders; the models evaluated in its original paper performed poorly on the paired tests.^[20] ARO expanded tests of attributes, relations, and order to more than 50,000 cases and found weaknesses in the contrastive models it evaluated.^[21] SugarCrepe then showed that several compositionality benchmarks contained linguistic artifacts: blind models without access to the image could outperform VLMs on many of the tested tasks. Its authors designed adversarially refined negatives to reduce those shortcuts.^[22]

Reproducible comparison

A defensible model comparison records at least:

• the exact checkpoint or service version and evaluation date;
• the image preprocessing, resolution, tiling, and frame-sampling rules;
• the prompt, demonstrations, system message, and answer format;
• decoding settings and the number of repeated runs;
• whether OCR, retrieval, tools, or external knowledge were available;
• the dataset version, split, exclusions, and contamination checks;
• the answer parser, judge model if any, and metric implementation;
• uncertainty intervals or variation across examples when available.

An open-ended answer may require semantic grading, but an automated judge can introduce its own preferences and errors. Exact match is reproducible but can reject a correct paraphrase. Multiple-choice accuracy is easy to calculate but can be affected by option order and guessing. Publishing raw outputs and evaluation code allows others to inspect these trade-offs.

Public benchmark data may have been present in a model's pretraining corpus, especially when the training data are undisclosed. A high score alone cannot distinguish learned task competence from exposure to related examples. Image ablations, new or private test items, temporal holdouts, and tests with controlled transformations provide complementary evidence, but none is a universal contamination detector.

Limitations and risks

Hallucination and weak visual grounding

A generative VLM can produce plausible text that is unsupported by the image. CHAIR was introduced after its authors found that standard captioning metrics did not adequately capture object hallucination; the metric checks whether mentioned objects are present according to image annotations.^[23] POPE evaluated object-presence questions and found that the tested large VLMs could favor objects that commonly co-occurred with visible content or appeared frequently in instructions.^[24]

These methods cover particular error types. CHAIR depends on available object annotations, and POPE converts generation into a polling protocol. Neither detects every wrong attribute, relation, transcription, count, or external fact. Deployment evaluation should define the factual units that matter for the task and test them directly.

Composition, spatial relations, and visual primitives

Whole-image alignment can succeed without preserving exact relations among every object and word. Winoground, ARO, and SugarCrepe each probe aspects of compositionality, while also demonstrating that the benchmark itself must be checked for shortcuts.^[20][21][22] Their results apply to the models and protocols tested; they do not support the claim that all VLMs are equivalent to bags of words.

Research published in 2026 tested three generative VLMs on 51 tests drawn from six neuropsychological and experimental batteries. The authors reported strong object recognition alongside deficits on several low- and mid-level visual abilities, including tasks involving orientation, position, continuity, and occlusion.^[32] This is a bounded comparison with human normative data, not a clinical diagnosis of a model or a complete account of machine vision.

Modality gap

Contrastive training does not necessarily mix image and text embeddings into one identical distribution. Liang and colleagues measured a separation between modalities in several contrastive models and analyzed how initialization, optimization, and the loss temperature contributed in their experiments.^[25] The term "modality gap" refers to this studied geometric effect. It should not be used as a catch-all explanation for every VLM error or assumed to have the same form in fusion and generative architectures.

Calibration and uncertainty

Output confidence and correctness can diverge. An ICML 2024 study evaluated calibration across VLM architectures, datasets, distribution shifts, and label sets. It found that the tested models were not inherently calibrated, while temperature scaling improved calibration in the studied zero-shot classification settings.^[27] A 2026 TMLR study tested late-2024 and early-2025 VLMs on anomaly detection and inherently ambiguous classification and found that newer or larger models did not eliminate these uncertainty challenges.^[28]

Calibration results are task-specific. A calibrated class probability does not automatically calibrate a free-form answer, and a model's verbal confidence is not guaranteed to equal an empirical probability. Consequential systems need task-level error measurement, abstention policies, and human escalation rather than confidence language alone.

Data coverage and social bias

Web image-text data reflect which languages, locations, people, and descriptions were available and retained. CLIP's authors documented sensitivity to class design and reported demographic performance disparities in several probes.^[5] Dataset filters can remove some unwanted content while changing coverage in other ways.^[13]

Multilingual support is especially uneven. MVL-SIB evaluates topical image-text matching in 205 languages. In the models tested by its authors, cross-modal performance declined disproportionately relative to text-only performance for lower-resource languages, with chance-level results for some languages.^[26] This benchmark does not represent every language use or culture, but it shows why English-only averages cannot establish multilingual reliability.

Security

Visual content can contain instructions as well as scene information. A 2025 Nature Communications study embedded visible or low-visibility instructions in oncology images and tested four commercial VLMs that met the study's inclusion threshold. Across 594 evaluated attacks, all four were susceptible under at least some tested conditions.^[29] The result is specific to those models, dates, prompts, and medical images, but it demonstrates that untrusted pixels can act as an attack channel.

A VLM that can call tools or act on private data should therefore treat image-derived instructions as untrusted input. Permissions, data isolation, logging, output validation, and approval for consequential actions limit the impact of a compromised response. A content filter alone does not establish a security boundary between an image's data and any text the model reads within it.

Resolution, efficiency, and reproducibility

More visual tokens can preserve detail but increase memory and computation in attention layers. Compression and tiling can reduce cost while changing spatial context. Published parameter counts also omit parts of a service pipeline, such as external OCR, retrieval, or proprietary routing.

For closed models, architecture, training data, and post-training details may be undisclosed, and a service can change without a new paper. Claims about a named service should therefore include the access date and version identifier when available. For open-weight models, weights alone may still be insufficient to reproduce results without the preprocessing code, prompt templates, evaluation parser, and exact dependencies.

Deployment considerations

VLMs are used in retrieval, accessibility support, media analysis, document workflows, and scientific or medical research. Suitability depends on the cost of an error and the controls around the model, not only on a benchmark average.

Before deployment, a system owner should define which visual facts the model must extract, which errors require abstention, and what data may leave the local environment. Evaluation data should reflect the actual image sources, languages, resolutions, and failure costs. Tests should include blank or irrelevant images, misleading text inside images, corrupted files, repeated queries, and cases where the correct response is to decline.

Human review is most useful when reviewers can inspect the original visual evidence and the model's extracted claims. A human-in-the-loop label alone is insufficient if reviewers lack time, expertise, or authority to reject an output. In high-consequence settings, the system should preserve the source image, prompt, model version, and result so that an error can be reconstructed.

Relationship to adjacent fields

VLM research overlaps with computer vision, natural language processing, contrastive learning, instruction tuning, and transfer learning. It also overlaps with multimodal generation, but an image generator that maps text to pixels is not necessarily called a VLM in the same sense as an image-text encoder or visual question-answering model.

The most informative description states the visual input, language input or output, architecture, training objectives, and evaluation protocol. That description remains meaningful when terminology or model branding changes.

See also

• Multimodal model
• CLIP
• Vision Transformer
• Visual question answering
• OCR Models
• MMMU
• MathVista
• Vision-language-action model

References

1. ^Vinyals, O., Toshev, A., Bengio, S., & Erhan, D. (2015). "Show and Tell: A Neural Image Caption Generator." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 3156-3164. openaccess.thecvf.com/..._and_Tell_2015_CVPR_paper
2. ^Antol, S. et al. (2015). "VQA: Visual Question Answering." Proceedings of the IEEE International Conference on Computer Vision, 2425-2433. openaccess.thecvf.com/..._Question_ICCV_2015_paper
3. ^Lu, J., Batra, D., Parikh, D., & Lee, S. (2019). "ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks." Advances in Neural Information Processing Systems 32. papers.nips.cc/...eae257e44aa9d5bade97baf-Abstract
4. ^Tan, H., & Bansal, M. (2019). "LXMERT: Learning Cross-Modality Encoder Representations from Transformers." Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, 5100-5111. aclanthology.org/D19-1514
5. ^Radford, A. et al. (2021). "Learning Transferable Visual Models From Natural Language Supervision." Proceedings of the 38th International Conference on Machine Learning, 8748-8763. proceedings.mlr.press/...radford21a
6. ^Jia, C. et al. (2021). "Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision." Proceedings of the 38th International Conference on Machine Learning, 4904-4916. proceedings.mlr.press/...jia21b
7. ^Li, J., Li, D., Xiong, C., & Hoi, S. (2022). "BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation." Proceedings of the 39th International Conference on Machine Learning, 12888-12900. proceedings.mlr.press/...li22n
8. ^Alayrac, J.-B. et al. (2022). "Flamingo: a Visual Language Model for Few-Shot Learning." Advances in Neural Information Processing Systems 35. papers.nips.cc/...cbb411a7d800-Abstract-Conference
9. ^Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., & Wu, Y. (2022). "CoCa: Contrastive Captioners are Image-Text Foundation Models." Transactions on Machine Learning Research. research.google/...re-image-text-foundation-models
10. ^Zhai, X., Mustafa, B., Kolesnikov, A., & Beyer, L. (2023). "Sigmoid Loss for Language Image Pre-Training." Proceedings of the IEEE/CVF International Conference on Computer Vision, 11975-11986. openaccess.thecvf.com/...-Training_ICCV_2023_paper
11. ^Li, J., Li, D., Savarese, S., & Hoi, S. (2023). "BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models." Proceedings of the 40th International Conference on Machine Learning, 19730-19742. proceedings.mlr.press/...li23q
12. ^Liu, H., Li, C., Wu, Q., & Lee, Y. J. (2023). "Visual Instruction Tuning." Advances in Neural Information Processing Systems 36. papers.nips.cc/...faf369fe6de0-Abstract-Conference
13. ^Gadre, S. Y. et al. (2023). "DataComp: In search of the next generation of multimodal datasets." Advances in Neural Information Processing Systems 36, Datasets and Benchmarks Track. papers.nips.cc/...Abstract-Datasets_and_Benchmarks
14. ^Goyal, Y. et al. (2017). "Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 6904-6913. openaccess.thecvf.com/...ing_the_v_CVPR_2017_paper
15. ^Singh, A. et al. (2019). "Towards VQA Models That Can Read." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8317-8326. openaccess.thecvf.com/..._Can_Read_CVPR_2019_paper
16. ^Mathew, M., Karatzas, D., & Jawahar, C. V. (2021). "DocVQA: A Dataset for VQA on Document Images." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2200-2209. openaccess.thecvf.com/...nt_Images_WACV_2021_paper
17. ^Masry, A., Long, D. X., Tan, J. Q., Joty, S., & Hoque, E. (2022). "ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning." Findings of the Association for Computational Linguistics: ACL 2022, 2263-2279. aclanthology.org/2022.findings-acl.177
18. ^Yue, X. et al. (2024). "MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9556-9567. openaccess.thecvf.com/...hmark_for_CVPR_2024_paper
19. ^Lu, P. et al. (2024). "MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts." International Conference on Learning Representations. proceedings.iclr.cc/...f1429d3-Abstract-Conference
20. ^Thrush, T. et al. (2022). "Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5238-5248. openaccess.thecvf.com/...tionality_CVPR_2022_paper
21. ^Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., & Zou, J. (2023). "When and Why Vision-Language Models Behave like Bags-Of-Words, and What to Do About It?" International Conference on Learning Representations. iclr.cc/...10875
22. ^Hsieh, C.-Y., Zhang, J., Ma, Z., Kembhavi, A., & Krishna, R. (2023). "SugarCrepe: Fixing Hackable Benchmarks for Vision-Language Compositionality." Advances in Neural Information Processing Systems 36, Datasets and Benchmarks Track. papers.nips.cc/...Abstract-Datasets_and_Benchmarks
23. ^Rohrbach, A., Hendricks, L. A., Burns, K., Darrell, T., & Saenko, K. (2018). "Object Hallucination in Image Captioning." Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 4035-4045. aclanthology.org/D18-1437
24. ^Li, Y. et al. (2023). "Evaluating Object Hallucination in Large Vision-Language Models." Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 292-305. aclanthology.org/2023.emnlp-main.20
25. ^Liang, V. W., Zhang, Y., Kwon, Y., Yeung, S., & Zou, J. (2022). "Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning." Advances in Neural Information Processing Systems 35. papers.nips.cc/...f588d57bc7c9-Abstract-Conference
26. ^Schmidt, F. D., Schneider, F., Biemann, C., & Glavas, G. (2025). "MVL-SIB: A Massively Multilingual Vision-Language Benchmark for Cross-Modal Topical Matching." Findings of the Association for Computational Linguistics: ACL 2025, 16285-16312. aclanthology.org/2025.findings-acl.838
27. ^Tu, W., Deng, W., Campbell, D., Gould, S., & Gedeon, T. (2024). "An Empirical Study Into What Matters for Calibrating Vision-Language Models." Proceedings of the 41st International Conference on Machine Learning, 48791-48808. proceedings.mlr.press/...tu24a
28. ^Wang, X., & Nalisnick, E. (2026). "Are vision language models robust to classic uncertainty challenges?" Transactions on Machine Learning Research. jmlr.org/...4lCSYCNfmo.bib
29. ^Clusmann, J. et al. (2025). "Prompt injection attacks on vision language models in oncology." Nature Communications, 16, 1239. nature.com/...s41467-024-55631-x
30. ^Yue, X. et al. (2025). "MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark." Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 15134-15186. aclanthology.org/2025.acl-long.736
31. ^Zitkovich, B. et al. (2023). "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control." Proceedings of the 7th Conference on Robot Learning, 2165-2183. proceedings.mlr.press/...zitkovich23a
32. ^Tangtartharakul, G., & Storrs, K. R. (2026). "Visual language models show widespread visual deficits on neuropsychological tests." Nature Machine Intelligence, 8, 209-219. nature.com/...s42256-026-01179-y