20240620/2307.06930v3.json

{
    "title": "mBLIP: Efficient Bootstrapping of Multilingual Vision-LLMs",
    "abstract": "Modular vision-language models (Vision-LLMs) align pretrained image encoders with (frozen) large language models (LLMs)\nand post-hoc condition LLMs to \u2018understand\u2019 the image input.\nWith the abundance of readily available high-quality English image-text data as well as strong monolingual English LLMs, the research focus has been on English-only Vision-LLMs.\nMultilingual vision-language models are still predominantly obtained via expensive end-to-end pretraining, resulting in comparatively smaller models, trained on limited multilingual image data supplemented with text-only multilingual corpora.\nWe present mBLIP, the first Vision-LLM leveraging multilingual LLMs, which we obtain in a computationally efficient manner on consumer-level hardware.\nTo this end, we re-align an image encoder previously tuned to an English LLM to a new, multilingual LLM using only a few million multilingual training examples derived from a mix of vision-and-language tasks, which we obtain by machine-translating high-quality English data to 95 languages.\nOn the IGLUE benchmark and XM3600, mBLIP yields results competitive with state-of-the-art models and it greatly outperforms strong English-only Vision-LLMs like Llava 1.5.\nWe release our model, code, and train data at https://github.com/gregor-ge/mBLIP.",
    "sections": [
        {
            "section_id": "1",
            "parent_section_id": null,
            "section_name": "Introduction",
            "text": "The success of model and data scaling in NLP from BERT (Devlin et al., 2019  ###reference_b17###) to more recent Large Language Models (LLMs) (Brown et al., 2020  ###reference_b3###; Zhang et al., 2022  ###reference_b76###; Touvron et al., 2023  ###reference_b59###, inter alia) has prompted similar endeavors in vision-language pretraining from \u2018small\u2019 BERT-size models (Chen et al., 2020  ###reference_b8###; Li et al., 2020  ###reference_b31###, 2021  ###reference_b29###, 2022  ###reference_b28###) trained on a few million image-text pairs to billion-parameter models trained with billions of examples (Wang et al., 2021  ###reference_b65###; Yu et al., 2022  ###reference_b71###; Wang et al., 2022  ###reference_b64###; Chen et al., 2022  ###reference_b7###, 2023  ###reference_b6###).\nThe prohibitive cost of such end-to-end (pre)training, however, has resulted in increased interest in efficient modular methods that leverage existing large language models (LLMs). These align the output of a pretrained image encoder to the LLM\u2019s input representation space (Tsimpoukelli et al., 2021  ###reference_b60###; Alayrac et al., 2022  ###reference_b1###; Li et al., 2023a  ###reference_b27###), resulting in a Vision-LLM.\nPretraining vision-language models from scratch requires a massive amount of high-quality image-text data, which is only available in English.\nBecause of this, multilingual pretraining of vision-language models (Ni et al., 2021  ###reference_b41###; Zhou et al., 2021  ###reference_b77###; Zeng et al., 2023  ###reference_b73###; Shan et al., 2022  ###reference_b54###; Li et al., 2023c  ###reference_b33###) commonly supplements limited-size multilingual image-text data with multilingual text-only data (the amount of which often surpasses that of image-text data) to achieve strong results, despite initialization with weights of multilingual text encoders such as XLM-R (Conneau et al., 2020  ###reference_b11###).\nIn this work, we recognize modular Vision-LLM methods as a potential solution to this problem, observing that: (1) once an image encoder is aligned to one LLM, it requires significantly less data to re-align it to another LLM (Zhang et al., 2023  ###reference_b75###; Zhu et al., 2023  ###reference_b78###) and (2) since image encoding is, in principle, language-agnostic, it may be possible to successfully re-align the image encoder to a strong multilingual LLM, even if it was initially aligned only with English image-text data.\nBased on these observations, we present mBLIP, the first massively multilingual modular Vision-LLM, which we obtain by (re-)aligning an image encoder to a multilingual LLM.\nPutting together a range of recent advances in multimodal representation learning, we efficiently bootstrap a massively multilingual Vision-LLM using only 2.5 million images (and without any additional multilingual text-only data), training only 124 million parameters on consumer-grade hardware. We achieve this efficiency by: 1) bootstrapping our model from a) an \u201cEnglish\u201d image encoder (Li et al., 2023a  ###reference_b27###), previously aligned to a monolingual English LLM and b) a strong instruction-tuned multilingual LLM (Xue et al., 2021  ###reference_b69###; Scao et al., 2022  ###reference_b50###; Muennighoff et al., 2022  ###reference_b40###); 2) leveraging recent advances in massively multilingual machine translation (Costa-juss\u00e0 et al., 2022  ###reference_b12###), which we use to translate high-quality English data\u2014both classic captions as well as task instructions (Dai et al., 2023  ###reference_b13###)\u2014to 95 languages; and finally 3) coupling parameter-efficient training methods (Hu et al., 2022  ###reference_b24###) together with quantization (Dettmers et al., 2022  ###reference_b15###, 2023  ###reference_b16###) to enable training on consumer-grade hardware.\nWe extensively evaluate mBLIP on different multilingual vision-language tasks to confirm the efficacy of our approach: for multilingual image captioning, mBLIP (with mT0-XL) surpasses (zero-shot) PaLI-X (a model with 55B parameters, trained with billions of examples) (Chen et al., 2023  ###reference_b6###) on the XM3600 (Thapliyal et al., 2022  ###reference_b58###).\nOn the visual reasoning and QA tasks of the IGLUE benchmark (Bugliarello et al., 2022  ###reference_b4###), mBLIP matches or surpasses the performance of state-of-the-art models, despite training far fewer parameters on far less pretraining data.\nWe consistently outperform state-of-the-art English Vision-LLMs outside of English, highlighting the multilingual prowess of our model."
        },
        {
            "section_id": "2",
            "parent_section_id": null,
            "section_name": "Related Work",
            "text": ""
        },
        {
            "section_id": "2.1",
            "parent_section_id": "2",
            "section_name": "LLMs and Images",
            "text": "The success of scaling up training data and model parameters has resulted in large vision-language models with billions of parameters (Wang et al., 2021  ###reference_b65###; Yu et al., 2022  ###reference_b71###; Wang et al., 2022  ###reference_b64###).\nHowever, with the number of parameters in single-digit billions, these are still an order of magnitude smaller than text-only models (Brown et al., 2020  ###reference_b3###); the compute necessary to pretrain comparably large vision-language models, however, is available only to select few (Chen et al., 2022  ###reference_b7###, 2023  ###reference_b6###).\nInstead, much of the vision-language research turned to approaches that can leverage the power of existing LLMs\nby training an image encoder to map an image into a sequence of tokens in the LLM embedding space (Tsimpoukelli et al., 2021  ###reference_b60###; Alayrac et al., 2022  ###reference_b1###; Li et al., 2023a  ###reference_b27###), while the LLM is kept as-is or is only partially tuned (Alayrac et al., 2022  ###reference_b1###).\nMost recently, the release of strong publicly available LLMs such as Llama (Touvron et al., 2023  ###reference_b59###) and the success of conversational instruction tuning (Ouyang et al., 2022  ###reference_b43###; Taori et al., 2023  ###reference_b57###; Chiang et al., 2023  ###reference_b9###; Xu et al., 2023  ###reference_b68###), has led to a body of work (Zhu et al., 2023  ###reference_b78###; Liu et al., 2023b  ###reference_b37###; Ye et al., 2023  ###reference_b70###; Dai et al., 2023  ###reference_b13###; Gao et al., 2023  ###reference_b21###; Liu et al., 2023a  ###reference_b36###; Bai et al., 2023  ###reference_b2###) that tries to replicate the vision-language skills of GPT-4 (OpenAI, 2023  ###reference_b42###). The vast majority of research focused on English, where both an abundance of high-quality image-text data and strong LLMs exist. To the best of our knowledge, we are the first to extend a massively multilingual LLM with \u201cvision capabilities\u201d."
        },
        {
            "section_id": "2.2",
            "parent_section_id": "2",
            "section_name": "Multilingual Vision-Language Models",
            "text": "While the majority of research on vision-language models targets English only, a number of multilingual models have been proposed too. M3P (Ni et al., 2021  ###reference_b41###), the first transformer-based (Vaswani et al., 2017  ###reference_b61###) multilingual vision-language model, adopts the architecture and pretraining objectives of English counterparts (Chen et al., 2020  ###reference_b8###; Li et al., 2020  ###reference_b31###).\nbut trains on (i) the code-switched image-text data\nin which words in English image captions are replaced with translations from various languages\nas well as (ii) additional text-only multilingual corpora.\nUC2 (Zhou et al., 2021  ###reference_b77###) uses a similar architecture and a mix of training objectives but instead of code-switching, it machine translates the 3M captions of CC3M (Sharma et al., 2018  ###reference_b55###) to 5 languages (German, French, Czech, Japanese, and Chinese).\nLi et al. (2023c  ###reference_b33###) and CCLM (Zeng et al., 2023  ###reference_b73###), which adopt the ALBEF architecture (Li et al., 2021  ###reference_b29###) that incorporates additional contrastive learning objectives, use the same translated CC3M data but they additionally supplement 19M parallel sentences (pairing English with all of the languages spanned by their respective downstream evaluation tasks).\nERNIE-UniX2 (Shan et al., 2022  ###reference_b54###), with an encoder-decoder architecture, adopts the same pretraining objectives but scales up the data to more translated captions and more text-only data (both aligned and monolingual).\nFinally, PaLI (Chen et al., 2022  ###reference_b7###) (17B parameters) and PaLI-X (Chen et al., 2023  ###reference_b6###) (55B parameters) represent two huge encoder-decoder models trained using a mixture of vision-and-language tasks, with billions of web-crawled multilingual captions, machine translated data, automatically extracted data (e.g., OCR and object detection), and generated visual QA (VQA) examples.\nWith the exception of the PaLI models and ERNIE-UniX2 \u2013 both of which are not publicly available \u2013 all other multilingual vision-language models represent encoder-only architectures, which cannot perform image captioning out of the box.\n###figure_1###"
        },
        {
            "section_id": "3",
            "parent_section_id": null,
            "section_name": "mBLIP",
            "text": "We first briefly describe the modular BLIP-2 architecture Li et al. (2023a  ###reference_b27###) which we adopt in this work, followed by the description of training tasks and data, which we translate to 95 languages."
        },
        {
            "section_id": "3.1",
            "parent_section_id": "3",
            "section_name": "Architecture",
            "text": "We follow the modular BLIP-2 architecture (Li et al., 2023a  ###reference_b27###) depicted in Figure 1  ###reference_###:\nA Query-Former (Q-Former) is an encoder-only transformer (Vaswani et al., 2017  ###reference_b61###) with 32 learned query tokens as input: it contextualizes the query tokens \u2013 via the cross-attention mechanism \u2013 with the representations of the image patches encoded by a large (frozen) Vision Transformer (ViT) (Dosovitskiy et al., 2020  ###reference_b18###). The visual tokens that are the output of the Q-Former are then projected into the LLM embedding space with a single linear projection matrix , with  and  as hidden dimensions (i.e., embedding dimensionality) of the Q-Former and LLM, respectively.\nDuring training, only the the Q-Former (including the 32 query tokens) and the linear projection  are updated; all ViT and LLM parameters are kept frozen.\nAlthough the Q-Former and projection have initially been aligned to a monolingual English LLM, they only produce visual tokens: we believe that as such they are not overly tailored to English and can therefore be effectively re-aligned to a different, multilingual LLM.\nBecause the LLM is frozen in the BLIP-2 training, its parameters cannot adapt to task-specific idiosyncrasies, e.g., in fine-tuning for VQA or for instruction-following (Dai et al., 2023  ###reference_b13###). Instead, task-specific fine-tuning of BLIP-2 requires that the text input is not just fed into the LLM but also into the Q-Former in order to enable encoding of task-specific visual information from the input. The Q-Former, however, is based on the English BERT Devlin et al. (2019  ###reference_b17###), preventing the application of this same approach in the multilingual setting (i.e., we cannot feed the text in other languages into the Q-Former nor efficiently make it massively multilingual, i.e., without a large multilingual pre-training effort). Because of this, we opt for a different approach: instead of feeding the text of the image-text instance (e.g., in VQA) to the Q-Former, we partially update the LLM with the parameter-efficient LoRA (Hu et al., 2022  ###reference_b24###), which trains low-rank reparametrization of the LLM matrices."
        },
        {
            "section_id": "3.2",
            "parent_section_id": "3",
            "section_name": "Training Tasks and Data",
            "text": "We create a small but high-quality mix of tasks for our re-alignment training. We start from existing high-quality English data and machine-translate it to 95 languages in order to obtain multilingual training data for re-alignment of the Q-Former to the multilingual LLM.111Training with only English data, even without LoRA, results in the LLM producing only English output.  We hypothesized that the re-alignment to a new LLM can be done with significantly less data than what is needed to train the original Q-Former (Zhu et al., 2023  ###reference_b78###; Zhang et al., 2023  ###reference_b75###). Accordingly, we create a small, high-quality English datasets and make it multilingual via MT rather than training with large-scale but very noisy multilingual image-caption datasets like LAION5B (Schuhmann et al., 2022  ###reference_b52###). In addition, in line with findings from language-only instruction-tuning (Sanh et al., 2022  ###reference_b49###; Muennighoff et al., 2022  ###reference_b40###; Chung et al., 2022  ###reference_b10###) and vision-language training Dai et al. (2023  ###reference_b13###); Liu et al. (2023b  ###reference_b37###, a  ###reference_b36###); Bai et al. (2023  ###reference_b2###), we expect the training on a mixture of vision-and-language tasks (as opposed to training only for image captioning), with different task instructions, to result in better generalization abilities of the\nmodel and improve its (zero-shot) downstream performance and usability.\nTask Mix: We select below the tasks and datasets used to create our training mix for re-alignment (naturally, we ensure that the data does not overlap with our downstream evaluation data; see \u00a74.1  ###reference_###). For every task, we create a set of instruction templates with which we generate the training examples (we provide the templates in \u00a7D.1  ###reference_### in the Appendix, along with additional details about the training data). In total, across all tasks, we use 5.1M examples encompassing 2.7M unique images.\n1. Image Captioning: We use MSCOCO (Lin et al., 2014  ###reference_b34###) along with 2.3 million examples sampled from the synthetic CapFilt dataset (Li et al., 2022  ###reference_b28###) with the noun phrase method by Liu et al. (2023b  ###reference_b37###) to ensure concept diversity. Additionally, we use LLaVA-Instruct-Detail (Liu et al., 2023b  ###reference_b37###), which contains longer and more detailed captions. \n2. Visual Question Answering and Generation: For VQA and the inverse task of question generation (given the answer, the model is supposed to produce the question), we use VQAv2 (Goyal et al., 2017  ###reference_b23###).\nAdditionally, we split the conversations from LLaVA-Instruct-Conversation into separate VQA pairs.\nWe use A-OKVQA (Schwenk et al., 2022  ###reference_b53###), a knowledge-intensive VQA dataset with rationales behind the answers, to create data for two additional task variants: 1) given the question, generate the answer and the rationale behind it, 2) given the question and the answer, generate the rationale.\nFinally, we use ImageNet (Deng et al., 2009  ###reference_b14###) with the multilingual labels from Babel-ImageNet (Geigle et al., 2023  ###reference_b22###) framed as an open-ended QA task (with questions like \u201cWhat is in the image?\u201d and no predefined answer choices). \n3. Matching: Inspired by image-text matching (Lu et al., 2019  ###reference_b39###), where an encoder has to classify if caption and image match, we propose a yes/no matching task so that the model learns what is and what is not in the image to reduce hallucinations when interrogating for image content (Li et al., 2023b  ###reference_b32###). For this, we use the Web CapFilt captions for \u201cstandard\u201d caption matching with hard negatives. We also use the ImageNet examples with multilingual class labels, where the model has to predict if a given class is in the image or not.\nMachine Translation:\nWe translate the above English data with NLLB (Costa-juss\u00e0 et al., 2022  ###reference_b12###) (nllb-200-distilled-1.3B), a recent massively multilingual MT model that exhibits strong performance also for low(er)-resource languages.\nTo extend the utility of mBLIP to languages beyond what is covered by existing multilingual evaluation benchmarks,\nwe translate the English data to all languages from the mC4 corpora (Xue et al., 2021  ###reference_b69###),222tensorflow.org/datasets/catalog/c4#c4multilingual  ###reference_g/c4#c4multilingual### excluding only a handful of languages not supported by NLLB.333Excluded are (ISO-1/3 codes): fy, haw, hmn, la, and co.\nOur final training dataset thus covers 96 languages (English and 95 translation languages).\nTranslating all English training instances to every target language would result in a 96 times larger dataset (w.r.t. the original English data) and, consequently, prohibitively expensive re-alignment training. We thus translate English instances to target languages in proportion to the languages\u2019 representation in mC4 (e.g., we translate 6% of English instances to German, because German represents 6% of the mC4 corpus).\nWe do not translate the short answers in A-OKVQA nor most VQAv2 examples444See \u00a7D.1  ###reference_### for details. In short, we limit to the top-1500 answers and use consistency with back-translations to filter incorrect translation. We also still use English half the time. because translating them without context is overly error-prone.\nOutput Language: Essential for multilingual models is control over the output language and minimizing language hallucinations (,i.e., output in an unwanted language) (Xue et al., 2021  ###reference_b69###; Vu et al., 2022  ###reference_b63###; Pfeiffer et al., 2023  ###reference_b45###; Li and Murray, 2023  ###reference_b30###).\nWe achieve this by combining English prompts that explicitly specify the target language (e.g., \u201cAnswer in French.\u201d) and translating the instructions for image captioning and LLaVA (Liu et al., 2023b  ###reference_b37###) to the target languages (other templates contain placeholders that make translation difficult)."
        },
        {
            "section_id": "4",
            "parent_section_id": null,
            "section_name": "Experiments",
            "text": ""
        },
        {
            "section_id": "4.1",
            "parent_section_id": "4",
            "section_name": "Evaluation Tasks and Setup",
            "text": "We evaluate our model on a range of languages on (1) classification-style VQA and image understanding tasks, where the model generates a short answer in response to a question or premise and (2) image captioning tasks, where the model describes an image. For VQA and image captioning, we ensured that no evaluation instances were used in re-alignment training. In contrast to VQA and image captioning, the model was not exposed to image understanding during re-alignment: these tasks thus test the model\u2019s cross-task generalization abilities.\nTo generate outputs, we use beam search with the beam width of  and a length penalty of  for classification-style tasks to encourage short answers. We provide the exact instruction-tuning templates for each task/dataset in \u00a7D.2  ###reference_###.\nImage Captioning:\nXM3600 (Thapliyal et al., 2022  ###reference_b58###) is a captioning dataset covering 36 languages, 3600 images, and 2 captions per image and language.\nxFlickrCo (Bugliarello et al., 2022  ###reference_b4###) combines the 1000 Flickr30k (Plummer et al., 2015  ###reference_b46###) test images with 1000 images from the MSCOCO (Lin et al., 2014  ###reference_b34###) test split555These captions were created from scratch and not by translating existing MSCOCO captions so this does not constitute leakage from the MSCOCO data of the training mix. and provides one new caption for each image in 8 languages. For the English xFlickrCo results, we use the standard Flickr30k test split (i.e., without MSCOCO images and with 5 reference captions per image).\nWe use CIDEr (Vedantam et al., 2015  ###reference_b62###) as the evaluation metric666Implementation: pycocoeval  ###reference_###\nFor Chinese, Japanese, and Thai, which do not use white space for tokenization, we use the default spaCy 3.5.3 segmenter for the respective languages; our results on those languages are thus not directly comparable to previous work \u2013 which, unfortunately, does not disclose the used tokenizer (Thapliyal et al., 2022  ###reference_b58###; Chen et al., 2022  ###reference_b7###, 2023  ###reference_b6###).\nVQA:\nwe leverage xGQA (Pfeiffer et al., 2022  ###reference_b44###) and MaXM (Changpinyo et al., 2022  ###reference_b5###), two VQA datasets with 8 and 7 languages, respectively. While answers in xGQA are in English (as only the original GQA (Hudson and Manning, 2019  ###reference_b25###) questions were translated), answers in MaXM are in the language of the question.\nWe evaluate our model in zero-shot inference (i.e., without any additional fine-tuning other than the VQA training included in the re-alignment mix) on both datasets. For xGQA, we additionally fine-tune the model on the training portion of the English GQA and perform cross-lingual zero-shot transfer.777Note that by zero-shot cross-lingual transfer here we refer to the fact that the model has been fine-tuned only on the English GQA data; in re-alignment training, however, it has been exposed to VQA from other datasets. We use exact match accuracy with open generation, that is, we do not constrain the generation to a fixed set of labels like, e.g., Zeng et al. (2023  ###reference_b73###). For MaXM, an exact match to any one of the answer candidates is correct, as proposed by Changpinyo et al. (2022  ###reference_b5###).\nImage Understanding:\nXVNLI (Bugliarello et al., 2022  ###reference_b4###; Xie et al., 2019  ###reference_b67###) is a visual entailment task that covers 5 languages: given an image and a statement, the model has to decide if the image entails, contradicts or is neutral to the statement. MaRVL (Liu et al., 2021  ###reference_b35###) is based on NLVR2 (Suhr et al., 2019  ###reference_b56###) with new images and concepts spanning different cultures in 6 languages: given two images, the model has to decide if a statement is true or false.\nWe separately encode the two images with the Q-Former and then concatenate their visual tokens together as input for the LLM.\nLike for xGQA, we evaluate the models on XVNLI and MaRVL with (1) zero-shot inference (i.e., no fine-tuning for XVNLI and MaRVL) and (2) supervised cross-lingual transfer: we fine-tune the re-aligned model on the English training portions (of XVNLI and NLVR2, respectively) and evaluate its performance on the test portions of target languages. We report the results in terms of exact match accuracy."
        },
        {
            "section_id": "4.2",
            "parent_section_id": "4",
            "section_name": "Implementation Details",
            "text": "Architecture: We initialize the mBLIP\u2019s ViT (EVA CLIP ViT-g/14 (Fang et al., 2022  ###reference_b19###)) and Q-Former with the BLIP-2 Flan-T5-XL checkpoint. For the multilingual LLM, we experiment with mT0-XL and BLOOMZ-7B (Muennighoff et al., 2022  ###reference_b40###), the instruction-tuned versions of mT5-XL (Xue et al., 2021  ###reference_b69###) and BLOOM-7B (Scao et al., 2022  ###reference_b50###).\nWe use 8/4-bit quantization (Dettmers et al., 2022  ###reference_b15###, 2023  ###reference_b16###).\nWarmup: Similar to Zhang et al. (2023  ###reference_b75###); Liu et al. (2023b  ###reference_b37###),\nwe first train\nonly the linear projection between the Q-Former and LLM. with 1M captions.\nRe-Alignment Training:\nWe train on the re-alignment task mixture for 80k steps (2 epochs), which takes 4 days (mT0) and 6 days (BLOOMZ) with 4 consumer-grade NVIDIA RTX 3090 cards.\nFine-tuning: We train 3 runs\u2014reporting their average\u2014and select the optimal checkpoint based only on the English validation data\nfor true zero-shot cross-lingual transfer Schmidt et al. (2022  ###reference_b51###).\nFull hyperparameters are listed in Appendix A  ###reference_###."
        },
        {
            "section_id": "4.3",
            "parent_section_id": "4",
            "section_name": "Results",
            "text": "We compare with various multilingual baselines: PaLI Chen et al. (2022  ###reference_b7###), PaLI-X Chen et al. (2023  ###reference_b6###), Thapliyal et al. (2022  ###reference_b58###), LMCap Ramos et al. (2023  ###reference_b47###), UC2 Zhou et al. (2021  ###reference_b77###), Li et al. (2023c  ###reference_b33###), CCLM Zeng et al. (2023  ###reference_b73###), Ernie-UniX2 Shan et al. (2022  ###reference_b54###), Changpinyo et al. (2022  ###reference_b5###); and also evaluate two strong English Vision-LLMs (InstructBLIP Dai et al. (2023  ###reference_b13###) and Llava 1.5 Liu et al. (2023a  ###reference_b36###) (LLM is Vicuna 1.5 Touvron et al. (2023  ###reference_b59###); Chiang et al. (2023  ###reference_b9###))).\nTable 1  ###reference_### summarizes our image captioning results.\nOn XM3600 (Table 1(a)  ###reference_st1###), mBLIP mT0 outperforms the (training-free) captioning pipeline LMCap (Ramos et al., 2023  ###reference_b47###) as well as PaLI-X (in zero-shot inference): these results are very encouraging, considering that PaLI-X trains orders of magnitude more parameters (55B vs.\u2009124M for mBLIP), on billions of multilingual vision-and-language examples. mBLIP, however, substantially trails the performance of the PaLI models fine-tuned on MSCOCO with full translations to all 35 languages (yielding  more training examples than we do from our entire re-alignment task mix).\nWhile mBLIP is also trained on MSCOCO with translated captions, PaLI models consume orders of magnitude more data in most languages, especially the low-resource ones.\nWith proportionally less mBLIP training for lower-resource languages (according to the language-specific corpus portions in mC4), this yields especially large gains for PaLI models for low-resource languages; mBLIP is more competitive for high-resource languages like Spanish or German.\nThe English models show strong English results (as expected) but fail for other languages as they either do not generate captions in the target language or, for high-resource languages like German where captioning works, still underperform mBLIP.\nWe additionally evaluate on xFlickrCo (Table 1(b)  ###reference_st2###).\nWhile we are the first to use it for multilingual captioning (in Bugliarello et al. (2022  ###reference_b4###), it is used for image-text retrieval), on the English Flickr30k captions, mBLIP achieves performance that is comparable to that of the English LLMs while outclassing them for other languages.\nFinally, between the two mBLIP models, the mT0 variant beats the BLOOMZ variant. We believe this is due to the fact that mT5 (the base LLM from which mT0 was derived) was trained on almost 3 times more text (1 trillion tokens vs. 366 billion) and in nearly twice as many languages as BLOOM (the LLM of BLOOMZ). On a handful of languages like Indonesian or Hindi, however, BLOOMZ outperforms mT0, suggesting that the choice of the mBLIP variant is language-specific.\nTable 2  ###reference_### summarizes the results on VQA and image understanding tasks.\nOn xGQA, mBLIP (zero-shot) outperforms the UC2 model that has been fine-tuned on the GQA data (Zhou et al., 2021  ###reference_b77###; Bugliarello et al., 2022  ###reference_b4###) for all target languages.\nWhen fine-tuned, our mBLIP variants are only outperformed by CCLM (large) (Zeng et al., 2023  ###reference_b73###); CCLM (large) trains nearly nine-times more parameters and leverages more multilingual pretraining data888CCLM is also initialized with the English X2-VLM (Zeng et al., 2022a  ###reference_b72###) which is trained on 1B images; the BLIP-2 weights, from which we start the mBLIP training, in contrast, were trained using only 129M images.. Crucially, however, CCLM resorts to constrained generation w.r.t. the available answers, which is an easier yet computationally much more demanding evaluation protocol than our open generation.\nmBLIP exhibits relatively poor zero-shot XVNLI performance, as it fails to predict the neutral class. After fine-tuning for XVNLI, however, mBLIP mT0 yields multilingual performance (over 4 languages) comparable to that of CCLM (large).\nThe MaRVL zero-shot performance of mBLIP variants is surprisingly good, considering that they were never trained for any task involving multiple images as input; Zero-shot performance of mBLIP mT0 on MaRVL is comparable to that of multiple fine-tuned baselines. When also fine-tuned, mBLIP achieves state-of-the-art MaRVL results, on par with CCLM (large).\nOn MAXM, mBLIP mT0 (zero-shot) performs comparably to the 1.5B parameter baseline model of Changpinyo et al. (2022  ###reference_b5###) but falls short of the performance of the huge PaLI-17B model.\nmBLIP BLOOMZ exhibits strong English performance, but surprisingly poor results for other languages. We should emphasize here that training on the translated VQAv2 answers is crucial: without it, the LLM consistently generate answers in English. Even though only 25% of examples in VQAv2 have non-English answers, this is already sufficient to eliminate language hallucination, where the model only answers in English regardless of the instruction language999Training with only English VQAv2 answers during re-alignment results in an mBLIP mT0 instances that achieves only 15.5% accuracy for 6-avg, due to the LLM predominantly generating English answers..\nThe English Vision-LLMs, like in captioning, show strong results for English but fall behind in other languages. This is particular evident in MAXM, which has non-English answers (unlike xGQA and XVNLI) that the models fail to consistently generate.\nFor high-resource languages like German, mBLIP still outperforms them, highlighting its strong multilingual capabilities.\n###figure_2### ###figure_3### ###figure_4### Looking at results for individual languages on the three IGLUE tasks in Figure 2  ###reference_###, we see that mBLIP with mT0 greatly improves cross-lingual transfer over prior work, especially for lower-resource languages:\nwhile CCLM and Ernie-UniX2 exhibit a gap of 20-25% on xGQA between the best and worst language (German and Bengali), the same gap is only 5% for our fine-tuned mBLIP. Similarly, on MaRVL, CCLM has a gap of 11% between Indonesian and Tamil, while the largest gap for mBLIP amounts to 2%. The same holds for XVNLI, but to a lesser degree: the largest gap between languages for mBLIP (mT0) is 4%, compared to 8% for CCLM/Ernie-UniX2.\nThe BLOOMZ-based variant, however, exhibits much weaker transfer ability and has in fact larger gaps than prior work; this highlights the importance of deriving mBLIP from a strong multilingual LLM."
        },
        {
            "section_id": "5",
            "parent_section_id": null,
            "section_name": "Ablation",
            "text": "We ablate the various components and design decisions for mBLIP, namely: 1) using our instruction mix compared to the \u2018classic\u2019 setting used for BLIP-2 with only image-caption data (using the 2M Web CapFilt examples as training data) and compared to the instruction mix translated following the mT5 language distribution, 2) using LoRA on (all) LLM matrices to better align the LLM to the visual input, and 3) using the warm-start where the projection between Q-Former and LLM is trained briefly in a preliminary stage before the full re-alignment training. We use the zero-shot results on xGQA, XVNLI, and XM3600 for evaluation.\nResults are shown in Table 3  ###reference_###. In \u00a7C.1  ###reference_###, we provide an additional ablation that investigates the effect of adding the matching tasks to re-alignment mix, demonstrating their effectiveness in reducing hallucinations.\nIn \u00a7C.2  ###reference_###, we consider the effect of our design choices on fine-tuned models (on xGQA)."
        },
        {
            "section_id": "6",
            "parent_section_id": null,
            "section_name": "Conclusion",
            "text": "In this work, we presented mBLIP, the first modular and massively multilingual vision-language model based on multilingual LLMs. Using a small task mix from quality English datasets, made massively multilingual by means of MT, we re-align an English BLIP-2 model to an instruction-tuned multilingual LLM. Our approach is highly efficient in compute and data requirements and \u2013 using recent engineering advances such as 8-bit quantization \u2013 can be trained in a few days on consumer-grade hardware (e.g., NVIDIA RTX 3090 cards). We extensively evaluate mBLIP on multilingual vision-language tasks covering image captioning, visual QA, and image understanding to confirm the efficacy of our approach. Results render mBLIP comparable or better than state-of-the-art multilingual vision-language models and strong English Vision-LLMs, despite the fact that we train only a fraction of their number of parameters and on far less data."
        }
    ],
    "appendix": [
        {
            "section_id": "Appendix 1",
            "parent_section_id": null,
            "section_name": "Appendix A Training and Evaluation Details",
            "text": "Training: We use AdamW (Loshchilov and Hutter, 2019  ###reference_b38###) with weight decay 0.1, learning rate 2e-4 for LoRA and 1e-5 for other parameters; 1000 warm-up steps before a cosine decay; batch size 128 (accomplished via gradient accumulation and checkpointing); we limit the max. target sequence length to 128. For LoRA, which we apply to all LLM matrices and not just the query and value matrices of self-attention heads, we set  and use dropout with the  rate.\nWarmup:\nWe use 1M captions to train for 8k steps with a learning rate of 5e-3 (and otherwise the same hyperparameters).\nFine-tuning:\nWe train 3 runs (seeds)\u2014reporting their average\u2014for 5/10/20 epochs and batch size 256/128/128 for xGQA/XVNLI/MaRVL, respectively. Other hyperparameters are identical as in re-alignment training.\nWe merge the LoRA weights obtained in instruction-based re-alignment training into the LLM before we execute LoRA fine-tuning for downstream tasks.\nImplementation: We use the HuggingFace Transformers (Wolf et al., 2020  ###reference_b66###) and PEFT101010https://github.com/huggingface/peft libraries for model implementation and LoRA, respectively.\n###figure_5###"
        },
        {
            "section_id": "Appendix 2",
            "parent_section_id": null,
            "section_name": "Appendix B Qualitative Analysis",
            "text": "In addition to the quantitative evaluation on multilingual datasets of previous sections, we perform a qualitative analysis to better understand the model\u2019s visual and multilingual capabilities.\nAs shown in Figure 3  ###reference_###, our model can understand instructions in a wide range of languages and describe diverse images, perform simple reasoning, and correctly ground images to world knowledge in those languages. We also see some limitations.\nThe capabilities decrease notably for lower-resource languages. The Urdu example is only a short sentence despite asking for a detailed description.\nSimilarly, the Azerbaijani caption is completely incorrect (and non-sensical), while the model produces a meaningful caption for that same image in many other languages.\nThe Romanian example shows the limitations of the model\u2019s world knowledge as the famous portrait of Vlad III is not recognized (neither when asked in Romanian nor in English with various prompts). Finally, the Croatian example shows the difficulty with controlling the output language that we also saw in the quantitative evaluation: despite being asked in Croatian, the model answers in (related but still distinct) Slovenian."
        },
        {
            "section_id": "Appendix 3",
            "parent_section_id": null,
            "section_name": "Appendix C Further Ablation Results",
            "text": "We introduce the matching tasks with the aim of reducing object hallucinations.\nWe evaluate the effectiveness of the measure using two hallucination metrics for English:\nPOPE (Li et al., 2023b  ###reference_b32###) uses interrogative questions (\u201cIs there X in the image?\u201d) with random, popular, and adversarial negative objects (using MSCOCO images and object annotations), reporting accuracy and the portion of \u2018yes\u2019 answers due to a yes-bias in most models.\nCHAIR (Rohrbach et al., 2018  ###reference_b48###) generates captions from MSCOCO images (we use 1k images from the validation split) and then counts hallucinated objects using MSCOCO object annotations. They report the ratio of hallucinated object instances Ci, that is of all occurring objects, how many are hallucinated, and the ratio of sentences with hallucinations Cs.\nWe generate both short (Prompt: Caption in English:) and long captions (Prompt: Describe the image in English with as much detail as possible.).\nWe train two models for 30k steps with and without the matching tasks and report results in Table 4  ###reference_###.\nThe matching tasks greatly improve results for POPE as they reduce the yes-bias but CHAIR metrics decrease only slightly.\nThis seems to indicate that while matching tasks help for the interrogative POPE questions, they do not noticeably decrease hallucinations when generating captions.\nLooking at supervised xGQA fine-tuning, we observe that all variants exhibit similar performance, regardless of the instruction-tuning (i.e., re-alignment) design. The variants re-aligned only via captioning (first two rows of Table 3  ###reference_###) yield even slightly better results than the variants for which VQA was included in the re-alignment training. Contradicting the findings of Dai et al. (2023  ###reference_b13###),\nour results suggest that more \u2018complex\u2019 instruction-based re-alignment involving a multitude of tasks brings limited gains (if any) for downstream task with large fine-tuning data."
        },
        {
            "section_id": "Appendix 4",
            "parent_section_id": null,
            "section_name": "Appendix D Training and Evaluation Data and Template Details",
            "text": "We present our instruction mix in more detail with Table 6  ###reference_### listing the datasets with additional information, and Table 7  ###reference_### listing the templates used to generate the examples.\nWe present the templates used for the different evaluation datasets in Table 8  ###reference_###.\nTemplates for XVNLI and MaRVL are selected using English validation zero-shot performance.\nXVNLI templates are based on Muennighoff et al. (2022  ###reference_b40###).\nWe use the same templates for training and inference."
        },
        {
            "section_id": "Appendix 5",
            "parent_section_id": null,
            "section_name": "Appendix E Image Attribution",
            "text": "Image attribution for Figure 3  ###reference_### in order of appearance from top-left to bottom-right:\nSagrada Familia:\nhttps://de.wikipedia.org/wiki/Datei:Sagrada_Familia_8-12-21_(1).jpg  ###reference_a_Familia_8-12-21_(1).jpg###.\nCanaan, CC BY-SA 4.0 https://creativecommons.org/licenses/by-sa/4.0  ###reference_/4.0###, via Wikimedia Commons\nGiza:\nhttps://commons.wikimedia.org/wiki/File:All_Gizah_Pyramids.jpg  ###reference_l_Gizah_Pyramids.jpg###.\nRicardo Liberato, CC BY-SA 2.0 https://creativecommons.org/licenses/by-sa/2.0  ###reference_/2.0###, via Wikimedia Commons\nOktoberfest Kutsche:\nhttps://de.wikipedia.org/wiki/Datei:Oktoberfest-Kutscher.jpg  ###reference_rfest-Kutscher.jpg###.\nHullbr3ach, CC BY-SA 2.5 https://creativecommons.org/licenses/by-sa/2.5  ###reference_/2.5###, via Wikimedia Commons\nGate of All Nations, Persepolis:\nhttps://commons.wikimedia.org/wiki/File:Gate_of_All_Nations,_Persepolis.jpg  ###reference_te_of_All_Nations,_Persepolis.jpg###.\nAlborzagros, CC BY-SA 3.0 https://creativecommons.org/licenses/by-sa/3.0  ###reference_/3.0###, via Wikimedia Commons\nLake saif ul malook:\nhttps://en.wikipedia.org/wiki/File:Lake-saif-ul-malook_Pakistan.jpg  ###reference_if-ul-malook_Pakistan.jpg###.\nAyesha.great, CC BY-SA 4.0 https://creativecommons.org/licenses/by-sa/4.0  ###reference_/4.0###, via Wikimedia Commons\nVlad III:\nhttps://en.wikipedia.org/wiki/File:Vlad_Tepes_002.jpg  ###reference_pes_002.jpg###.\nPortrait of Vlad III the Impaler\nSatellite:\nhttps://en.wikipedia.org/wiki/File:Jaz_Murian_satellite.jpg  ###reference_ian_satellite.jpg###.\nNASA, Public domain, via Wikimedia Commons\nKrk waterfalls:\nhttps://commons.wikimedia.org/wiki/File:Krk_waterfalls.jpg  ###reference_k_waterfalls.jpg###.\nVersion13 at English Wikipedia, Public domain, via Wikimedia Commons"
        },
        {
            "section_id": "Appendix 6",
            "parent_section_id": null,
            "section_name": "Appendix F Full Results",
            "text": ""
        }
    ],
    "tables": {
        "1": {
            "table_html": "<figure class=\"ltx_table\" id=\"S4.T1\">\n<div class=\"ltx_flex_figure ltx_flex_table\">\n<div class=\"ltx_flex_cell ltx_flex_size_1\">\n<figure class=\"ltx_table ltx_figure_panel ltx_align_center\" id=\"S4.T1.st1\">\n<div class=\"ltx_inline-block ltx_transformed_outer\" id=\"S4.T1.st1.4\" style=\"width:433.6pt;height:190.8pt;vertical-align:-0.9pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(-29.8pt,13.1pt) scale(0.879155868403134,0.879155868403134) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"S4.T1.st1.4.4\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.5.1\">\n<th class=\"ltx_td ltx_th ltx_th_row ltx_border_tt\" id=\"S4.T1.st1.4.4.5.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_border_tt\" id=\"S4.T1.st1.4.4.5.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_border_tt\" id=\"S4.T1.st1.4.4.5.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" colspan=\"2\" id=\"S4.T1.st1.4.4.5.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.5.1.4.1\">XM3600</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.6.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_column ltx_th_row\" id=\"S4.T1.st1.4.4.6.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.1.1\">Model</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st1.4.4.6.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.2.1\">Train P.</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st1.4.4.6.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.3.1\">Total P.</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st1.4.4.6.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.4.1\">en</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st1.4.4.6.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.5.1\">35-avg</span></th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.1.1.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st1.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<cite class=\"ltx_cite ltx_citemacro_citet\">Thapliyal et\u00a0al. (<a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib58\" title=\"\">2022</a>)</cite> \n</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st1.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.8B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st1.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.8B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st1.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">57.60</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st1.1.1.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">28.90</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.2.2.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.2.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-3B \n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.2.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.2.2.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.2.2.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">92.80</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.2.2.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">47.00</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.3.3.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.3.3.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-17B \n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.3.3.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.3.3.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.3.3.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.3.3.3.4.1\">98.10</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.3.3.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.3.3.3.5.1\">53.60</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.4\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.4.4.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-X \n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">94.20</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">53.10</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.7.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st1.4.4.7.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-X 0-shot</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.7.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.7.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.7.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">48.80</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.7.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">22.70</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.8.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.4.4.8.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">LMCap <cite class=\"ltx_cite ltx_citemacro_citep\">(Ramos et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib47\" title=\"\">2023</a>)</cite>\n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.8.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.8.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.8.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">45.20</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.8.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17.60</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.9.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st1.4.4.9.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">InstructBLIP Flan-T5-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.9.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">107M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.9.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.1B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.9.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">85.22</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.9.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.10</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.10.4\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.4.4.10.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Llava 1.5 7B</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.10.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.10.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7.3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.10.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.87</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.10.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">9.78</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.11.5\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st1.4.4.11.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP mT0-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.11.5.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.11.5.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.9B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.11.5.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">80.17</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.11.5.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">26.77</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.12.6\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP BLOOMZ-7B</th>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.3B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">76.40</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">21.87</td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption\"><span class=\"ltx_tag ltx_tag_table\">(a) </span>\nmBLIP outperforms\nall models except those\nfine-tuned on MSCOCO translated to all 36 languages (). Different tokenizers for <span class=\"ltx_text ltx_font_italic\" id=\"S4.T1.st1.8.1\">zh, ja, th</span> make results not perfectly comparable.</figcaption>\n</figure>\n</div>\n<div class=\"ltx_flex_break\"></div>\n<div class=\"ltx_flex_cell ltx_flex_size_1\">\n<figure class=\"ltx_table ltx_figure_panel ltx_align_center\" id=\"S4.T1.st2\">\n<div class=\"ltx_inline-block ltx_transformed_outer\" id=\"S4.T1.st2.1\" style=\"width:433.6pt;height:157.1pt;vertical-align:-0.0pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(67.8pt,-24.6pt) scale(1.45496268095054,1.45496268095054) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"S4.T1.st2.1.1\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.1.1\">\n<th class=\"ltx_td ltx_th ltx_th_row ltx_border_tt\" id=\"S4.T1.st2.1.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_border_tt\" id=\"S4.T1.st2.1.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_border_tt\" id=\"S4.T1.st2.1.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" colspan=\"2\" id=\"S4.T1.st2.1.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.1.1.4.1\">xFlickrCo</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.2.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_column ltx_th_row\" id=\"S4.T1.st2.1.1.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.1.1\">Model</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st2.1.1.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.2.1\">Train P.</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st2.1.1.2.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.3.1\">Total P.</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st2.1.1.2.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.4.1\">en</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st2.1.1.2.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.5.1\">7-avg</span></th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.3.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">InstructBLIP Flan-T5-XL</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">107M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.1B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.3.1.4.1\">84.71</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.46</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.4.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st2.1.1.4.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Llava 1.5 7B</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.4.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.4.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7.3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.4.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">64.47</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.4.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">22.23</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.5.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st2.1.1.5.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP mT0-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.5.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.5.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.9B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.5.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">77.00</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.5.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.5.3.5.1\">44.39</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.6.4\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP BLOOMZ-7B</th>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.3B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">76.75</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">42.11</td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption\"><span class=\"ltx_tag ltx_tag_table\">(b) </span>No multilingual baseline on xFlickrCo exists at the time of writing but mBLIP is competitive with English models.</figcaption>\n</figure>\n</div>\n</div>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 1: </span>\nCaptioning results (CIDEr) on XM3600 and xFlickrCo for English and other languages.\n</figcaption>\n</figure>",
            "capture": "(a) \nmBLIP outperforms\nall models except those\nfine-tuned on MSCOCO translated to all 36 languages (). Different tokenizers for zh, ja, th make results not perfectly comparable."
        },
        "2": {
            "table_html": "<figure class=\"ltx_table ltx_figure_panel ltx_align_center\" id=\"S4.T1.st1\">\n<div class=\"ltx_inline-block ltx_transformed_outer\" id=\"S4.T1.st1.4\" style=\"width:433.6pt;height:190.8pt;vertical-align:-0.9pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(-29.8pt,13.1pt) scale(0.879155868403134,0.879155868403134) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"S4.T1.st1.4.4\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.5.1\">\n<th class=\"ltx_td ltx_th ltx_th_row ltx_border_tt\" id=\"S4.T1.st1.4.4.5.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_border_tt\" id=\"S4.T1.st1.4.4.5.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_border_tt\" id=\"S4.T1.st1.4.4.5.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" colspan=\"2\" id=\"S4.T1.st1.4.4.5.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.5.1.4.1\">XM3600</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.6.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_column ltx_th_row\" id=\"S4.T1.st1.4.4.6.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.1.1\">Model</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st1.4.4.6.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.2.1\">Train P.</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st1.4.4.6.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.3.1\">Total P.</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st1.4.4.6.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.4.1\">en</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st1.4.4.6.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.4.4.6.2.5.1\">35-avg</span></th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.1.1.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st1.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<cite class=\"ltx_cite ltx_citemacro_citet\">Thapliyal et\u00a0al. (<a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib58\" title=\"\">2022</a>)</cite> \n</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st1.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.8B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st1.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.8B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st1.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">57.60</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st1.1.1.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">28.90</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.2.2.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.2.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-3B \n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.2.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.2.2.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.2.2.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">92.80</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.2.2.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">47.00</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.3.3.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.3.3.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-17B \n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.3.3.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.3.3.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.3.3.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.3.3.3.4.1\">98.10</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.3.3.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st1.3.3.3.5.1\">53.60</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.4\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.4.4.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-X \n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">94.20</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">53.10</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.7.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st1.4.4.7.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-X 0-shot</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.7.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.7.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.7.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">48.80</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.7.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">22.70</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.8.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.4.4.8.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">LMCap <cite class=\"ltx_cite ltx_citemacro_citep\">(Ramos et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib47\" title=\"\">2023</a>)</cite>\n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.8.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.8.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.8.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">45.20</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.8.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17.60</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.9.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st1.4.4.9.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">InstructBLIP Flan-T5-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.9.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">107M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.9.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.1B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.9.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">85.22</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.9.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.10</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.10.4\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st1.4.4.10.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Llava 1.5 7B</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.10.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.10.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7.3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.10.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.87</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.10.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">9.78</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.11.5\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st1.4.4.11.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP mT0-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.11.5.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.11.5.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.9B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.11.5.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">80.17</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st1.4.4.11.5.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">26.77</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st1.4.4.12.6\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP BLOOMZ-7B</th>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.3B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">76.40</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st1.4.4.12.6.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">21.87</td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption\"><span class=\"ltx_tag ltx_tag_table\">(a) </span>\nmBLIP outperforms\nall models except those\nfine-tuned on MSCOCO translated to all 36 languages (). Different tokenizers for <span class=\"ltx_text ltx_font_italic\" id=\"S4.T1.st1.8.1\">zh, ja, th</span> make results not perfectly comparable.</figcaption>\n</figure>",
            "capture": "(a) \nmBLIP outperforms\nall models except those\nfine-tuned on MSCOCO translated to all 36 languages (). Different tokenizers for zh, ja, th make results not perfectly comparable."
        },
        "3": {
            "table_html": "<figure class=\"ltx_table ltx_figure_panel ltx_align_center\" id=\"S4.T1.st2\">\n<div class=\"ltx_inline-block ltx_transformed_outer\" id=\"S4.T1.st2.1\" style=\"width:433.6pt;height:157.1pt;vertical-align:-0.0pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(67.8pt,-24.6pt) scale(1.45496268095054,1.45496268095054) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"S4.T1.st2.1.1\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.1.1\">\n<th class=\"ltx_td ltx_th ltx_th_row ltx_border_tt\" id=\"S4.T1.st2.1.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_border_tt\" id=\"S4.T1.st2.1.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_border_tt\" id=\"S4.T1.st2.1.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" colspan=\"2\" id=\"S4.T1.st2.1.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.1.1.4.1\">xFlickrCo</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.2.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_column ltx_th_row\" id=\"S4.T1.st2.1.1.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.1.1\">Model</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st2.1.1.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.2.1\">Train P.</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st2.1.1.2.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.3.1\">Total P.</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st2.1.1.2.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.4.1\">en</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S4.T1.st2.1.1.2.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.2.2.5.1\">7-avg</span></th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.3.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">InstructBLIP Flan-T5-XL</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">107M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.1B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.3.1.4.1\">84.71</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T1.st2.1.1.3.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.46</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.4.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T1.st2.1.1.4.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Llava 1.5 7B</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.4.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.4.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7.3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.4.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">64.47</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.4.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">22.23</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.5.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T1.st2.1.1.5.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP mT0-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.5.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.5.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.9B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.5.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">77.00</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T1.st2.1.1.5.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T1.st2.1.1.5.3.5.1\">44.39</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T1.st2.1.1.6.4\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP BLOOMZ-7B</th>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.3B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">76.75</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T1.st2.1.1.6.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">42.11</td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption\"><span class=\"ltx_tag ltx_tag_table\">(b) </span>No multilingual baseline on xFlickrCo exists at the time of writing but mBLIP is competitive with English models.</figcaption>\n</figure>",
            "capture": "(b) No multilingual baseline on xFlickrCo exists at the time of writing but mBLIP is competitive with English models."
        },
        "4": {
            "table_html": "<figure class=\"ltx_table\" id=\"S4.T2\">\n<div class=\"ltx_inline-block ltx_align_center ltx_transformed_outer\" id=\"S4.T2.3\" style=\"width:433.6pt;height:213.9pt;vertical-align:-0.7pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(-112.7pt,55.4pt) scale(0.658023468638938,0.658023468638938) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"S4.T2.3.3\">\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.4.1\">\n<th class=\"ltx_td ltx_th ltx_th_row ltx_border_tt\" id=\"S4.T2.3.3.4.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_border_tt\" id=\"S4.T2.3.3.4.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_tt\" id=\"S4.T2.3.3.4.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_center ltx_border_tt\" colspan=\"2\" id=\"S4.T2.3.3.4.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.4.1.4.1\">XVNLI</span></td>\n<td class=\"ltx_td ltx_align_center ltx_border_tt\" colspan=\"2\" id=\"S4.T2.3.3.4.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.4.1.5.1\">MaRVL</span></td>\n<td class=\"ltx_td ltx_align_center ltx_border_tt\" colspan=\"2\" id=\"S4.T2.3.3.4.1.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.4.1.6.1\">xGQA</span></td>\n<td class=\"ltx_td ltx_align_center ltx_border_tt\" colspan=\"2\" id=\"S4.T2.3.3.4.1.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.4.1.7.1\">MaXM</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.5.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.3.3.5.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.1.1\">Model</span></th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.2.1\">Train P.</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.3.1\">Total P.</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.4.1\">en</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.5.1\">4-avg</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.6.1\">en</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.7.1\">5-avg</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.8.1\">en</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.9.1\">7-avg</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.10.1\">en</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.5.2.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.5.2.11.1\">6-avg</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.6.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" colspan=\"3\" id=\"S4.T2.3.3.6.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.6.3.1.1\">Fine-tuned on train split</span></th>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.6.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.6.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.6.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.6.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.6.3.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.6.3.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.6.3.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.6.3.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.7.4\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T2.3.3.7.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">UC2 <cite class=\"ltx_cite ltx_citemacro_citep\">(Bugliarello et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib4\" title=\"\">2022</a>)</cite>\n</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">270M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">270M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">76.38</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">62.05</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">70.56</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">57.28</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.19</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">29.35</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.3.3.7.4.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.8.5\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.3.3.8.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><cite class=\"ltx_cite ltx_citemacro_citet\">Li et\u00a0al. (<a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib33\" title=\"\">2023c</a>)</cite></th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">330M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">330M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">69.50</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">62.10</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">42.10</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.8.5.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.1.1.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">CCLM (4M) \n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">520M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">520M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">73.32</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">83.22</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">67.17</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">46.24</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.1.1.1.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.9.6\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.3.3.9.6.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">CCLM base</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">420M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">420M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">74.78</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">68.49</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">48.12</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.9.6.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.10.7\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.3.3.10.7.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">CCLM large</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">970M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">970M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.10.7.5.1\">78.95</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">74.83</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.10.7.9.1\">56.25</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.10.7.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.11.8\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.3.3.11.8.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Ernie-UniX2</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">910M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">910M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.11.8.4.1\">87.73</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">77.42</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">56.68</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">45.25</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.11.8.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.12.9\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T2.3.3.12.9.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP mT0-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.9B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">82.41</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">76.41</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">85.20</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.12.9.7.1\">75.13</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">56.54</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">47.71</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.12.9.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.13.10\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.3.3.13.10.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP BLOOMZ-7B</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">75.45</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">66.96</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.13.10.6.1\">86.69</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">73.94</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.13.10.8.1\">57.89</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">44.91</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.13.10.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.14.11\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" colspan=\"3\" id=\"S4.T2.3.3.14.11.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.14.11.1.1\">Zero-shot</span></th>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.14.11.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.14.11.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.14.11.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.14.11.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.14.11.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.14.11.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.14.11.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_border_t\" id=\"S4.T2.3.3.14.11.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.2.2.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T2.2.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<cite class=\"ltx_cite ltx_citemacro_citet\">Changpinyo et\u00a0al. (<a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib5\" title=\"\">2022</a>)</cite> \n</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.5B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.5B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">41.50</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">39.44</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">36.60</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S4.T2.2.2.2.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">42.42</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.3.3.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">PaLI-17B \n</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">54.20</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.3.9.1\">50.77</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.3.10.1\">56.40</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.3.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.3.11.1\">57.27</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.15.12\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T2.3.3.15.12.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">InstructBLIP Flan-T5-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">107M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.1B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.15.12.4.1\">62.09</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">48.65</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">48.23</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">18.63</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.03</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.15.12.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.4</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.16.13\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"S4.T2.3.3.16.13.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Llava 1.5 7B *</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7.3B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">56.43</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">49.33</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2014</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">*<span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.16.13.8.1\">57.37</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">*27.53</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">52.01</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.16.13.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">16.22</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.17.14\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"S4.T2.3.3.17.14.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP mT0-XL</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.9B</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">60.61</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.17.14.5.1\">57.65</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.17.14.6.1\">67.26</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.3.3.17.14.7.1\">66.66</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">42.55</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">39.20</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">47.99</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S4.T2.3.3.17.14.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">41.04</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S4.T2.3.3.18.15\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_bb\" id=\"S4.T2.3.3.18.15.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP BLOOMZ-7B</th>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">124M</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.3B</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">58.26</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.46</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">62.26</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">58.61</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">43.35</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">37.73</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.70</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S4.T2.3.3.18.15.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">27.91</td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 2: </span>\nVQA and image understanding results for English and averaged over all other languages: The metric is (exact match) accuracy with open generation for mBLIP &amp; PaLI and constrained generation to a set of labels for CCLM on xGQA.\n<span class=\"ltx_text ltx_font_bold\" id=\"S4.T2.9.1\">Bold</span> indicates the best score in each column.\n: From <cite class=\"ltx_cite ltx_citemacro_cite\">Zeng et\u00a0al. (<a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib74\" title=\"\">2022b</a>)</cite> v1\u2009(arXiv).\n: Fine-tuned on VQAv2 translated to all MaXM &amp; xGQA languages.\n*: GQA included in training data.\n</figcaption>\n</figure>",
            "capture": "Table 2: \nVQA and image understanding results for English and averaged over all other languages: The metric is (exact match) accuracy with open generation for mBLIP & PaLI and constrained generation to a set of labels for CCLM on xGQA.\nBold indicates the best score in each column.\n: From Zeng et\u00a0al. (2022b) v1\u2009(arXiv).\n: Fine-tuned on VQAv2 translated to all MaXM & xGQA languages.\n*: GQA included in training data.\n"
        },
        "5": {
            "table_html": "<figure class=\"ltx_table\" id=\"S5.T3\">\n<div class=\"ltx_inline-block ltx_align_center ltx_transformed_outer\" id=\"S5.T3.1\" style=\"width:433.6pt;height:214.1pt;vertical-align:-0.0pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(52.8pt,-26.1pt) scale(1.32167419605461,1.32167419605461) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"S5.T3.1.1\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.1.1\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_th_row ltx_border_tt\" id=\"S5.T3.1.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.1.1.1.1\">Task</span></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_th_row ltx_border_tt\" id=\"S5.T3.1.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.1.1.2.1\">LoRA</span></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_th_row ltx_border_tt\" id=\"S5.T3.1.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.1.1.3.1\">Warm-</span></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" colspan=\"2\" id=\"S5.T3.1.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.1.1.4.1\">xGQA</span></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" colspan=\"2\" id=\"S5.T3.1.1.1.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.1.1.5.1\">XVNLI</span></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" colspan=\"2\" id=\"S5.T3.1.1.1.1.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.1.1.6.1\">XM3600</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.2.2\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_th_row\" id=\"S5.T3.1.1.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.2.2.1.1\">Mix</span></th>\n<th class=\"ltx_td ltx_th ltx_th_column ltx_th_row\" id=\"S5.T3.1.1.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_th_row\" id=\"S5.T3.1.1.2.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.2.2.3.1\">start</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S5.T3.1.1.2.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.2.2.4.1\">en</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S5.T3.1.1.2.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.2.2.5.1\">avg</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S5.T3.1.1.2.2.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.2.2.6.1\">en</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S5.T3.1.1.2.2.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.2.2.7.1\">avg</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S5.T3.1.1.2.2.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.2.2.8.1\">en</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"S5.T3.1.1.2.2.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.2.2.9.1\">avg</span></th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.3.1\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row ltx_border_t\" id=\"S5.T3.1.1.3.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row ltx_border_t\" id=\"S5.T3.1.1.3.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row ltx_border_t\" id=\"S5.T3.1.1.3.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S5.T3.1.1.3.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">26.92</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S5.T3.1.1.3.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">9.43</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S5.T3.1.1.3.1.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">34.17</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S5.T3.1.1.3.1.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">35.26</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S5.T3.1.1.3.1.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.3.1.8.1\">86.78</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"S5.T3.1.1.3.1.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">22.01</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.4.2\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.4.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.4.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">all</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.4.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.4.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.51</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.4.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.00</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.4.2.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">33.04</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.4.2.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">25.72</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.4.2.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">85.53</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.4.2.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">24.69</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.5.3\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.5.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.5.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.5.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.5.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">37.33</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.5.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">33.77</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.5.3.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">52.02</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.5.3.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">54.26</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.5.3.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">84.14</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.5.3.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">21.35</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.6.4\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.6.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.6.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">q,v</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.6.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.6.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">39.83</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.6.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">36.50</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.6.4.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">57.91</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.6.4.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.22</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.6.4.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">81.45</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.6.4.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">23.46</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.7.5\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.7.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.7.5.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">all</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.7.5.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.7.5.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">40.89</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.7.5.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">37.88</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.7.5.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">57.74</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.7.5.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">54.50</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.7.5.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">80.68</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.7.5.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">24.38</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.8.6\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.8.6.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mT5</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.8.6.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">all</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row\" id=\"S5.T3.1.1.8.6.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.8.6.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">40.91</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.8.6.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">37.67</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.8.6.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">58.00</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.8.6.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">54.96</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.8.6.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">80.13</td>\n<td class=\"ltx_td ltx_align_right\" id=\"S5.T3.1.1.8.6.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.8.6.9.1\">25.85</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"S5.T3.1.1.9.7\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row ltx_border_bb\" id=\"S5.T3.1.1.9.7.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row ltx_border_bb\" id=\"S5.T3.1.1.9.7.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">all</th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_row ltx_border_bb\" id=\"S5.T3.1.1.9.7.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</th>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S5.T3.1.1.9.7.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.9.7.4.1\">41.98</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S5.T3.1.1.9.7.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.9.7.5.1\">38.46</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S5.T3.1.1.9.7.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.9.7.6.1\">58.87</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S5.T3.1.1.9.7.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"S5.T3.1.1.9.7.7.1\">56.28</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S5.T3.1.1.9.7.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">81.51</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"S5.T3.1.1.9.7.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">25.02</td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 3: </span>\nAblations for mBLIP (mT0) w.r.t.: (i) instruction mix (\u2713) vs. only captions (\u2717) (i.e., the 2M Web CapFilt examples) vs. instruction mix using the mT5 distribution (mT5), (ii) LoRA (no LoRA \u2717, standard LoRA on query&amp;value matrices, LoRA on all matrices), and (iii) using the warm-start where the projection between Q-Former and LLM is trained alone first. All model variants are trained (i.e., re-aligned) for 30k steps.\n</figcaption>\n</figure>",
            "capture": "Table 3: \nAblations for mBLIP (mT0) w.r.t.: (i) instruction mix (\u2713) vs. only captions (\u2717) (i.e., the 2M Web CapFilt examples) vs. instruction mix using the mT5 distribution (mT5), (ii) LoRA (no LoRA \u2717, standard LoRA on query&value matrices, LoRA on all matrices), and (iii) using the warm-start where the projection between Q-Former and LLM is trained alone first. All model variants are trained (i.e., re-aligned) for 30k steps.\n"
        },
        "6": {
            "table_html": "<figure class=\"ltx_table\" id=\"A3.T4\">\n<div class=\"ltx_inline-block ltx_align_center ltx_transformed_outer\" id=\"A3.T4.4\" style=\"width:433.6pt;height:92.8pt;vertical-align:-0.0pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(6.5pt,-1.4pt) scale(1.03092657842552,1.03092657842552) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"A3.T4.4.4\">\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A3.T4.4.4.5.1\">\n<th class=\"ltx_td ltx_th ltx_th_row ltx_border_tt\" id=\"A3.T4.4.4.5.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_center ltx_border_tt\" colspan=\"6\" id=\"A3.T4.4.4.5.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.5.1.2.1\">POPE</span></td>\n<td class=\"ltx_td ltx_align_center ltx_border_tt\" colspan=\"4\" id=\"A3.T4.4.4.5.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.5.1.3.1\">CHAIR</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A3.T4.4.4.6.2\">\n<th class=\"ltx_td ltx_th ltx_th_row\" id=\"A3.T4.4.4.6.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_center\" colspan=\"2\" id=\"A3.T4.4.4.6.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.6.2.2.1\">random</span></td>\n<td class=\"ltx_td ltx_align_center\" colspan=\"2\" id=\"A3.T4.4.4.6.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.6.2.3.1\">popular</span></td>\n<td class=\"ltx_td ltx_align_center\" colspan=\"2\" id=\"A3.T4.4.4.6.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.6.2.4.1\">adversarial</span></td>\n<td class=\"ltx_td ltx_align_center\" colspan=\"2\" id=\"A3.T4.4.4.6.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.6.2.5.1\">short</span></td>\n<td class=\"ltx_td ltx_align_center\" colspan=\"2\" id=\"A3.T4.4.4.6.2.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.6.2.6.1\">long</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A3.T4.4.4.4\">\n<th class=\"ltx_td ltx_th ltx_th_row\" id=\"A3.T4.4.4.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.4.4.4.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.4.6.1\">acc</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.4.4.4.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.4.7.1\">yes</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.4.4.4.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.4.8.1\">acc</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.4.4.4.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.4.9.1\">yes</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.4.4.4.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.4.10.1\">acc</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.4.4.4.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.4.11.1\">yes</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.1.1.1.1.1\">C<sub class=\"ltx_sub\" id=\"A3.T4.1.1.1.1.1.1\"><span class=\"ltx_text ltx_font_medium ltx_font_italic\" id=\"A3.T4.1.1.1.1.1.1.1\">i</span></sub></span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.2.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.2.2.2.2.1\">C<sub class=\"ltx_sub\" id=\"A3.T4.2.2.2.2.1.1\"><span class=\"ltx_text ltx_font_medium ltx_font_italic\" id=\"A3.T4.2.2.2.2.1.1.1\">s</span></sub></span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.3.3.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.3.3.3.3.1\">C<sub class=\"ltx_sub\" id=\"A3.T4.3.3.3.3.1.1\"><span class=\"ltx_text ltx_font_medium ltx_font_italic\" id=\"A3.T4.3.3.3.3.1.1.1\">i</span></sub></span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T4.4.4.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T4.4.4.4.4.1\">C<sub class=\"ltx_sub\" id=\"A3.T4.4.4.4.4.1.1\"><span class=\"ltx_text ltx_font_medium ltx_font_italic\" id=\"A3.T4.4.4.4.4.1.1.1\">s</span></sub></span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A3.T4.4.4.7.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"A3.T4.4.4.7.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">without matching</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">71.00</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">74%</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">70.40</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">75%</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">63.70</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">81%</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">3.10</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.50</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">14.90</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T4.4.4.7.3.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">54.70</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A3.T4.4.4.8.4\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_bb\" id=\"A3.T4.4.4.8.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">with matching</th>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">87.30</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">48%</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">83.30</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">52%</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">76.10</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">59%</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">2.40</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">3.50</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">14.10</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T4.4.4.8.4.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">50.50</td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 4: </span>\nEffect of decision tasks on object hallucination evaluated with POPE <cite class=\"ltx_cite ltx_citemacro_citep\">(Li et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib32\" title=\"\">2023b</a>)</cite> and CHAIR <cite class=\"ltx_cite ltx_citemacro_citep\">(Rohrbach et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib48\" title=\"\">2018</a>)</cite> metrics.\nPOPE results improve because the yes-bias is reduced but CHAIR metrics for both short and long captions barely decrease (lower is better).\n</figcaption>\n</figure>",
            "capture": "Table 4: \nEffect of decision tasks on object hallucination evaluated with POPE (Li et\u00a0al., 2023b) and CHAIR (Rohrbach et\u00a0al., 2018) metrics.\nPOPE results improve because the yes-bias is reduced but CHAIR metrics for both short and long captions barely decrease (lower is better).\n"
        },
        "7": {
            "table_html": "<figure class=\"ltx_table\" id=\"A3.T5\">\n<div class=\"ltx_inline-block ltx_align_center ltx_transformed_outer\" id=\"A3.T5.1\" style=\"width:433.6pt;height:247.8pt;vertical-align:-0.0pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(122.3pt,-69.9pt) scale(2.29461970248135,2.29461970248135) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"A3.T5.1.1\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"A3.T5.1.1.1.1\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" id=\"A3.T5.1.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.1.1.1.1\">Task</span></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" id=\"A3.T5.1.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.1.1.2.1\">LoRA</span></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" id=\"A3.T5.1.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.1.1.3.1\">Warm-</span></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column ltx_border_tt\" colspan=\"2\" id=\"A3.T5.1.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.1.1.4.1\">xGQA (finetune)</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"A3.T5.1.1.2.2\">\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column\" id=\"A3.T5.1.1.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.2.2.1.1\">Mix</span></th>\n<th class=\"ltx_td ltx_th ltx_th_column\" id=\"A3.T5.1.1.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<th class=\"ltx_td ltx_align_center ltx_th ltx_th_column\" id=\"A3.T5.1.1.2.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.2.2.3.1\">start</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"A3.T5.1.1.2.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.2.2.4.1\">en</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column\" id=\"A3.T5.1.1.2.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.2.2.5.1\">avg</span></th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A3.T5.1.1.3.1\">\n<td class=\"ltx_td ltx_align_center ltx_border_t\" id=\"A3.T5.1.1.3.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</td>\n<td class=\"ltx_td ltx_align_center ltx_border_t\" id=\"A3.T5.1.1.3.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</td>\n<td class=\"ltx_td ltx_align_center ltx_border_t\" id=\"A3.T5.1.1.3.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T5.1.1.3.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.3.1.4.1\">56.68</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A3.T5.1.1.3.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">46.50</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A3.T5.1.1.4.2\">\n<td class=\"ltx_td ltx_align_center\" id=\"A3.T5.1.1.4.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</td>\n<td class=\"ltx_td ltx_align_center\" id=\"A3.T5.1.1.4.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">all</td>\n<td class=\"ltx_td ltx_align_center\" id=\"A3.T5.1.1.4.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T5.1.1.4.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">56.55</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T5.1.1.4.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">44.78</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A3.T5.1.1.5.3\">\n<td class=\"ltx_td ltx_align_center\" id=\"A3.T5.1.1.5.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</td>\n<td class=\"ltx_td ltx_align_center\" id=\"A3.T5.1.1.5.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2717</td>\n<td class=\"ltx_td ltx_align_center\" id=\"A3.T5.1.1.5.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T5.1.1.5.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.72</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A3.T5.1.1.5.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">45.36</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A3.T5.1.1.6.4\">\n<td class=\"ltx_td ltx_align_center ltx_border_bb\" id=\"A3.T5.1.1.6.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</td>\n<td class=\"ltx_td ltx_align_center ltx_border_bb\" id=\"A3.T5.1.1.6.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">all</td>\n<td class=\"ltx_td ltx_align_center ltx_border_bb\" id=\"A3.T5.1.1.6.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\u2713</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T5.1.1.6.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">56.47</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A3.T5.1.1.6.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A3.T5.1.1.6.4.5.1\">46.84</span></td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 5: </span>\nAblations for mBLIP (mT0) w.r.t.: (i) instruction mix (\u2713) vs. only captions (\u2717) (i.e., the 2M Web CapFilt examples) vs. instruction mix using the mT5 distribution (mT5), (ii) LoRA (no LoRA \u2717, standard LoRA on query&amp;value matrices, LoRA on all matrices), and (iii) using the warm-start where the projection between Q-Former and LLM is trained alone first. All model variants are trained (i.e., re-aligned) for 30k steps.\n</figcaption>\n</figure>",
            "capture": "Table 5: \nAblations for mBLIP (mT0) w.r.t.: (i) instruction mix (\u2713) vs. only captions (\u2717) (i.e., the 2M Web CapFilt examples) vs. instruction mix using the mT5 distribution (mT5), (ii) LoRA (no LoRA \u2717, standard LoRA on query&value matrices, LoRA on all matrices), and (iii) using the warm-start where the projection between Q-Former and LLM is trained alone first. All model variants are trained (i.e., re-aligned) for 30k steps.\n"
        },
        "8": {
            "table_html": "<figure class=\"ltx_table\" id=\"A4.T6\">\n<div class=\"ltx_inline-block ltx_align_center ltx_transformed_outer\" id=\"A4.T6.8\" style=\"width:433.6pt;height:803.5pt;vertical-align:-0.9pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(-37.3pt,69.0pt) scale(0.853361152376225,0.853361152376225) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"A4.T6.8.8\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"A4.T6.8.8.9.1\">\n<th class=\"ltx_td ltx_align_justify ltx_align_top ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T6.8.8.9.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.9.1.1.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.9.1.1.1.1\" style=\"width:142.3pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T6.8.8.9.1.1.1.1.1\">Dataset</span></span>\n</span>\n</th>\n<th class=\"ltx_td ltx_align_justify ltx_align_top ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T6.8.8.9.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.9.1.2.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.9.1.2.1.1\" style=\"width:71.1pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T6.8.8.9.1.2.1.1.1\">Tasks</span></span>\n</span>\n</th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T6.8.8.9.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T6.8.8.9.1.3.1\">#Images</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T6.8.8.9.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T6.8.8.9.1.4.1\">#Examples</span></th>\n<th class=\"ltx_td ltx_align_justify ltx_align_top ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T6.8.8.9.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.9.1.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.9.1.5.1.1\" style=\"width:142.3pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T6.8.8.9.1.5.1.1.1\">Details</span></span>\n</span>\n</th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A4.T6.8.8.10.1\">\n<td class=\"ltx_td ltx_align_justify ltx_align_top ltx_border_t\" id=\"A4.T6.8.8.10.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.10.1.1.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.10.1.1.1.1\" style=\"width:142.3pt;\">Web CapFilt <cite class=\"ltx_cite ltx_citemacro_citep\">(Li et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib28\" title=\"\">2022</a>)</cite></span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top ltx_border_t\" id=\"A4.T6.8.8.10.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.10.1.2.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.10.1.2.1.1\" style=\"width:71.1pt;\">Image captioning</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A4.T6.8.8.10.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">2.27m</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A4.T6.8.8.10.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">2.27m</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top ltx_border_t\" id=\"A4.T6.8.8.10.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.10.1.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.10.1.5.1.1\" style=\"width:142.3pt;\">Subset of the CC3M+CC12M+SBU Web CapFilt dataset<span class=\"ltx_note ltx_role_footnote\" id=\"footnote11\"><sup class=\"ltx_note_mark\">11</sup><span class=\"ltx_note_outer\"><span class=\"ltx_note_content\"><sup class=\"ltx_note_mark\">11</sup><span class=\"ltx_tag ltx_tag_note\">11</span><a class=\"ltx_ref ltx_url ltx_font_typewriter\" href=\"https://storage.googleapis.com/sfr-vision-language-research/BLIP/datasets/ccs_synthetic_filtered_large.json\" title=\"\">https://storage.googleapis.com/sfr-vision-language-research/BLIP/datasets/ccs_synthetic_filtered_large.json</a></span></span></span>. Like <cite class=\"ltx_cite ltx_citemacro_citet\">Liu et\u00a0al. (<a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib37\" title=\"\">2023b</a>)</cite>, we use spaCy to extract noun phrases and then sample from every phrase with at least 10 occurrences at most 30 captions for a subset covering diverse concepts.</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.8.8.11.2\">\n<td class=\"ltx_td ltx_align_top\" id=\"A4.T6.8.8.11.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.11.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.11.2.2.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.11.2.2.1.1\" style=\"width:71.1pt;\">Caption Matching</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.8.8.11.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">600k</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.8.8.11.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">600k</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.11.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.11.2.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.11.2.5.1.1\" style=\"width:142.3pt;\">Subset of our image captioning data. We use the CLIP ViT-L/14 by <cite class=\"ltx_cite ltx_citemacro_citet\">Gadre et\u00a0al. (<a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib20\" title=\"\">2023</a>)</cite> to encode images and text to find similar examples for hard negatives. We match every image randomly with the correct caption (50% of the time) or with equal probability a random caption or the 3/10/30/100/300 most similar caption for a mix of very hard to random negatives.</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.1.1.1\">\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.1.1.1.2.1\">\n<span class=\"ltx_p\" id=\"A4.T6.1.1.1.2.1.1\" style=\"width:142.3pt;\">MSCOCO <cite class=\"ltx_cite ltx_citemacro_citep\">(Lin et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib34\" title=\"\">2014</a>)</cite></span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.1.1.1.3.1\">\n<span class=\"ltx_p\" id=\"A4.T6.1.1.1.3.1.1\" style=\"width:71.1pt;\">Image Captioning</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">83k\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">414k</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.1.1.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.1.1.1.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.1.1.1.5.1.1\" style=\"width:142.3pt;\">Karpathy training split of MSCOCO <cite class=\"ltx_cite ltx_citemacro_citep\">(Karpathy and Fei-Fei, <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib26\" title=\"\">2017</a>)</cite> with 5 captions per image.</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.4.4.4\">\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.4.4.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.4.4.4.4.1\">\n<span class=\"ltx_p\" id=\"A4.T6.4.4.4.4.1.1\" style=\"width:142.3pt;\">VQAv2 <cite class=\"ltx_cite ltx_citemacro_citep\">(Goyal et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib23\" title=\"\">2017</a>)</cite></span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.4.4.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.4.4.4.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.4.4.4.5.1.1\" style=\"width:71.1pt;\">VQA, VQG</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.2.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">83k\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.3.3.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">2443k</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.4.4.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.4.4.4.3.1\">\n<span class=\"ltx_p\" id=\"A4.T6.4.4.4.3.1.1\" style=\"width:142.3pt;\">Question-answer pairs with 5 questions per image. For VQA and VQG, each example is translated to a different language to increase language diversity.\nWe use Google Translate to translate the most common 1500 answers to the 95 languages. We then back-translate them to English and keep only the translations where the back-translation is the original answer; this is to ensure that the answer is (likely) translated correctly. We randomly use either the translated or English answer when generating examples. 83k of the 443k examples have non-English answers.</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.6.6.6\">\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.6.6.6.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.6.6.6.3.1\">\n<span class=\"ltx_p\" id=\"A4.T6.6.6.6.3.1.1\" style=\"width:142.3pt;\">A-OKVQA <cite class=\"ltx_cite ltx_citemacro_citep\">(Schwenk et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib53\" title=\"\">2022</a>)</cite></span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.6.6.6.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.6.6.6.4.1\">\n<span class=\"ltx_p\" id=\"A4.T6.6.6.6.4.1.1\" style=\"width:71.1pt;\">Rational generation, VQA with rational</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.5.5.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">11k\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.6.6.6.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">233k</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.6.6.6.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.6.6.6.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.6.6.6.5.1.1\" style=\"width:142.3pt;\">Knowledge-intense VQA questions with additional answer rationals. We generate examples for all three given rationales. We only use the subset of the training split overlapping with the MSCOCO training split. A-OKVQA examples are not translated to any language.</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.7.7.7\">\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.7.7.7.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.7.7.7.2.1\">\n<span class=\"ltx_p\" id=\"A4.T6.7.7.7.2.1.1\" style=\"width:142.3pt;\">LLaVA <cite class=\"ltx_cite ltx_citemacro_citep\">(Liu et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib37\" title=\"\">2023b</a>)</cite> detail</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.7.7.7.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.7.7.7.3.1\">\n<span class=\"ltx_p\" id=\"A4.T6.7.7.7.3.1.1\" style=\"width:71.1pt;\">Image captioning</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.7.7.7.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">23k\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.7.7.7.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">23k</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.7.7.7.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.7.7.7.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.7.7.7.5.1.1\" style=\"width:142.3pt;\">Subset of LLaVA instructions with detailed multi-sentence image captions.</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.8.8.8\">\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.8.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.8.2.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.8.2.1.1\" style=\"width:142.3pt;\">LLaVA <cite class=\"ltx_cite ltx_citemacro_citep\">(Liu et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib37\" title=\"\">2023b</a>)</cite> conversations</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.8.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.8.3.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.8.3.1.1\" style=\"width:71.1pt;\">VQA</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.8.8.8.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">56k\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.8.8.8.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">219k</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.8.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.8.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.8.5.1.1\" style=\"width:142.3pt;\">Subset of LLaVA instructions with multi-turn dialog; we split the dialogs into independent pairs and keep all pairs with an answer length of max. 3 sentences.</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.8.8.12.3\">\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.12.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.12.3.1.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.12.3.1.1.1\" style=\"width:142.3pt;\">ImageNet <cite class=\"ltx_cite ltx_citemacro_citep\">(Deng et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib14\" title=\"\">2009</a>)</cite> and Babel-ImageNet <cite class=\"ltx_cite ltx_citemacro_citep\">(Geigle et\u00a0al., <a class=\"ltx_ref\" href=\"https://arxiv.org/html/2307.06930v3#bib.bib22\" title=\"\">2023</a>)</cite></span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.12.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.12.3.2.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.12.3.2.1.1\" style=\"width:71.1pt;\">VQA</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.8.8.12.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">300k</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.8.8.12.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">300k</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.12.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.12.3.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.12.3.5.1.1\" style=\"width:142.3pt;\">Image classification framed as open-ended VQA tasks (i.e., no answer options are given). Babel-ImageNet provides partial translations of the ImageNet classes to the 95 languages. We select one image for every class+language combination (that is, we do not use the full training set).</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.8.8.13.4\">\n<td class=\"ltx_td ltx_align_top\" id=\"A4.T6.8.8.13.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.13.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.13.4.2.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.13.4.2.1.1\" style=\"width:71.1pt;\">Matching</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.8.8.13.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">300k</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A4.T6.8.8.13.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">300k</td>\n<td class=\"ltx_td ltx_align_justify ltx_align_top\" id=\"A4.T6.8.8.13.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.13.4.5.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.13.4.5.1.1\" style=\"width:142.3pt;\">The model has to decide if a given ImageNet class is correctly in the image. We use the correct label or a random label with equal probability. This uses the same images as the VQA examples but shuffles the image-language pairs.</span>\n</span>\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T6.8.8.14.5\">\n<td class=\"ltx_td ltx_align_justify ltx_align_top ltx_border_bb ltx_border_t\" id=\"A4.T6.8.8.14.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">\n<span class=\"ltx_inline-block ltx_align_top\" id=\"A4.T6.8.8.14.5.1.1\">\n<span class=\"ltx_p\" id=\"A4.T6.8.8.14.5.1.1.1\" style=\"width:142.3pt;\">Total</span>\n</span>\n</td>\n<td class=\"ltx_td ltx_align_top ltx_border_bb ltx_border_t\" id=\"A4.T6.8.8.14.5.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb ltx_border_t\" id=\"A4.T6.8.8.14.5.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">2.65m</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb ltx_border_t\" id=\"A4.T6.8.8.14.5.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">5.1m</td>\n<td class=\"ltx_td ltx_align_top ltx_border_bb ltx_border_t\" id=\"A4.T6.8.8.14.5.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 6: </span>\nDetailed information about the datasets used for training.\n: Dataset uses MSCOCO images.\n</figcaption>\n</figure>",
            "capture": "Table 6: \nDetailed information about the datasets used for training.\n: Dataset uses MSCOCO images.\n"
        },
        "9": {
            "table_html": "<figure class=\"ltx_table\" id=\"A4.T7\">\n<div class=\"ltx_inline-block ltx_align_center ltx_transformed_outer\" id=\"A4.T7.5\" style=\"width:433.6pt;height:845.8pt;vertical-align:-0.7pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(-108.2pt,210.9pt) scale(0.667021094063388,0.667021094063388) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"A4.T7.5.5\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.6.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T7.5.5.6.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T7.5.5.6.1.1.1\">Task</span></th>\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T7.5.5.6.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T7.5.5.6.1.2.1\">Templates</span></th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.7.1\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.7.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Image Captioning</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.7.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Caption the image in $LANGUAGE.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.8.2\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.8.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.8.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Short $LANGUAGE image caption:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.9.3\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.9.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.9.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Image caption (in $LANGUAGE):</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.10.4\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.10.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.10.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Briefly describe the image in $LANGUAGE.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.11.5\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.11.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.11.5.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Write a short $LANGUAGE image description.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.12.6\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.12.6.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.12.6.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Summarize the image in $LANGUAGE.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.1.1.1\">\n<td class=\"ltx_td\" id=\"A4.T7.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Caption the image.\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.2.2.2\">\n<td class=\"ltx_td\" id=\"A4.T7.2.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.2.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Short image caption:\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.3.3.3\">\n<td class=\"ltx_td\" id=\"A4.T7.3.3.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.3.3.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Briefly describe the image.\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.4.4.4\">\n<td class=\"ltx_td\" id=\"A4.T7.4.4.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.4.4.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Write a short image description.\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.5\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.5.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Summarize the image.\n</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.13.7\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.13.7.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Caption Matching</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.13.7.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Does \"$CAPTION\" accurately describe the image? | Yes, it does. | No, it does not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.14.8\">\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.14.8.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Question | Yes Answer | No Answer</td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.14.8.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Does the caption \"$CAPTION\" fit the picture? | Yes, it does. | No, it does not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.15.9\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.15.9.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.15.9.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Does \"$CAPTION\" correctly summarize the image? | Yes, it does. | No, it does not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.16.10\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.16.10.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.16.10.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Is \"$CAPTION\" a good image description? | Yes, it is. | No, it is not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.17.11\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.17.11.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.17.11.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Is \"$CAPTION\" a correct caption for the picture? | Yes, it is. | No, it is not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.18.12\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.18.12.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.18.12.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Is the caption \"$CAPTION\" a good match for the image? | Yes, it is. | No, it is not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.19.13\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.19.13.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.19.13.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Decide if the following caption accurately describes the image: $CAPTION. Answer: | Yes, it does. | No, it does not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.20.14\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.20.14.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.20.14.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Is this caption a good match for the picture? $CAPTION. Answer: | Yes, it is. | No, it is not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.21.15\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.21.15.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.21.15.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Decide if this caption is a correct summary of the image: $CAPTION. | Yes, it is. | No, it is not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.22.16\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.22.16.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.22.16.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Would \"$CAPTION\" be a good image summary? | Yes, it would. | No, it would not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.23.17\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.23.17.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.23.17.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Would the caption \"$CAPTION\" fit the picture? | Yes, it would. | No, it would not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.24.18\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.24.18.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.24.18.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Could you use \"$CAPTION\" as a caption for the image? | Yes, you could. | No, you could not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.25.19\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.25.19.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">VQA</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.25.19.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">$QUESTION. Short English answer:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.26.20\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.26.20.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.26.20.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Question: $QUESTION. Brief answer (in $LANGUAGE):</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.27.21\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.27.21.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.27.21.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Give a short answer in $LANGUAGE to the following question. $QUESTION</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.28.22\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.28.22.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.28.22.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Answer the provided question in $LANGUAGE with three words or less. $QUESTION</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.29.23\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.29.23.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.29.23.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">What is the $LANGUAGE answer to this question? $QUESTION</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.30.24\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.30.24.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.30.24.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Briefly answer in $LANGUAGE. $QUESTION</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.31.25\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.31.25.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">VQG</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.31.25.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Given the image, generate a question in $LANGUAGE whose answer is: $ANSWER. Question:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.32.26\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.32.26.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.32.26.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Based on the image, create a question (in $LANGUAGE) for which the answer is \"$ANSWER\".</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.33.27\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.33.27.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.33.27.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">From the image provided, come up with a $LANGUAGE question that leads to the reply: $ANSWER. Question:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.34.28\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.34.28.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.34.28.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">What is a $LANGUAGE question for the image with the answer \"$ANSWER\"?</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.35.29\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.35.29.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.35.29.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Given the image, what would be a $LANGUAGE question that has as answer \"$ANSWER\"?</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.36.30\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.36.30.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">VQA with rational (instruction templates)</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.36.30.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Reason the answer to the following question. $QUESTION</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.37.31\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.37.31.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.37.31.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Use reasoning to come to an answer for this question. $QUESTION</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.38.32\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.38.32.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.38.32.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Think step-by-step to answer this question. $QUESTION</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.39.33\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.39.33.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.39.33.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Answer the following question and explain your answer. $QUESTION</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.40.34\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.40.34.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.40.34.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">$QUESTION What is the answer and why?</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.41.35\">\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.41.35.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">VQA with rational (label templates)</td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.41.35.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">$ANSWER. So the answer is $RATIONAL</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.42.36\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.42.36.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.42.36.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">$ANSWER so $RATIONAL</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.43.37\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.43.37.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.43.37.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">$RATIONAL. This means the answer is $ANSWER</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.44.38\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.44.38.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.44.38.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">The answer is $ANSWER because $RATIONAL.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.45.39\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.45.39.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.45.39.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">$ANSWER because $RATIONAL.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.46.40\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.46.40.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Rational Generation</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.46.40.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Question: $QUESTION Answer: $ANSWER. Explanation:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.47.41\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.47.41.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.47.41.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Question: $QUESTION: Answer: $ANSWER. The reason is because</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.48.42\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.48.42.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.48.42.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">The answer to the question \"$QUESTION\" is \"$ANSWER\". Why?</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.49.43\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.49.43.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.49.43.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Why is the answer to the question \"$QUESTION\" \"$ANSWER\"?</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.50.44\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.50.44.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.50.44.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Explain why the answer to the question \"$QUESTION\" is \"$ANSWER\"</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.51.45\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.51.45.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">ImageNet Classification</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.51.45.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">What is the main focus of the image? Short $LANGUAGE answer:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.52.46\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.52.46.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.52.46.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">What is in the image? Answer briefly in $LANGUAGE.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.53.47\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.53.47.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.53.47.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">This is an image of what? Answer briefly in $LANGUAGE.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.54.48\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.54.48.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.54.48.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">What is the central object in the image? Give a short $LANGUAGE answer.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.55.49\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.55.49.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.55.49.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">The focus of the image is on what? Short $LANGUAGE answer:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.56.50\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.56.50.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.56.50.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Question: This is an image of what? Answer briefly in $LANGUAGE.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.57.51\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.57.51.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.57.51.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">What is at the center of this picture? Short $LANGUAGE answer:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.58.52\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.58.52.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.58.52.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Give a short answer in $LANGUAGE to the following question. What is the main thing shown in the image?</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.59.53\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.59.53.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.59.53.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Complete the sentence in $LANGUAGE. This is a photo of a</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.60.54\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.60.54.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.60.54.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Name the main thing of this photo in $LANGUAGE:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.61.55\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.61.55.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.61.55.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">In less than 3 words in $LANGUAGE, what can be seen in this image?</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.62.56\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.62.56.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">ImageNet Matching</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T7.5.5.62.56.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Does this image show a $LABEL? | Yes, it does. | No, it does not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.63.57\">\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.63.57.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Question | Yes Answer | No Answer</td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.63.57.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Is there a $LABEL? | Yes, there is. | No, there is not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.64.58\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.64.58.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.64.58.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Are there any $LABEL in the picture? | Yes, there are. | No, there are not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.65.59\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.65.59.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.65.59.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Does the image contain a $LABEL? | Yes, it does. | No, it does not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.66.60\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.66.60.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.66.60.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Yes or no, there is a $LABEL in the photo. | Yes | No</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.67.61\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.67.61.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.67.61.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Yes or no, there is a $LABEL visible in the image. | Yes | No</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.68.62\">\n<td class=\"ltx_td\" id=\"A4.T7.5.5.68.62.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T7.5.5.68.62.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Does this picture have a $LABEL in it? | Yes, it does. | No, it does not.</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T7.5.5.69.63\">\n<td class=\"ltx_td ltx_border_bb\" id=\"A4.T7.5.5.69.63.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n<td class=\"ltx_td ltx_align_left ltx_border_bb\" id=\"A4.T7.5.5.69.63.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">Can you see a $LABEL in the image? | Yes, you can. | No, you can not.</td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 7: </span>\nTemplates used for the training examples. For each example, we randomly select one template. LLaVA examples are used as is since they are already in instruction form. : Template is translated to the 95 languages.\n</figcaption>\n</figure>",
            "capture": "Table 7: \nTemplates used for the training examples. For each example, we randomly select one template. LLaVA examples are used as is since they are already in instruction form. : Template is translated to the 95 languages.\n"
        },
        "10": {
            "table_html": "<figure class=\"ltx_table\" id=\"A4.T8\">\n<table class=\"ltx_tabular ltx_centering ltx_guessed_headers ltx_align_middle\" id=\"A4.T8.1\">\n<thead class=\"ltx_thead\">\n<tr class=\"ltx_tr\" id=\"A4.T8.1.1.1\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T8.1.1.1.1\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T8.1.1.1.1.1\">Dataset</span></th>\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_column ltx_border_tt\" id=\"A4.T8.1.1.1.2\"><span class=\"ltx_text ltx_font_bold\" id=\"A4.T8.1.1.1.2.1\">Template</span></th>\n</tr>\n</thead>\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A4.T8.1.2.1\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T8.1.2.1.1\">xFlickrCo, XM3600</td>\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A4.T8.1.2.1.2\">Caption in $LANGUAGE:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T8.1.3.2\">\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T8.1.3.2.1\">xGQA, MaXM</td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T8.1.3.2.2\">Question: $QUESTION Short answer in $LANGUAGE:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T8.1.4.3\">\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T8.1.4.3.1\">XVNLI</td>\n<td class=\"ltx_td ltx_align_left\" id=\"A4.T8.1.4.3.2\">Is it guaranteed true that \"$HYPOTHESIS\"? Yes, no, or maybe? Answer in English:</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A4.T8.1.5.4\">\n<td class=\"ltx_td ltx_align_left ltx_border_bb\" id=\"A4.T8.1.5.4.1\">MaRVL</td>\n<td class=\"ltx_td ltx_align_left ltx_border_bb\" id=\"A4.T8.1.5.4.2\">Based on the two images, is it correct to say \"$STATEMENT\"? Yes or no? Answer in English:</td>\n</tr>\n</tbody>\n</table>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 8: </span>\nTemplates used for evaluation.\nXVNLI labels \u2018entailment\u2019, \u2018contradiction\u2019, and \u2018neutral\u2019 are remapped to \u2018yes\u2019, \u2018no\u2019, \u2018maybe\u2019, respectively; MaRVL labels \u2018true\u2019 &amp; \u2018false\u2019 are remapped to \u2018yes\u2019, \u2018no\u2019, respectively.\n</figcaption>\n</figure>",
            "capture": "Table 8: \nTemplates used for evaluation.\nXVNLI labels \u2018entailment\u2019, \u2018contradiction\u2019, and \u2018neutral\u2019 are remapped to \u2018yes\u2019, \u2018no\u2019, \u2018maybe\u2019, respectively; MaRVL labels \u2018true\u2019 & \u2018false\u2019 are remapped to \u2018yes\u2019, \u2018no\u2019, respectively.\n"
        },
        "11": {
            "table_html": "<figure class=\"ltx_table\" id=\"A6.T9\">\n<table class=\"ltx_tabular ltx_centering ltx_guessed_headers ltx_align_middle\" id=\"A6.T9.1\">\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A6.T9.1.1.1\">\n<td class=\"ltx_td ltx_border_tt\" id=\"A6.T9.1.1.1.1\"></td>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T9.1.1.1.2\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T9.1.1.1.2.1\">bn</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T9.1.1.1.3\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T9.1.1.1.3.1\">de</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T9.1.1.1.4\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T9.1.1.1.4.1\">id</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T9.1.1.1.5\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T9.1.1.1.5.1\">ko</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T9.1.1.1.6\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T9.1.1.1.6.1\">pt</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T9.1.1.1.7\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T9.1.1.1.7.1\">ru</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T9.1.1.1.8\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T9.1.1.1.8.1\">zh</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T9.1.2.2\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A6.T9.1.2.2.1\">mBLIP mT0-XL (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T9.1.2.2.2\">38.51</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T9.1.2.2.3\">40.53</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T9.1.2.2.4\">38.34</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T9.1.2.2.5\">38.31</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T9.1.2.2.6\">40.15</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T9.1.2.2.7\">39.59</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T9.1.2.2.8\">38.99</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T9.1.3.3\">\n<td class=\"ltx_td ltx_align_left\" id=\"A6.T9.1.3.3.1\">mBLIP mT0-XL (finetuned)</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.3.3.2\">45.21</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.3.3.3\">50.32</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.3.3.4\">46.80</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.3.3.5\">46.28</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.3.3.6\">49.12</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.3.3.7\">48.94</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.3.3.8\">47.28</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T9.1.4.4\">\n<td class=\"ltx_td ltx_align_left\" id=\"A6.T9.1.4.4.1\">mBLIP BLOOMZ-7B (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.4.4.2\">38.96</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.4.4.3\">37.04</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.4.4.4\">39.99</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.4.4.5\">29.06</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.4.4.6\">41.78</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.4.4.7\">37.55</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T9.1.4.4.8\">39.72</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T9.1.5.5\">\n<td class=\"ltx_td ltx_align_left ltx_border_bb\" id=\"A6.T9.1.5.5.1\">mBLIP BLOOMZ-7B (finetuned)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T9.1.5.5.2\">46.90</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T9.1.5.5.3\">42.86</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T9.1.5.5.4\">48.01</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T9.1.5.5.5\">31.56</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T9.1.5.5.6\">51.99</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T9.1.5.5.7\">43.44</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T9.1.5.5.8\">49.64</td>\n</tr>\n</tbody>\n</table>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 9: </span>\nResults in all languages for xGQA. Finetuned results are averaged over 3 seeds.\n</figcaption>\n</figure>",
            "capture": "Table 9: \nResults in all languages for xGQA. Finetuned results are averaged over 3 seeds.\n"
        },
        "12": {
            "table_html": "<figure class=\"ltx_table\" id=\"A6.T10\">\n<table class=\"ltx_tabular ltx_centering ltx_guessed_headers ltx_align_middle\" id=\"A6.T10.1\">\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A6.T10.1.1.1\">\n<td class=\"ltx_td ltx_border_tt\" id=\"A6.T10.1.1.1.1\"></td>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T10.1.1.1.2\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T10.1.1.1.2.1\">ar</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T10.1.1.1.3\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T10.1.1.1.3.1\">es</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T10.1.1.1.4\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T10.1.1.1.4.1\">fr</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T10.1.1.1.5\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T10.1.1.1.5.1\">ru</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T10.1.2.2\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A6.T10.1.2.2.1\">mBLIP mT0-XL (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T10.1.2.2.2\">56.26</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T10.1.2.2.3\">57.57</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T10.1.2.2.4\">58.52</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T10.1.2.2.5\">58.26</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T10.1.3.3\">\n<td class=\"ltx_td ltx_align_left\" id=\"A6.T10.1.3.3.1\">mBLIP mT0-XL (finetuned)</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T10.1.3.3.2\">73.80</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T10.1.3.3.3\">77.62</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T10.1.3.3.4\">76.87</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T10.1.3.3.5\">77.33</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T10.1.4.4\">\n<td class=\"ltx_td ltx_align_left\" id=\"A6.T10.1.4.4.1\">mBLIP BLOOMZ-7B (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T10.1.4.4.2\">56.26</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T10.1.4.4.3\">56.17</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T10.1.4.4.4\">57.74</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T10.1.4.4.5\">51.65</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T10.1.5.5\">\n<td class=\"ltx_td ltx_align_left ltx_border_bb\" id=\"A6.T10.1.5.5.1\">mBLIP BLOOMZ-7B (finetuned)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T10.1.5.5.2\">68.90</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T10.1.5.5.3\">68.81</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T10.1.5.5.4\">71.57</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T10.1.5.5.5\">58.55</td>\n</tr>\n</tbody>\n</table>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 10: </span>\nResults in all languages for XVNLI. Finetuned results are averaged over 3 seeds.\n</figcaption>\n</figure>",
            "capture": "Table 10: \nResults in all languages for XVNLI. Finetuned results are averaged over 3 seeds.\n"
        },
        "13": {
            "table_html": "<figure class=\"ltx_table\" id=\"A6.T11\">\n<table class=\"ltx_tabular ltx_centering ltx_guessed_headers ltx_align_middle\" id=\"A6.T11.1\">\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A6.T11.1.1.1\">\n<td class=\"ltx_td ltx_border_tt\" id=\"A6.T11.1.1.1.1\"></td>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T11.1.1.1.2\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T11.1.1.1.2.1\">id</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T11.1.1.1.3\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T11.1.1.1.3.1\">sw</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T11.1.1.1.4\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T11.1.1.1.4.1\">ta</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T11.1.1.1.5\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T11.1.1.1.5.1\">tr</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T11.1.1.1.6\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T11.1.1.1.6.1\">zh</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T11.1.2.2\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A6.T11.1.2.2.1\">mBLIP mT0-XL (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T11.1.2.2.2\">64.89</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T11.1.2.2.3\">64.80</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T11.1.2.2.4\">69.65</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T11.1.2.2.5\">68.05</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T11.1.2.2.6\">65.91</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T11.1.3.3\">\n<td class=\"ltx_td ltx_align_left\" id=\"A6.T11.1.3.3.1\">mBLIP mT0-XL (finetuned)</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.3.3.2\">75.09</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.3.3.3\">74.61</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.3.3.4\">75.93</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.3.3.5\">74.32</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.3.3.6\">75.72</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T11.1.4.4\">\n<td class=\"ltx_td ltx_align_left\" id=\"A6.T11.1.4.4.1\">mBLIP BLOOMZ-7B (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.4.4.2\">59.13</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.4.4.3\">56.23</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.4.4.4\">60.31</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.4.4.5\">57.71</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T11.1.4.4.6\">59.68</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T11.1.5.5\">\n<td class=\"ltx_td ltx_align_left ltx_border_bb\" id=\"A6.T11.1.5.5.1\">mBLIP BLOOMZ-7B (finetuned)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T11.1.5.5.2\">80.08</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T11.1.5.5.3\">69.71</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T11.1.5.5.4\">77.38</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T11.1.5.5.5\">61.38</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T11.1.5.5.6\">81.16</td>\n</tr>\n</tbody>\n</table>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 11: </span>\nResults in all languages for MaRVL. Finetuned results are averaged over 3 seeds.\n</figcaption>\n</figure>",
            "capture": "Table 11: \nResults in all languages for MaRVL. Finetuned results are averaged over 3 seeds.\n"
        },
        "14": {
            "table_html": "<figure class=\"ltx_table\" id=\"A6.T12\">\n<table class=\"ltx_tabular ltx_centering ltx_guessed_headers ltx_align_middle\" id=\"A6.T12.1\">\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A6.T12.1.1.1\">\n<td class=\"ltx_td ltx_border_tt\" id=\"A6.T12.1.1.1.1\"></td>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T12.1.1.1.2\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T12.1.1.1.2.1\">fr</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T12.1.1.1.3\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T12.1.1.1.3.1\">hi</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T12.1.1.1.4\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T12.1.1.1.4.1\">iw</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T12.1.1.1.5\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T12.1.1.1.5.1\">ro</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T12.1.1.1.6\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T12.1.1.1.6.1\">th</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T12.1.1.1.7\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T12.1.1.1.7.1\">zh</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T12.1.2.2\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A6.T12.1.2.2.1\">mBLIP mT0-XL (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T12.1.2.2.2\">40.61</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T12.1.2.2.3\">48.30</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T12.1.2.2.4\">35.56</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T12.1.2.2.5\">41.74</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T12.1.2.2.6\">53.97</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T12.1.2.2.7\">26.06</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T12.1.3.3\">\n<td class=\"ltx_td ltx_align_left ltx_border_bb\" id=\"A6.T12.1.3.3.1\">mBLIP BLOOMZ-7B (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T12.1.3.3.2\">22.87</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T12.1.3.3.3\">52.38</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T12.1.3.3.4\">18.41</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T12.1.3.3.5\">31.83</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T12.1.3.3.6\">17.22</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T12.1.3.3.7\">24.76</td>\n</tr>\n</tbody>\n</table>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 12: </span>\nResults in all languages for MaXM.\n</figcaption>\n</figure>",
            "capture": "Table 12: \nResults in all languages for MaXM.\n"
        },
        "15": {
            "table_html": "<figure class=\"ltx_table\" id=\"A6.T13\">\n<table class=\"ltx_tabular ltx_centering ltx_guessed_headers ltx_align_middle\" id=\"A6.T13.1\">\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A6.T13.1.1.1\">\n<td class=\"ltx_td ltx_border_tt\" id=\"A6.T13.1.1.1.1\"></td>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T13.1.1.1.2\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T13.1.1.1.2.1\">de</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T13.1.1.1.3\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T13.1.1.1.3.1\">es</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T13.1.1.1.4\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T13.1.1.1.4.1\">id</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T13.1.1.1.5\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T13.1.1.1.5.1\">ja</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T13.1.1.1.6\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T13.1.1.1.6.1\">ru</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T13.1.1.1.7\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T13.1.1.1.7.1\">tr</span></th>\n<th class=\"ltx_td ltx_align_right ltx_th ltx_th_column ltx_border_tt\" id=\"A6.T13.1.1.1.8\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T13.1.1.1.8.1\">zh</span></th>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T13.1.2.2\">\n<td class=\"ltx_td ltx_align_left ltx_border_t\" id=\"A6.T13.1.2.2.1\">mBLIP mT0-XL (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T13.1.2.2.2\">58.23</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T13.1.2.2.3\">64.86</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T13.1.2.2.4\">47.44</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T13.1.2.2.5\">33.27</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T13.1.2.2.6\">41.77</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T13.1.2.2.7\">35.18</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T13.1.2.2.8\">29.98</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T13.1.3.3\">\n<td class=\"ltx_td ltx_align_left ltx_border_bb\" id=\"A6.T13.1.3.3.1\">mBLIP BLOOMZ-7B (zero-shot)</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T13.1.3.3.2\">50.50</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T13.1.3.3.3\">64.89</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T13.1.3.3.4\">54.42</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T13.1.3.3.5\">29.10</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T13.1.3.3.6\">38.36</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T13.1.3.3.7\">25.08</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T13.1.3.3.8\">32.42</td>\n</tr>\n</tbody>\n</table>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 13: </span>\nResults in all languages for xFlickrCo.\n</figcaption>\n</figure>",
            "capture": "Table 13: \nResults in all languages for xFlickrCo.\n"
        },
        "16": {
            "table_html": "<figure class=\"ltx_table\" id=\"A6.T14\">\n<div class=\"ltx_inline-block ltx_align_center ltx_transformed_outer\" id=\"A6.T14.1\" style=\"width:433.6pt;height:122.2pt;vertical-align:-0.7pt;\"><span class=\"ltx_transformed_inner\" style=\"transform:translate(-72.3pt,20.3pt) scale(0.749809755469849,0.749809755469849) ;\">\n<table class=\"ltx_tabular ltx_guessed_headers ltx_align_middle\" id=\"A6.T14.1.1\">\n<tbody class=\"ltx_tbody\">\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.1.1\">\n<th class=\"ltx_td ltx_th ltx_th_row ltx_border_tt\" id=\"A6.T14.1.1.1.1.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.2.1\">ar</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.3.1\">bn</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.4.1\">cs</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.5.1\">da</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.6.1\">de</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.7.1\">el</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.8.1\">es</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.9.1\">fa</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.10.1\">fi</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.11.1\">fil</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.12.1\">fr</span></td>\n<td class=\"ltx_td ltx_align_right ltx_border_tt\" id=\"A6.T14.1.1.1.1.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.1.1.13.1\">he</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.2.2\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row ltx_border_t\" id=\"A6.T14.1.1.2.2.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP mT0-XL (zero-shot)</th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">21.13</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">11.30</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">31.84</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">44.19</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">32.48</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">23.36</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">62.61</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.00</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">16.78</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17.71</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">57.64</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.2.2.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">18.69</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.3.3\">\n<th class=\"ltx_td ltx_align_left ltx_th ltx_th_row\" id=\"A6.T14.1.1.3.3.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">mBLIP BLOOMZ-7B (zero-shot)</th>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">27.78</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">16.12</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">21.77</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">25.25</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">30.04</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">14.12</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">60.03</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">13.84</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.69</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.99</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">60.42</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.3.3.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">7.16</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.4.4\">\n<th class=\"ltx_td ltx_th ltx_th_row\" id=\"A6.T14.1.1.4.4.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.2.1\">hi</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.3.1\">hr</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.4.1\">hu</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.5.1\">id</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.6.1\">it</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.7.1\">ja</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.8.1\">ko</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.9.1\">mi</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.10.1\">nl</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.11.1\">no</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.12.1\">pl</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.4.4.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.4.4.13.1\">pt</span></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.5.5\">\n<th class=\"ltx_td ltx_th ltx_th_row\" id=\"A6.T14.1.1.5.5.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">16.07</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">5.18</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">21.54</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">38.53</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">45.19</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">33.23</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">10.39</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">4.09</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">55.72</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">46.15</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">31.22</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.5.5.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">53.13</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.6.6\">\n<th class=\"ltx_td ltx_th ltx_th_row\" id=\"A6.T14.1.1.6.6.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">24.91</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">2.13</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">10.99</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">45.29</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">42.40</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">25.43</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">2.54</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.02</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">45.54</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">25.01</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">20.65</td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.6.6.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">47.79</td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.7.7\">\n<th class=\"ltx_td ltx_th ltx_th_row\" id=\"A6.T14.1.1.7.7.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.2.1\">quz</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.3.1\">ro</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.4.1\">ru</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.5.1\">sv</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.6.1\">sw</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.7.1\">te</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.8.1\">th</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.9.1\">tr</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.10.1\">uk</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.11.1\">vi</span></td>\n<td class=\"ltx_td ltx_align_right\" id=\"A6.T14.1.1.7.7.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"><span class=\"ltx_text ltx_font_bold\" id=\"A6.T14.1.1.7.7.12.1\">zh</span></td>\n<td class=\"ltx_td\" id=\"A6.T14.1.1.7.7.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.8.8\">\n<th class=\"ltx_td ltx_th ltx_th_row\" id=\"A6.T14.1.1.8.8.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">1.08</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">21.71</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">27.25</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">48.38</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">11.76</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">11.20</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">41.93</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">22.64</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.00</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">39.24</td>\n<td class=\"ltx_td ltx_align_right ltx_border_t\" id=\"A6.T14.1.1.8.8.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">13.48</td>\n<td class=\"ltx_td ltx_border_t\" id=\"A6.T14.1.1.8.8.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n</tr>\n<tr class=\"ltx_tr\" id=\"A6.T14.1.1.9.9\">\n<th class=\"ltx_td ltx_th ltx_th_row ltx_border_bb\" id=\"A6.T14.1.1.9.9.1\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></th>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.2\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">0.02</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.3\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">17.62</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.4\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">22.83</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.5\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">31.77</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.6\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.45</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.7\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.65</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.8\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.16</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.9\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">14.21</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.10\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">8.97</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.11\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">54.29</td>\n<td class=\"ltx_td ltx_align_right ltx_border_bb\" id=\"A6.T14.1.1.9.9.12\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\">14.65</td>\n<td class=\"ltx_td ltx_border_bb\" id=\"A6.T14.1.1.9.9.13\" style=\"padding-top:-0.15pt;padding-bottom:-0.15pt;\"></td>\n</tr>\n</tbody>\n</table>\n</span></div>\n<figcaption class=\"ltx_caption ltx_centering\"><span class=\"ltx_tag ltx_tag_table\">Table 14: </span>\nResults in all languages for XM3600.\n</figcaption>\n</figure>",
            "capture": "Table 14: \nResults in all languages for XM3600.\n"
        }
    },
    "image_paths": {
        "1": {
            "figure_path": "2307.06930v3_figure_1.png",
            "caption": "Figure 1: The mBLIP architecture: A Q-Former encodes the image in learned query tokens which are projected to the LLM space. We initialize the Q-Former from a BLIP-2 model and re-align it to the multilingual LLM with a multilingual task mix. The image encoder and LLM (aside from LoRA weights) are frozen during training.",
            "url": "http://arxiv.org/html/2307.06930v3/x1.png"
        },
        "2(a)": {
            "figure_path": "2307.06930v3_figure_2(a).png",
            "caption": "(a) xGQA\nFigure 2: Cross-lingual transfer of models fine-tuned on English.\nThe smaller gap of mBLIP mT0 between high- and low-resource\nlanguages suggests better transfer capabilities. (CCLM 4M from Zeng et al. (2022b) v1 on arXiv.)",
            "url": "http://arxiv.org/html/2307.06930v3/x2.png"
        },
        "2(b)": {
            "figure_path": "2307.06930v3_figure_2(b).png",
            "caption": "(b) MaRVL\nFigure 2: Cross-lingual transfer of models fine-tuned on English.\nThe smaller gap of mBLIP mT0 between high- and low-resource\nlanguages suggests better transfer capabilities. (CCLM 4M from Zeng et al. (2022b) v1 on arXiv.)",
            "url": "http://arxiv.org/html/2307.06930v3/x3.png"
        },
        "2(c)": {
            "figure_path": "2307.06930v3_figure_2(c).png",
            "caption": "(c) XVNLI\nFigure 2: Cross-lingual transfer of models fine-tuned on English.\nThe smaller gap of mBLIP mT0 between high- and low-resource\nlanguages suggests better transfer capabilities. (CCLM 4M from Zeng et al. (2022b) v1 on arXiv.)",
            "url": "http://arxiv.org/html/2307.06930v3/x4.png"
        },
        "3": {
            "figure_path": "2307.06930v3_figure_3.png",
            "caption": "Figure 3: Multilingual examples (with translations from Google Translate in parentheses). While the first row shows that the model can handle captioning and QA in diverse languages, the second row shows some failure cases.\nWe use beam search (5 beams) with a repetition penalty of 1.5.",
            "url": "http://arxiv.org/html/2307.06930v3/x5.png"
        }
    },
    "validation": true,
    "references": [
        {
            "1": {
                "title": "Flamingo: a\nVisual Language Model for Few-Shot Learning.",
                "author": "Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr,\nYana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds,\nRoman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina\nSamangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew\nBrock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo\nBarreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. 2022.",
                "venue": "CoRR, abs/2204.14198.",
                "url": "https://doi.org/10.48550/arXiv.2204.14198"
            }
        },
        {
            "2": {
                "title": "Qwen-VL: A\nFrontier Large Vision-Language Model with Versatile Abilities.",
                "author": "Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang\nLin, Chang Zhou, and Jingren Zhou. 2023.",
                "venue": "CoRR, abs/2308.12966.",
                "url": "https://doi.org/10.48550/ARXIV.2308.12966"
            }
        },
        {
            "3": {
                "title": "Language Models are\nFew-Shot Learners.",
                "author": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan,\nPrafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda\nAskell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan,\nRewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter,\nChristopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray,\nBenjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford,\nIlya Sutskever, and Dario Amodei. 2020.",
                "venue": "arXiv:2005.14165 [cs].",
                "url": "http://arxiv.org/abs/2005.14165"
            }
        },
        {
            "4": {
                "title": "IGLUE: A Benchmark for Transfer Learning across Modalities,\nTasks, and Languages.",
                "author": "Emanuele Bugliarello, Fangyu Liu, Jonas Pfeiffer, Siva Reddy, Desmond Elliott,\nEdoardo Maria Ponti, and Ivan Vulic. 2022.",
                "venue": "In International Conference on Machine Learning, ICML\n2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of\nProceedings of Machine Learning Research, pages 2370\u20132392.\nPMLR.",
                "url": "https://proceedings.mlr.press/v162/bugliarello22a.html"
            }
        },
        {
            "5": {
                "title": "MaXM: Towards multilingual visual question answering.",
                "author": "Soravit Changpinyo, Linting Xue, Idan Szpektor, Ashish V. Thapliyal, Julien\nAmelot, Michal Yarom, Xi Chen, and Radu Soricut. 2022.",
                "venue": "arXiv preprint arXiv:2209.05401.",
                "url": null
            }
        },
        {
            "6": {
                "title": "PaLI-X: On\nScaling up a Multilingual Vision and Language Model.",
                "author": "Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo,\nJialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak\nShakeri, Mostafa Dehghani, Daniel Salz, Mario Lucic, Michael Tschannen, Arsha\nNagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina\nPietrzyk, Marvin Ritter, A. J. Piergiovanni, Matthias Minderer, Filip\nPavetic, Austin Waters, Gang Li, Ibrahim Alabdulmohsin, Lucas Beyer, Julien\nAmelot, Kenton Lee, Andreas Peter Steiner, Yang Li, Daniel Keysers, Anurag\nArnab, Yuanzhong Xu, Keran Rong, Alexander Kolesnikov, Mojtaba Seyedhosseini,\nAnelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. 2023.",
                "venue": "CoRR, abs/2305.18565.",
                "url": "https://doi.org/10.48550/arXiv.2305.18565"
            }
        },
        {
            "7": {
                "title": "PaLI: A\nJointly-Scaled Multilingual Language-Image Model.",
                "author": "Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski,\nDaniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer,\nAlexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari,\nGaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo,\nMojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas\nSteiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. 2022.",
                "venue": "CoRR, abs/2209.06794.",
                "url": "https://doi.org/10.48550/arXiv.2209.06794"
            }
        },
        {
            "8": {
                "title": "UNITER:\nUNiversal Image-TExt Representation Learning.",
                "author": "Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan,\nYu Cheng, and Jingjing Liu. 2020.",
                "venue": "In Computer Vision - ECCV 2020 - 16th European\nConference, Glasgow, UK, August 23-28, 2020, Proceedings, Part\nXXX, volume 12375 of Lecture Notes in Computer Science, pages\n104\u2013120. Springer.",
                "url": "https://doi.org/10.1007/978-3-030-58577-8_7"
            }
        },
        {
            "9": {
                "title": "Vicuna: An\nopen-source chatbot impressing gpt-4 with 90%* chatgpt quality.",
                "author": "Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin\nZheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and\nEric P. Xing. 2023.",
                "venue": null,
                "url": "https://lmsys.org/blog/2023-03-30-vicuna/"
            }
        },
        {
            "10": {
                "title": "Scaling\nInstruction-Finetuned Language Models.",
                "author": "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus,\nEric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson,\nShixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha\nChowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Y. Zhao, Yanping\nHuang, Andrew M. Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob\nDevlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022.",
                "venue": "CoRR, abs/2210.11416.",
                "url": "https://doi.org/10.48550/arXiv.2210.11416"
            }
        },
        {
            "11": {
                "title": "Unsupervised\nCross-lingual Representation Learning at Scale.",
                "author": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume\nWenzek, Francisco Guzm\u00e1n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and\nVeselin Stoyanov. 2020.",
                "venue": "In Proceedings of the 58th Annual Meeting of the\nAssociation for Computational Linguistics, ACL 2020, Online, July\n5-10, 2020, pages 8440\u20138451. Association for Computational Linguistics.",
                "url": "https://doi.org/10.18653/v1/2020.acl-main.747"
            }
        },
        {
            "12": {
                "title": "No Language\nLeft Behind: Scaling Human-Centered Machine Translation.",
                "author": "Marta R. Costa-juss\u00e0, James Cross, Onur \u00c7elebi, Maha Elbayad, Kenneth\nHeafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean\nMaillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula,\nLo\u00efc Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman,\nSemarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau\nTran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela\nFan, Cynthia Gao, Vedanuj Goswami, Francisco Guzm\u00e1n, Philipp Koehn,\nAlexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and\nJeff Wang. 2022.",
                "venue": "CoRR, abs/2207.04672.",
                "url": "https://doi.org/10.48550/arXiv.2207.04672"
            }
        },
        {
            "13": {
                "title": "InstructBLIP:\nTowards General-purpose Vision-Language Models with Instruction\nTuning.",
                "author": "Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao,\nWeisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. 2023.",
                "venue": "CoRR, abs/2305.06500.",
                "url": "https://doi.org/10.48550/arXiv.2305.06500"
            }
        },
        {
            "14": {
                "title": "ImageNet: A\nlarge-scale hierarchical image database.",
                "author": "J. Deng, W. Dong, R. Socher, L. Li, Kai Li, and Li Fei-Fei. 2009.",
                "venue": "In 2009 IEEE Conference on Computer Vision and\nPattern Recognition, pages 248\u2013255.",
                "url": "https://doi.org/10.1109/CVPR.2009.5206848"
            }
        },
        {
            "15": {
                "title": "LLM.int8():\n8-bit Matrix Multiplication for Transformers at Scale.",
                "author": "Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022.",
                "venue": "CoRR, abs/2208.07339.",
                "url": "https://doi.org/10.48550/arXiv.2208.07339"
            }
        },
        {
            "16": {
                "title": "QLoRA:\nEfficient Finetuning of Quantized LLMs.",
                "author": "Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023.",
                "venue": "CoRR, abs/2305.14314.",
                "url": "https://doi.org/10.48550/arXiv.2305.14314"
            }
        },
        {
            "17": {
                "title": "BERT: Pre-training\nof Deep Bidirectional Transformers for Language Understanding.",
                "author": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019.",
                "venue": "In Proceedings of the 2019 Conference of the North\nAmerican Chapter of the Association for Computational Linguistics:\nHuman Language Technologies, NAACL-HLT 2019, Minneapolis, MN,\nUSA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages\n4171\u20134186. Association for Computational Linguistics.",
                "url": "https://doi.org/10.18653/v1/n19-1423"
            }
        },
        {
            "18": {
                "title": "An Image is Worth 16x16\nWords: Transformers for Image Recognition at Scale.",
                "author": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn,\nXiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg\nHeigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2020.",
                "venue": "arXiv preprint, abs/2010.11929.",
                "url": "https://arxiv.org/abs/2010.11929"
            }
        },
        {
            "19": {
                "title": "EVA: exploring\nthe limits of masked visual representation learning at scale.",
                "author": "Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun\nHuang, Xinlong Wang, and Yue Cao. 2022.",
                "venue": "CoRR, abs/2211.07636.",
                "url": "https://doi.org/10.48550/arXiv.2211.07636"
            }
        },
        {
            "20": {
                "title": "DataComp: In\nsearch of the next generation of multimodal datasets.",
                "author": "Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios\nSmyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu\nZhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M. Pratt, Vivek\nRamanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu,\nMehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner,\nShuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh,\nAlex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig\nSchmidt. 2023.",
                "venue": "CoRR, abs/2304.14108.",
                "url": "https://doi.org/10.48550/arXiv.2304.14108"
            }
        },
        {
            "21": {
                "title": "LLaMA-Adapter\nV2: Parameter-Efficient Visual Instruction Model.",
                "author": "Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei\nZhang, Pan Lu, Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. 2023.",
                "venue": "CoRR, abs/2304.15010.",
                "url": "https://doi.org/10.48550/arXiv.2304.15010"
            }
        },
        {
            "22": {
                "title": "Babel-ImageNet:\nMassively Multilingual Evaluation of Vision-and-Language\nRepresentations.",
                "author": "Gregor Geigle, Radu Timofte, and Goran Glavas. 2023.",
                "venue": "CoRR, abs/2306.08658.",
                "url": "https://doi.org/10.48550/arXiv.2306.08658"
            }
        },
        {
            "23": {
                "title": "Making the V in\nVQA Matter: Elevating the Role of Image Understanding in Visual\nQuestion Answering.",
                "author": "Y. Goyal, T. Khot, D. Summers-Stay, D. Batra, and D. Parikh. 2017.",
                "venue": "In 2017 IEEE Conference on Computer Vision and\nPattern Recognition (CVPR), pages 6325\u20136334.",
                "url": "https://doi.org/10.1109/CVPR.2017.670"
            }
        },
        {
            "24": {
                "title": "LoRA:\nLow-Rank Adaptation of Large Language Models.",
                "author": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean\nWang, Lu Wang, and Weizhu Chen. 2022.",
                "venue": "In The Tenth International Conference on Learning\nRepresentations, ICLR 2022, Virtual Event, April 25-29, 2022.\nOpenReview.net.",
                "url": "https://openreview.net/forum?id=nZeVKeeFYf9"
            }
        },
        {
            "25": {
                "title": "GQA: A New\nDataset for Real-World Visual Reasoning and Compositional\nQuestion Answering.",
                "author": "Drew A. Hudson and Christopher D. Manning. 2019.",
                "venue": "In IEEE Conference on Computer Vision and Pattern\nRecognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019,\npages 6700\u20136709. Computer Vision Foundation / IEEE.",
                "url": "https://doi.org/10.1109/CVPR.2019.00686"
            }
        },
        {
            "26": {
                "title": "Deep\nVisual-Semantic Alignments for Generating Image Descriptions.",
                "author": "A. Karpathy and L. Fei-Fei. 2017.",
                "venue": "IEEE Transactions on Pattern Analysis and Machine\nIntelligence, 39(4):664\u2013676.",
                "url": "https://doi.org/10.1109/TPAMI.2016.2598339"
            }
        },
        {
            "27": {
                "title": "BLIP-2:\nBootstrapping Language-Image Pre-training with Frozen Image\nEncoders and Large Language Models.",
                "author": "Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi.\n2023a.",
                "venue": "CoRR, abs/2301.12597.",
                "url": "https://doi.org/10.48550/arXiv.2301.12597"
            }
        },
        {
            "28": {
                "title": "BLIP:\nBootstrapping Language-Image Pre-training for Unified\nVision-Language Understanding and Generation.",
                "author": "Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. 2022.",
                "venue": "In International Conference on Machine Learning, ICML\n2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of\nProceedings of Machine Learning Research, pages 12888\u201312900.\nPMLR.",
                "url": "https://proceedings.mlr.press/v162/li22n.html"
            }
        },
        {
            "29": {
                "title": "Align before Fuse:\nVision and Language Representation Learning with Momentum\nDistillation.",
                "author": "Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq R. Joty,\nCaiming Xiong, and Steven C. H. Hoi. 2021.",
                "venue": "arXiv preprint, abs/2107.07651.",
                "url": "https://arxiv.org/abs/2107.07651"
            }
        },
        {
            "30": {
                "title": "Why Does\nZero-Shot Cross-Lingual Generation Fail? An Explanation and a\nSolution.",
                "author": "Tianjian Li and Kenton Murray. 2023.",
                "venue": "CoRR, abs/2305.17325.",
                "url": "https://doi.org/10.48550/arXiv.2305.17325"
            }
        },
        {
            "31": {
                "title": "Oscar:\nObject-Semantics Aligned Pre-training for Vision-Language\nTasks.",
                "author": "Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan\nWang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. 2020.",
                "venue": "In Computer Vision - ECCV 2020 - 16th European\nConference, Glasgow, UK, August 23-28, 2020, Proceedings, Part\nXXX, volume 12375 of Lecture Notes in Computer Science, pages\n121\u2013137. Springer.",
                "url": "https://doi.org/10.1007/978-3-030-58577-8_8"
            }
        },
        {
            "32": {
                "title": "Evaluating\nObject Hallucination in Large Vision-Language Models.",
                "author": "Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen.\n2023b.",
                "venue": "CoRR, abs/2305.10355.",
                "url": "https://doi.org/10.48550/arXiv.2305.10355"
            }
        },
        {
            "33": {
                "title": "Unifying\ncross-lingual and cross-modal modeling towards weakly supervised multilingual\nvision-language pre-training.",
                "author": "Zejun Li, Zhihao Fan, Jingjing Chen, Qi Zhang, Xuanjing Huang, and Zhongyu Wei.\n2023c.",
                "venue": "In Proceedings of the 61st Annual Meeting of the Association\nfor Computational Linguistics (Volume 1: Long Papers), pages 5939\u20135958,\nToronto, Canada. Association for Computational Linguistics.",
                "url": "https://doi.org/10.18653/v1/2023.acl-long.327"
            }
        },
        {
            "34": {
                "title": "Microsoft\nCOCO: Common Objects in Context.",
                "author": "Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva\nRamanan, Piotr Doll\u00e1r, and C. Lawrence Zitnick. 2014.",
                "venue": "In Computer Vision - ECCV 2014 - 13th European\nConference, Zurich, Switzerland, September 6-12, 2014, Proceedings,\nPart V, volume 8693 of Lecture Notes in Computer Science,\npages 740\u2013755. Springer.",
                "url": "https://doi.org/10.1007/978-3-319-10602-1_48"
            }
        },
        {
            "35": {
                "title": "Visually\nGrounded Reasoning across Languages and Cultures.",
                "author": "Fangyu Liu, Emanuele Bugliarello, Edoardo Maria Ponti, Siva Reddy, Nigel\nCollier, and Desmond Elliott. 2021.",
                "venue": "In Proceedings of the 2021 Conference on Empirical\nMethods in Natural Language Processing, EMNLP 2021, Virtual\nEvent / Punta Cana, Dominican Republic, 7-11 November, 2021,\npages 10467\u201310485. Association for Computational Linguistics.",
                "url": "https://doi.org/10.18653/v1/2021.emnlp-main.818"
            }
        },
        {
            "36": {
                "title": "Improved\nBaselines with Visual Instruction Tuning.",
                "author": "Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023a.",
                "venue": "CoRR, abs/2310.03744.",
                "url": "https://doi.org/10.48550/ARXIV.2310.03744"
            }
        },
        {
            "37": {
                "title": "Visual\nInstruction Tuning.",
                "author": "Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023b.",
                "venue": "CoRR, abs/2304.08485.",
                "url": "https://doi.org/10.48550/arXiv.2304.08485"
            }
        },
        {
            "38": {
                "title": "Decoupled\nWeight Decay Regularization.",
                "author": "Ilya Loshchilov and Frank Hutter. 2019.",
                "venue": "In 7th International Conference on Learning\nRepresentations, ICLR 2019, New Orleans, LA, USA, May 6-9,\n2019. OpenReview.net.",
                "url": "https://openreview.net/forum?id=Bkg6RiCqY7"
            }
        },
        {
            "39": {
                "title": "ViLBERT: Pretraining Task-Agnostic Visiolinguistic\nRepresentations for Vision-and-Language Tasks.",
                "author": "Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019.",
                "venue": "In Advances in Neural Information Processing Systems\n32: Annual Conference on Neural Information Processing Systems\n2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada,\npages 13\u201323.",
                "url": "https://proceedings.neurips.cc/paper/2019/hash/c74d97b01eae257e44aa9d5bade97baf-Abstract.html"
            }
        },
        {
            "40": {
                "title": "Crosslingual\nGeneralization through Multitask Finetuning.",
                "author": "Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella\nBiderman, Teven Le Scao, M. Saiful Bari, Sheng Shen, Zheng Xin Yong, Hailey\nSchoelkopf, Xiangru Tang, Dragomir Radev, Alham Fikri Aji, Khalid Almubarak,\nSamuel Albanie, Zaid Alyafeai, Albert Webson, Edward Raff, and Colin Raffel.\n2022.",
                "venue": "CoRR, abs/2211.01786.",
                "url": "https://doi.org/10.48550/arXiv.2211.01786"
            }
        },
        {
            "41": {
                "title": "M3P:\nLearning Universal Representations via Multitask Multilingual\nMultimodal Pre-Training.",
                "author": "Minheng Ni, Haoyang Huang, Lin Su, Edward Cui, Taroon Bharti, Lijuan Wang,\nDongdong Zhang, and Nan Duan. 2021.",
                "venue": "In IEEE Conference on Computer Vision and Pattern\nRecognition, CVPR 2021, virtual, June 19-25, 2021, pages 3977\u20133986.\nComputer Vision Foundation / IEEE.",
                "url": "https://doi.org/10.1109/CVPR46437.2021.00397"
            }
        },
        {
            "42": {
                "title": "GPT-4 technical\nreport.",
                "author": "OpenAI. 2023.",
                "venue": "CoRR, abs/2303.08774.",
                "url": "https://doi.org/10.48550/arXiv.2303.08774"
            }
        },
        {
            "43": {
                "title": "Training language models to follow instructions with human feedback.",
                "author": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela\nMishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John\nSchulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda\nAskell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. 2022.",
                "venue": "In NeurIPS.",
                "url": "http://papers.nips.cc/paper_files/paper/2022/hash/b1efde53be364a73914f58805a001731-Abstract-Conference.html"
            }
        },
        {
            "44": {
                "title": "xGQA:\nCross-Lingual Visual Question Answering.",
                "author": "Jonas Pfeiffer, Gregor Geigle, Aishwarya Kamath, Jan-Martin O. Steitz, Stefan\nRoth, Ivan Vulic, and Iryna Gurevych. 2022.",
                "venue": "In Findings of the Association for Computational\nLinguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages\n2497\u20132511. Association for Computational Linguistics.",
                "url": "https://doi.org/10.18653/v1/2022.findings-acl.196"
            }
        },
        {
            "45": {
                "title": "mmT5: Modular\nMultilingual Pre-Training Solves Source Language\nHallucinations.",
                "author": "Jonas Pfeiffer, Francesco Piccinno, Massimo Nicosia, Xinyi Wang, Machel Reid,\nand Sebastian Ruder. 2023.",
                "venue": "CoRR, abs/2305.14224.",
                "url": "https://doi.org/10.48550/arXiv.2305.14224"
            }
        },
        {
            "46": {
                "title": "Flickr30k Entities:\nCollecting Region-to-Phrase Correspondences for Richer\nImage-to-Sentence Models.",
                "author": "Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia\nHockenmaier, and Svetlana Lazebnik. 2015.",
                "venue": "In 2015 IEEE International Conference on Computer\nVision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages\n2641\u20132649.",
                "url": "https://doi.org/10.1109/ICCV.2015.303"
            }
        },
        {
            "47": {
                "title": "LMCap:\nFew-shot Multilingual Image Captioning by Retrieval Augmented\nLanguage Model Prompting.",
                "author": "Rita Ramos, Bruno Martins, and Desmond Elliott. 2023.",
                "venue": "CoRR, abs/2305.19821.",
                "url": "https://doi.org/10.48550/arXiv.2305.19821"
            }
        },
        {
            "48": {
                "title": "Object Hallucination\nin Image Captioning.",
                "author": "Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate\nSaenko. 2018.",
                "venue": "In Proceedings of the 2018 Conference on Empirical\nMethods in Natural Language Processing, Brussels, Belgium,\nOctober 31 - November 4, 2018, pages 4035\u20134045. Association for\nComputational Linguistics.",
                "url": "https://doi.org/10.18653/v1/d18-1437"
            }
        },
        {
            "49": {
                "title": "Multitask\nPrompted Training Enables Zero-Shot Task Generalization.",
                "author": "Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika,\nZaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey,\nM. Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza\nSzczechla, Taewoon Kim, Gunjan Chhablani, Nihal V. Nayak, Debajyoti Datta,\nJonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen,\nZheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj,\nJos Rozen, Abheesht Sharma, Andrea Santilli, Thibault F\u00e9vry, Jason Alan\nFries, Ryan Teehan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and\nAlexander M. Rush. 2022.",
                "venue": "In The Tenth International Conference on Learning\nRepresentations, ICLR 2022, Virtual Event, April 25-29, 2022.\nOpenReview.net.",
                "url": "https://openreview.net/forum?id=9Vrb9D0WI4"
            }
        },
        {
            "50": {
                "title": "BLOOM: A\n176B-Parameter Open-Access Multilingual Language Model.",
                "author": "Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic,\nDaniel Hesslow, Roman Castagn\u00e9, Alexandra Sasha Luccioni, Fran\u00e7ois Yvon,\nMatthias Gall\u00e9, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert\nWebson, Pawan Sasanka Ammanamanchi, Thomas Wang, Beno\u00eet Sagot, Niklas\nMuennighoff, Albert Villanova del Moral, Olatunji Ruwase, Rachel Bawden, Stas\nBekman, Angelina McMillan-Major, Iz Beltagy, Huu Nguyen, Lucile Saulnier,\nSamson Tan, Pedro Ortiz Suarez, Victor Sanh, Hugo Lauren\u00e7on, Yacine Jernite,\nJulien Launay, Margaret Mitchell, Colin Raffel, Aaron Gokaslan, Adi Simhi,\nAitor Soroa, Alham Fikri Aji, Amit Alfassy, Anna Rogers, Ariel Kreisberg\nNitzav, Canwen Xu, Chenghao Mou, Chris Emezue, Christopher Klamm, Colin\nLeong, Daniel van Strien, David Ifeoluwa Adelani, and et al. 2022.",
                "venue": "CoRR, abs/2211.05100.",
                "url": "https://doi.org/10.48550/arXiv.2211.05100"
            }
        },
        {
            "51": {
                "title": "Don\u2019t stop fine-tuning: On training regimes for few-shot\ncross-lingual transfer with multilingual language models.",
                "author": "Fabian David Schmidt, Ivan Vuli\u0107, and Goran Glava\u0161. 2022.",
                "venue": "In Proceedings of the 2022 Conference on Empirical Methods in\nNatural Language Processing, pages 10725\u201310742.",
                "url": null
            }
        },
        {
            "52": {
                "title": "LAION-5B: An\nopen large-scale dataset for training next generation image-text models.",
                "author": "Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross\nWightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell\nWortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig\nSchmidt, Robert Kaczmarczyk, and Jenia Jitsev. 2022.",
                "venue": "CoRR, abs/2210.08402.",
                "url": "https://doi.org/10.48550/arXiv.2210.08402"
            }
        },
        {
            "53": {
                "title": "A-OKVQA: A\nbenchmark for visual question answering using world knowledge.",
                "author": "Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and\nRoozbeh Mottaghi. 2022.",
                "venue": "In Computer Vision - ECCV 2022 - 17th European Conference,\nTel Aviv, Israel, October 23-27, 2022, Proceedings, Part VIII, volume\n13668 of Lecture Notes in Computer Science, pages 146\u2013162. Springer.",
                "url": "https://doi.org/10.1007/978-3-031-20074-8_9"
            }
        },
        {
            "54": {
                "title": "ERNIE-UniX2:\nA Unified Cross-lingual Cross-modal Framework for Understanding\nand Generation.",
                "author": "Bin Shan, Yaqian Han, Weichong Yin, Shuohuan Wang, Yu Sun, Hao Tian, Hua Wu,\nand Haifeng Wang. 2022.",
                "venue": "CoRR, abs/2211.04861.",
                "url": "https://doi.org/10.48550/arXiv.2211.04861"
            }
        },
        {
            "55": {
                "title": "Conceptual Captions:\nA Cleaned, Hypernymed, Image Alt-text Dataset For Automatic\nImage Captioning.",
                "author": "Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. 2018.",
                "venue": "In Proceedings of the 56th Annual Meeting of the\nAssociation for Computational Linguistics (Volume 1: Long\nPapers), pages 2556\u20132565, Melbourne, Australia. Association for\nComputational Linguistics.",
                "url": "https://doi.org/10.18653/v1/P18-1238"
            }
        },
        {
            "56": {
                "title": "A Corpus for\nReasoning about Natural Language Grounded in Photographs.",
                "author": "Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi.\n2019.",
                "venue": "In Proceedings of the 57th Conference of the Association\nfor Computational Linguistics, ACL 2019, Florence, Italy, July\n28- August 2, 2019, Volume 1: Long Papers, pages 6418\u20136428.\nAssociation for Computational Linguistics.",
                "url": "https://doi.org/10.18653/v1/p19-1644"
            }
        },
        {
            "57": {
                "title": "Stanford alpaca: An instruction-following llama model.",
                "author": "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos\nGuestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023.",
                "venue": "https://github.com/tatsu-lab/stanford_alpaca.",
                "url": null
            }
        },
        {
            "58": {
                "title": "Crossmodal-3600:\nA Massively Multilingual Multimodal Evaluation Dataset.",
                "author": "Ashish V. Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. 2022.",
                "venue": "In Proceedings of the 2022 Conference on Empirical\nMethods in Natural Language Processing, EMNLP 2022, Abu Dhabi,\nUnited Arab Emirates, December 7-11, 2022, pages 715\u2013729.\nAssociation for Computational Linguistics.",
                "url": "https://aclanthology.org/2022.emnlp-main.45"
            }
        },
        {
            "59": {
                "title": "LLaMA: Open\nand Efficient Foundation Language Models.",
                "author": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne\nLachaux, Timoth\u00e9e Lacroix, Baptiste Rozi\u00e8re, Naman Goyal, Eric Hambro,\nFaisal Azhar, Aur\u00e9lien Rodriguez, Armand Joulin, Edouard Grave, and\nGuillaume Lample. 2023.",
                "venue": "CoRR, abs/2302.13971.",
                "url": "https://doi.org/10.48550/arXiv.2302.13971"
            }
        },
        {
            "60": {
                "title": "Multimodal Few-Shot\nLearning with Frozen Language Models.",
                "author": "Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals,\nand Felix Hill. 2021.",
                "venue": "arXiv:2106.13884 [cs].",
                "url": "http://arxiv.org/abs/2106.13884"
            }
        },
        {
            "61": {
                "title": "Attention is All you Need.",
                "author": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,\nAidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017.",
                "venue": "In Advances in Neural Information Processing Systems\n30: Annual Conference on Neural Information Processing Systems\n2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998\u20136008.",
                "url": "https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html"
            }
        },
        {
            "62": {
                "title": "Cider:\nConsensus-based image description evaluation.",
                "author": "Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. 2015.",
                "venue": "In IEEE Conference on Computer Vision and Pattern\nRecognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages\n4566\u20134575. IEEE Computer Society.",
                "url": "https://doi.org/10.1109/CVPR.2015.7299087"
            }
        },
        {
            "63": {
                "title": "Overcoming\nCatastrophic Forgetting in Zero-Shot Cross-Lingual Generation.",
                "author": "Tu Vu, Aditya Barua, Brian Lester, Daniel Cer, Mohit Iyyer, and Noah Constant.\n2022.",
                "venue": "In Proceedings of the 2022 Conference on Empirical\nMethods in Natural Language Processing, EMNLP 2022, Abu Dhabi,\nUnited Arab Emirates, December 7-11, 2022, pages 9279\u20139300.\nAssociation for Computational Linguistics.",
                "url": "https://aclanthology.org/2022.emnlp-main.630"
            }
        },
        {
            "64": {
                "title": "Image as a foreign\nlanguage: Beit pretraining for all vision and vision-language tasks.",
                "author": "Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti\nAggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei.\n2022.",
                "venue": "CoRR, abs/2208.10442.",
                "url": "https://doi.org/10.48550/arXiv.2208.10442"
            }
        },
        {
            "65": {
                "title": "SimVLM: Simple Visual\nLanguage Model Pretraining with Weak Supervision.",
                "author": "Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao.\n2021.",
                "venue": "arXiv:2108.10904 [cs].",
                "url": "http://arxiv.org/abs/2108.10904"
            }
        },
        {
            "66": {
                "title": "Transformers: State-of-the-art natural language processing.",
                "author": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue,\nAnthony Moi, Pierric Cistac, Tim Rault, R\u00e9mi Louf, Morgan Funtowicz, Joe\nDavison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien\nPlu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest,\nand Alexander M. Rush. 2020.",
                "venue": "In Proceedings of the 2020 Conference on Empirical Methods in\nNatural Language Processing: System Demonstrations, pages 38\u201345, Online.\nAssociation for Computational Linguistics.",
                "url": "https://www.aclweb.org/anthology/2020.emnlp-demos.6"
            }
        },
        {
            "67": {
                "title": "Visual Entailment: A\nNovel Task for Fine-Grained Image Understanding.",
                "author": "Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. 2019.",
                "venue": "arXiv preprint, abs/1901.06706.",
                "url": "http://arxiv.org/abs/1901.06706"
            }
        },
        {
            "68": {
                "title": "Baize: An\nOpen-Source Chat Model with Parameter-Efficient Tuning on\nSelf-Chat Data.",
                "author": "Canwen Xu, Daya Guo, Nan Duan, and Julian J. McAuley. 2023.",
                "venue": "CoRR, abs/2304.01196.",
                "url": "https://doi.org/10.48550/arXiv.2304.01196"
            }
        },
        {
            "69": {
                "title": "mT5: A\nMassively Multilingual Pre-trained Text-to-Text Transformer.",
                "author": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya\nSiddhant, Aditya Barua, and Colin Raffel. 2021.",
                "venue": "In Proceedings of the 2021 Conference of the North\nAmerican Chapter of the Association for Computational Linguistics:\nHuman Language Technologies, NAACL-HLT 2021, Online, June 6-11,\n2021, pages 483\u2013498. Association for Computational Linguistics.",
                "url": "https://doi.org/10.18653/v1/2021.naacl-main.41"
            }
        },
        {
            "70": {
                "title": "mPLUG-Owl:\nModularization Empowers Large Language Models with\nMultimodality.",
                "author": "Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang\nWang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong\nChen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. 2023.",
                "venue": "CoRR, abs/2304.14178.",
                "url": "https://doi.org/10.48550/arXiv.2304.14178"
            }
        },
        {
            "71": {
                "title": "Coca: Contrastive\ncaptioners are image-text foundation models.",
                "author": "Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and\nYonghui Wu. 2022.",
                "venue": "Transactions on Machine Learning Research, 2022.",
                "url": "https://openreview.net/forum?id=Ee277P3AYC"
            }
        },
        {
            "72": {
                "title": "X\\(^\\mbox2\\)-VLM:\nAll-In-One Pre-trained Model For Vision-Language Tasks.",
                "author": "Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu\nZhou. 2022a.",
                "venue": "CoRR, abs/2211.12402.",
                "url": "https://doi.org/10.48550/arXiv.2211.12402"
            }
        },
        {
            "73": {
                "title": "Cross-View\nLanguage Modeling: Towards Unified Cross-Lingual Cross-Modal\nPre-training.",
                "author": "Yan Zeng, Wangchunshu Zhou, Ao Luo, Ziming Cheng, and Xinsong Zhang. 2023.",
                "venue": "In Proceedings of the 61st Annual Meeting of the\nAssociation for Computational Linguistics (Volume 1: Long\nPapers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages\n5731\u20135746. Association for Computational Linguistics.",
                "url": "https://doi.org/10.18653/v1/2023.acl-long.315"
            }
        },
        {
            "74": {
                "title": "Cross-View\nLanguage Modeling: Towards Unified Cross-Lingual Cross-Modal\nPre-training.",
                "author": "Yan Zeng, Wangchunshu Zhou, Ao Luo, and Xinsong Zhang. 2022b.",
                "venue": "CoRR, abs/2206.00621.",
                "url": "https://doi.org/10.48550/arXiv.2206.00621"
            }
        },
        {
            "75": {
                "title": "Transfer Visual\nPrompt Generator across LLMs.",
                "author": "Ao Zhang, Hao Fei, Yuan Yao, Wei Ji, Li Li, Zhiyuan Liu, and Tat-Seng Chua.\n2023.",
                "venue": "CoRR, abs/2305.01278.",
                "url": "https://doi.org/10.48550/arXiv.2305.01278"
            }
        },
        {
            "76": {
                "title": "Opt: Open pre-trained transformer language models.",
                "author": "Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui\nChen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022.",
                "venue": "arXiv preprint arXiv:2205.01068.",
                "url": null
            }
        },
        {
            "77": {
                "title": "UC2:\nUniversal Cross-Lingual Cross-Modal Vision-and-Language\nPre-Training.",
                "author": "Mingyang Zhou, Luowei Zhou, Shuohang Wang, Yu Cheng, Linjie Li, Zhou Yu, and\nJingjing Liu. 2021.",
                "venue": "In IEEE Conference on Computer Vision and Pattern\nRecognition, CVPR 2021, virtual, June 19-25, 2021, pages 4155\u20134165.\nComputer Vision Foundation / IEEE.",
                "url": "https://doi.org/10.1109/CVPR46437.2021.00414"
            }
        },
        {
            "78": {
                "title": "MiniGPT-4:\nEnhancing Vision-Language Understanding with Advanced Large\nLanguage Models.",
                "author": "Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023.",
                "venue": "CoRR, abs/2304.10592.",
                "url": "https://doi.org/10.48550/arXiv.2304.10592"
            }
        }
    ],
    "url": "http://arxiv.org/html/2307.06930v3"
}