| Task | Input | Output | User History |
| LaMP-1 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please choose one of the following two references that is more relevant to the user's input title: [1] {reference1}; [2] {reference2}. Please just answer with “[1” or “[2” without explanation. “title”: {title}. | [1] | “title”: {title}\n“abstract”: {abstract} |
| LaMP-2 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please select the tag from [sci-fi, based on a book, comedy ... ] that is most relevant to the user's input description. Please just answer with the tag name without explanation. “description”: {description}; “tag”: | comedy | “description”: {description};\n“tag”: {tag} |
| LaMP-3 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, what is the score of the following review on a scale of 1 to 5? just answer with 1, 2, 3, 4, or 5 without further explanation. “review”: {review}; “score”: | 5 | “review”: {review}\n“score”: {score} |
| LaMP-4 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please generate a title for the given user's input text. Please generate it in the following format: {"title": "generated title"} without explanation, and use only English. “text”: {text}; “title”: | {“title”: Finding Happiness \nAfter Divorce - It Can Happen} | “text”: {text}\n“title”: {title} |
| LaMP-5 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please generate a title for the given user's input abstract. Please generate it in the following format: {"title": "generated title"} without explanation, and use only English. “abstract”: {abstract}; “title”: | {“title”: Link-Reliability Based \nTwo-Hop Routing for \nWireless Sensor Networks.} | “abstract”: {abstract}\n“title”: {title} |
| LaMP-7 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the style pattern of the historical tweets provided, please paraphrase the user's input tweet without any explanation before or after it. Please generate it in the following format: {"tweet": "generated tweet"} without explanation, and use only English. “tweet”: {tweet}. | {“tweet”:lilxcutiesworld the \ndanny picture is GOOD!! \nI really like it.} | “tweet”: {tweet} |
",
+ "bbox": [
+ 93,
+ 178,
+ 906,
+ 621
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "5.4.4 Impact of the Number of Retrieved Users. Since we enhance personalized text generation by introducing collaborative filtering, we further explored how much collaborative information to introduce, specifically the impact of the number of retrieved users on the results, as shown in Figure 8. In LaMP-1, retrieving too few or too many users leads to poorer performance, with the best results at 4 users. In LaMP-5, the performance improves as the number of users increases. This highlights the importance of introducing collaborative filtering, but it also indicates that excessive introduction can lead to decreased effectiveness.",
+ "bbox": [
+ 81,
+ 631,
+ 482,
+ 770
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "5.4.5 Impact of the Number of Retrieved Documents. We also analyzed the impact of the number of retrieved documents, $k$ , on the results, as shown in Figure 9. It can be observed that as the number of retrieved documents increases, performance improves, indicating the importance of retrieving user history to reflect user preferences for enhancing LLM-generated results. Since more documents lead to longer prompts and slower LLM generation, we chose $k = 5$ for our experiments.",
+ "bbox": [
+ 81,
+ 785,
+ 482,
+ 896
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "6 Conclusion",
+ "text_level": 1,
+ "bbox": [
+ 514,
+ 630,
+ 638,
+ 643
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "In this paper, we propose CFRAG, which adapts collaborative filtering into RAG to personalize LLMs. To introduce collaborative information without explicit user labels and retrieve documents that support personalized LLM generation, we first train user embeddings through contrastive learning to retrieve similar users. Then, we design the personalized retriever and reranker that considers user preferences during retrieval and fine-tune them using LLM feedback. The results on the Language Model Personalization (LaMP) benchmark validate the effectiveness of CFRAG. The experimental analysis also confirms the effectiveness of each module within CFRAG.",
+ "bbox": [
+ 511,
+ 648,
+ 915,
+ 800
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "A Appendix: Prompts",
+ "text_level": 1,
+ "bbox": [
+ 514,
+ 849,
+ 709,
+ 864
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "We provide detailed formats for the inputs, outputs, and user histories for the LLM across different datasets, as shown in Table 4.",
+ "bbox": [
+ 513,
+ 867,
+ 915,
+ 895
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "header",
+ "text": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation",
+ "bbox": [
+ 83,
+ 75,
+ 532,
+ 87
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "header",
+ "text": "SIGIR '25, July 13-18, 2025, Padua, Italy.",
+ "bbox": [
+ 718,
+ 75,
+ 913,
+ 87
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "References",
+ "text_level": 1,
+ "bbox": [
+ 84,
+ 104,
+ 176,
+ 119
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[1] AI@Meta. 2024. Llama 3 Model Card. (2024). https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md",
+ "[2] Akari Asai, Zegiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. [n.d.]. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. In The Twelfth International Conference on Learning Representations.",
+ "[3] Sebastian Borgeaud, Arthur Mensch, et al. 2022. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, PMLR, 2206-2240.",
+ "[4] Jin Chen, Zheng Liu, et al. 2024. When large language models meet personalization: Perspectives of challenges and opportunities. World Wide Web 27, 4 (2024), 42.",
+ "[5] Sunhao Dai, Ninglu Shao, et al. 2023. Uncovering chatgpt's capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems. 1126-1132.",
+ "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers).",
+ "[7] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 6491-6501.",
+ "[8] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023).",
+ "[9] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. In International conference on machine learning. PMLR, 3929-3938.",
+ "[10] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639-648.",
+ "[11] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web. 173-182.",
+ "[12] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. [n.d.]. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations.",
+ "[13] Gautier Izacard and Edouard Grave. 2021. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 874-880.",
+ "[14] Gautier Izacard, Patrick Lewis, et al. 2022. Few-shot learning with retrieval augmented language models. arXiv preprint arXiv:2208.03299 1, 2 (2022), 4.",
+ "[15] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. 2020. A survey on contrastive self-supervised learning. Technologies 9, 1 (2020), 2.",
+ "[16] Joel Jang, Seungone Kim, et al. 2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv preprint arXiv:2310.11564 (2023).",
+ "[17] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In International Conference on Machine Learning. PMLR, 15696-15707.",
+ "[18] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).",
+ "[19] Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer 42, 8 (2009), 30-37.",
+ "[20] Patrick Lewis, Ethan Perez, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33 (2020), 9459-9474.",
+ "[21] Cheng Li, Mingyang Zhang, Qiaozhu Mei, Yaqing Wang, Spurthi Amba Hombaiah, Yi Liang, and Michael Bendersky. 2023. Teach LLMs to Personalize-An Approach inspired by Writing Education. arXiv preprint arXiv:2308.07968 (2023).",
+ "[22] Junyi Li, Tianyi Tang, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2024. Pre-trained language models for text generation: A survey. Comput. Surveys 56, 9 (2024), 1-39.",
+ "[23] Xinyu Li, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133 (2024).",
+ "[24] Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out. 74-81.",
+ "[25] Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. [n.d.]. RA-DIT: Retrieval-Augmented Dual Instruction Tuning. In The Twelfth International Conference on Learning Representations.",
+ "[26] Yinhan Liu. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)."
+ ],
+ "bbox": [
+ 86,
+ 122,
+ 482,
+ 896
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[27] Sheshera Mysore, Zhuoran Lu, et al. 2023. Pearl: Personalizing large language model writing assistants with generation-calibrated retrievers. arXiv preprint arXiv:2311.09180 (2023).",
+ "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018).",
+ "[29] Chris Richardson, Yao Zhang, Kellen Gillespie, Sudipta Kar, Arshdeep Singh, Zeynab Raeesy, Omar Zia Khan, and Abhinav Sethy. 2023. Integrating summarization and retrieval for enhanced personalization via large language models. arXiv preprint arXiv:2310.20081 (2023).",
+ "[30] Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp 109 (1995), 109.",
+ "[31] Alireza Salemi, Surya Kallumadi, and Hamed Zamani. 2024. Optimization methods for personalizing large language models through retrieval augmentation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 752-762.",
+ "[32] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2023. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406 (2023).",
+ "[33] Chenglei Shen, Xiao Zhang, Teng Shi, Changshuo Zhang, Guofu Xie, and Jun Xu. 2024. A survey of controllable learning: Methods and applications in information retrieval. arXiv preprint arXiv:2407.06083 (2024).",
+ "[34] Teng Shi, Zihua Si, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Dewei Leng, Yanan Niu, and Yang Song. 2024. UniSAR: Modeling User Transition Behaviors between Search and Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1029-1039.",
+ "[35] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2024. REPLUG: Retrieval-Augmented Black-Box Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 8364-8377.",
+ "[36] Zhaoxuan Tan, Zheyuan Liu, and Meng Jiang. 2024. Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts. arXiv preprint arXiv:2406.10471 (2024).",
+ "[37] Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2024. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. arXiv:2402.04401 [cs.CL] https://arxiv.org/abs/2402.04401",
+ "[38] Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Wu Jian, and Yuning Jiang. 2025. Think Before Recommend: Unleashing the Latent Reasoning Power for Sequential Recommendation. arXiv:2503.22675 [cs.IR] https://arxiv.org/abs/2503.22675",
+ "[39] A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017).",
+ "[40] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval. 165-174.",
+ "[41] Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al. 2024. A survey on large language models for recommendation. World Wide Web 27, 5 (2024), 60.",
+ "[42] Xinghao Wu, Xuefeng Liu, Jianwei Niu, Haolin Wang, Shaojie Tang, and Guogang Zhu. 2024. FedLoRA: When Personalized Federated Learning Meets Low-Rank Adaptation. (2024).",
+ "[43] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. 2024. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems 36 (2024).",
+ "[44] Zhuofeng Wu, Sinong Wang, Jiatao Gu, Madian Khabsa, Fei Sun, and Hao Ma. 2020. Clear: Contrastive learning for sentence representation. arXiv preprint arXiv:2012.15466 (2020).",
+ "[45] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597 [cs.CL]",
+ "[46] Hong-Jian Xue, Xinyu Dai, Jianbing Zhang, Shujian Huang, and Jiajun Chen. 2017. Deep matrix factorization models for recommender systems.. In IJCAI, Vol. 17. Melbourne, Australia, 3203-3209.",
+ "[47] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng-peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671 (2024).",
+ "[48] Changshuo Zhang, Teng Shi, Xiao Zhang, Qi Liu, Ruobing Xie, Jun Xu, and Ji-Rong Wen. 2024. Modeling Domain and Feedback Transitions for Cross-Domain Sequential Recommendation. arXiv preprint arXiv:2408.08209 (2024).",
+ "[49] Changshuo Zhang, Teng Shi, Xiao Zhang, Yanping Zheng, Ruobing Xie, Qi Liu, Jun Xu, and Ji-Rong Wen. 2024. QAGCF: Graph Collaborative Filtering for Q&A Recommendation. arXiv preprint arXiv:2406.04828 (2024).",
+ "[50] Changshuo Zhang, Xiao Zhang, Teng Shi, Jun Xu, and Ji-Rong Wen. 2025. Test-Time Alignment for Tracking User Interest Shifts in Sequential Recommendation. arXiv:2504.01489 [cs.IR] https://arxiv.org/abs/2504.01489"
+ ],
+ "bbox": [
+ 516,
+ 108,
+ 911,
+ 893
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "header",
+ "text": "SIGIR '25, July 13-18, 2025, Padua, Italy.",
+ "bbox": [
+ 84,
+ 75,
+ 279,
+ 87
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "header",
+ "text": "Teng Shi et al.",
+ "bbox": [
+ 841,
+ 75,
+ 911,
+ 87
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[51] Kepu Zhang, Teng Shi, Sunhao Dai, Xiao Zhang, Yinfeng Li, Jing Lu, Xiaoxue Zang, Yang Song, and Jun Xu. 2024. SAQRec: Aligning Recommender Systems to User Satisfaction via Questionnaire Feedback. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 3165-3175.",
+ "[52] Xiao Zhang, Teng Shi, Jun Xu, Zhenhua Dong, and Ji-Rong Wen. 2024. Model-Agnostic Causal Embedding Learning for Counterfactually Group-Fair Recommendation. IEEE Transactions on Knowledge and Data Engineering (2024).",
+ "[53] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219 (2023)."
+ ],
+ "bbox": [
+ 84,
+ 108,
+ 483,
+ 220
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[54] Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2024. Dense text retrieval based on pretrained language models: A survey. ACM Transactions on Information Systems 42, 4 (2024), 1-60.",
+ "[55] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223 (2023).",
+ "[56] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107 (2023).",
+ "[57] Yuchen Zhuang, Haotian Sun, Yue Yu, Qifan Wang, Chao Zhang, and Bo Dai. 2024. HYDRA: Model Factorization Framework for Black-Box LLM Personalization. arXiv preprint arXiv:2406.02888 (2024)."
+ ],
+ "bbox": [
+ 516,
+ 108,
+ 913,
+ 229
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "header",
+ "text": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation",
+ "bbox": [
+ 84,
+ 75,
+ 532,
+ 87
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "header",
+ "text": "SIGIR '25, July 13-18, 2025, Padua, Italy.",
+ "bbox": [
+ 718,
+ 75,
+ 911,
+ 87
+ ],
+ "page_idx": 10
+ }
+]
\ No newline at end of file
diff --git a/data/2025/2504_05xxx/2504.05731/1d2e0a8a-e48a-4cb6-8ec0-35774101eda9_model.json b/data/2025/2504_05xxx/2504.05731/1d2e0a8a-e48a-4cb6-8ec0-35774101eda9_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..04855b23b8abd6965a2cd8214ee8d75c17c2b331
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/1d2e0a8a-e48a-4cb6-8ec0-35774101eda9_model.json
@@ -0,0 +1,3104 @@
+[
+ [
+ {
+ "type": "aside_text",
+ "bbox": [
+ 0.023,
+ 0.28,
+ 0.061,
+ 0.701
+ ],
+ "angle": 270,
+ "content": "arXiv:2504.05731v1 [cs.IR] 8 Apr 2025"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.086,
+ 0.101,
+ 0.915,
+ 0.15
+ ],
+ "angle": 0,
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.26,
+ 0.158,
+ 0.335,
+ 0.174
+ ],
+ "angle": 0,
+ "content": "Teng Shi"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.204,
+ 0.176,
+ 0.394,
+ 0.19
+ ],
+ "angle": 0,
+ "content": "Renmin University of China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.25,
+ 0.191,
+ 0.347,
+ 0.205
+ ],
+ "angle": 0,
+ "content": "Beijing, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.231,
+ 0.206,
+ 0.366,
+ 0.22
+ ],
+ "angle": 0,
+ "content": "shiteng@ruc.edu.cn"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.24,
+ 0.249,
+ 0.356,
+ 0.266
+ ],
+ "angle": 0,
+ "content": "Xiaoxue Zang"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.255,
+ 0.267,
+ 0.341,
+ 0.284
+ ],
+ "angle": 0,
+ "content": "Kai Zheng"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.196,
+ 0.284,
+ 0.402,
+ 0.298
+ ],
+ "angle": 0,
+ "content": "Kuaishou Technology Co., Ltd."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.249,
+ 0.299,
+ 0.346,
+ 0.313
+ ],
+ "angle": 0,
+ "content": "Beijing, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.238,
+ 0.314,
+ 0.359,
+ 0.327
+ ],
+ "angle": 0,
+ "content": "xxic666@126.com"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.223,
+ 0.329,
+ 0.373,
+ 0.344
+ ],
+ "angle": 0,
+ "content": "zhengk92@gmail.com"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.084,
+ 0.353,
+ 0.158,
+ 0.367
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.371,
+ 0.483,
+ 0.718
+ ],
+ "angle": 0,
+ "content": "Recently, the personalization of Large Language Models (LLMs) to generate content that aligns with individual user preferences has garnered widespread attention. Personalized Retrieval-Augmented Generation (RAG), which retrieves relevant documents from the user's history to reflect their preferences and enhance LLM generation, is one commonly used approach for personalization. However, existing personalized RAG methods do not consider that the histories of similar users can also assist in personalized generation for the current user, meaning that collaborative information between users can also benefit personalized generation. Inspired by the application of collaborative filtering in recommender systems, we propose a method called CFRAG, which adapts Collaborative Filtering to RAG for personalized text generation. However, this presents two challenges: (1) how to incorporate collaborative information without explicit user similarity labels? (2) how to retrieve documents that support personalized LLM generation? For Challenge 1, we use contrastive learning to train user embeddings to retrieve similar users and introduce collaborative information. For Challenge 2, we design a personalized retriever and reranker to retrieve the top-\\(k\\) documents from these users' histories. We take into account the user's preference during retrieval and reranking. Then we leverage feedback from the LLM to fine-tune the personalized retriever and reranker, enabling them to retrieve documents that meet the personalized generation needs of the LLM. Experimental results on the Language Model Personalization (LaMP) benchmark"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.732,
+ 0.483,
+ 0.765
+ ],
+ "angle": 0,
+ "content": "*Corresponding authors. Work partially done at Engineering Research Center of Next-Generation Intelligent Search and Recommendation, Ministry of Education. \nWork done when Teng Shi was the intern at Kuaishou."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.78,
+ 0.483,
+ 0.853
+ ],
+ "angle": 0,
+ "content": "Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.086,
+ 0.854,
+ 0.192,
+ 0.864
+ ],
+ "angle": 0,
+ "content": "SIGIR '25, Padua, Italy."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.085,
+ 0.865,
+ 0.473,
+ 0.885
+ ],
+ "angle": 0,
+ "content": "© 2025 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 979-8-4007-1592-1/25/07"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.086,
+ 0.885,
+ 0.29,
+ 0.896
+ ],
+ "angle": 0,
+ "content": "https://doi.org/10.1145/XXXXXX.XXXXXXX"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.667,
+ 0.158,
+ 0.732,
+ 0.174
+ ],
+ "angle": 0,
+ "content": "Jun Xu*"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.653,
+ 0.176,
+ 0.749,
+ 0.193
+ ],
+ "angle": 0,
+ "content": "Xiao Zhang"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.607,
+ 0.193,
+ 0.796,
+ 0.207
+ ],
+ "angle": 0,
+ "content": "Renmin University of China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.653,
+ 0.208,
+ 0.75,
+ 0.222
+ ],
+ "angle": 0,
+ "content": "Beijing, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.6,
+ 0.223,
+ 0.801,
+ 0.237
+ ],
+ "angle": 0,
+ "content": "{junxu,zhangx89}@ruc.edu.cn"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.657,
+ 0.249,
+ 0.744,
+ 0.266
+ ],
+ "angle": 0,
+ "content": "Yang Song"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.671,
+ 0.268,
+ 0.729,
+ 0.281
+ ],
+ "angle": 0,
+ "content": "Han Li"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.598,
+ 0.284,
+ 0.804,
+ 0.298
+ ],
+ "angle": 0,
+ "content": "Kuaishou Technology Co., Ltd."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.652,
+ 0.299,
+ 0.749,
+ 0.313
+ ],
+ "angle": 0,
+ "content": "Beijing, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.652,
+ 0.314,
+ 0.749,
+ 0.328
+ ],
+ "angle": 0,
+ "content": "ys@sonyis.me"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.329,
+ 0.78,
+ 0.342
+ ],
+ "angle": 0,
+ "content": "lihan08@kuaishou.com"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.354,
+ 0.914,
+ 0.383
+ ],
+ "angle": 0,
+ "content": "validate the effectiveness of CFRAG. Further analysis confirms the importance of incorporating collaborative information."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.395,
+ 0.634,
+ 0.411
+ ],
+ "angle": 0,
+ "content": "CCS Concepts"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.414,
+ 0.928,
+ 0.442
+ ],
+ "angle": 0,
+ "content": "- Information systems \\(\\rightarrow\\) Personalization; - Computing methodologies \\(\\rightarrow\\) Natural language generation."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.455,
+ 0.6,
+ 0.47
+ ],
+ "angle": 0,
+ "content": "Keywords"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.473,
+ 0.915,
+ 0.5
+ ],
+ "angle": 0,
+ "content": "Large language model; Personalization; Retrieval augmented generation"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.509,
+ 0.662,
+ 0.52
+ ],
+ "angle": 0,
+ "content": "ACM Reference Format:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.522,
+ 0.915,
+ 0.597
+ ],
+ "angle": 0,
+ "content": "Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Han Li. 2025. Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '25), July 13-18, 2025, Padua, Italy. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/XXXXXX.XXXXXXX"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.615,
+ 0.651,
+ 0.629
+ ],
+ "angle": 0,
+ "content": "1 Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.633,
+ 0.915,
+ 0.716
+ ],
+ "angle": 0,
+ "content": "Personalizing Large Language Models (LLMs) [55] to generate personalized outputs tailored to individual user preferences has emerged as a significant and rapidly growing field [16, 23, 29, 31, 32, 36, 37, 57]. Personalized Retrieval-Augmented Generation (RAG) [8] has become a commonly used approach for personalizing LLMs [29, 31, 32, 57]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.717,
+ 0.915,
+ 0.897
+ ],
+ "angle": 0,
+ "content": "The process of existing personalized RAG methods typically involves retrieving similar documents from the user's historical behaviors based on the user's input query, then concatenating these documents with the query as a prompt input to the LLM for generation. Although effective, this approach is limited to retrieving only the current user's history, neglecting collaborative information. Users with similar histories tend to be more alike, and the information from these similar users can also aid in personalizing generation for the current user. As shown in the example in Figure 1, the upper part illustrates the results of the existing RAG method, which retrieves documents from the current user's history. We can only infer from these results that \"She\" in the user's input refers to \"Hillary Clinton\". In contrast, the lower part demonstrates"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.085,
+ 0.076,
+ 0.28,
+ 0.087
+ ],
+ "angle": 0,
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.843,
+ 0.076,
+ 0.913,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "Teng Shi et al."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.107,
+ 0.104,
+ 0.885,
+ 0.349
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.082,
+ 0.352,
+ 0.915,
+ 0.422
+ ],
+ "angle": 0,
+ "content": "Figure 1: An example from the LaMP-4 dataset [32]. The task of LaMP-4 is to generate personalized news headlines based on user input. This example illustrates the benefit of collaborative information for LLM personalization: (a) The top shows results retrieved by the existing RAG method from the current user's history, where we can only infer that \"She\" in the user's input refers to \"Hillary Clinton\". (b) The bottom shows results retrieved by our method from similar users' histories, allowing us to infer further that \"his\" in the user's input refers to \"Donald Trump\" thus enabling the generation of a more accurate result."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.432,
+ 0.483,
+ 0.515
+ ],
+ "angle": 0,
+ "content": "our method, which retrieves documents from the history of similar users. In this case, we can further infer that \"his\" in the user's input refers to \"Donald Trump\", leading to a better generation result. From this example, we can see that incorporating collaborative information allows the retrieval of more diverse documents, helping the LLM generate results that better meet the user's needs."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.515,
+ 0.483,
+ 0.709
+ ],
+ "angle": 0,
+ "content": "Inspired by the application of collaborative filtering in recommender systems [11, 40, 46], we propose to adapt collaborative information into RAG to personalize LLMs. However, adapting collaborative filtering to personalized RAG presents two challenges. Challenge 1: How to incorporate collaborative information. Without explicit labels indicating which users are similar, which users' information should be selected to help personalize generation for the current user? Challenge 2: How to retrieve documents that support personalized LLM generation, rather than relying on traditional semantic relevance? Pre-trained dense retrieval models [54] only retrieve based on the semantic relevance between the query and document. Directly using these models for retrieval may not necessarily result in content that allows the LLM to generate outputs that meet the user's needs [25, 35]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.709,
+ 0.483,
+ 0.89
+ ],
+ "angle": 0,
+ "content": "To address the above challenges, this paper proposes a method named CFRAG which adapts Collaborative Filtering to personalized Retrieval Augmented Generation. Firstly, to address Challenge 1, since there are no explicit user similarity labels, we use contrastive learning [15, 44] to train user embeddings for retrieving similar users to introduce collaborative information. Specifically, we apply different data augmentation methods to the user's history to obtain different views, and then treat different views of the same user's history as positive samples for each other. Then we use contrastive learning on different views to train the user embeddings. Secondly, for Challenge 2, we designed a personalized retriever and reranker to retrieve the top-\\(k\\) documents from the histories of the retrieved users. In both retrieval and reranking, in addition to the semantic"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.432,
+ 0.915,
+ 0.543
+ ],
+ "angle": 0,
+ "content": "relevance between the query and documents, we also considered the user's preferences for different documents to enable personalized retrieval. Additionally, we further fine-tune the retriever and reranker based on the feedback from the LLM to ensure that the retrieved documents better support the personalized LLM generation. Finally, the top-\\(k\\) documents are concatenated with the user's input query to form a prompt, which is then fed into the LLM for personalized generation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.543,
+ 0.915,
+ 0.613
+ ],
+ "angle": 0,
+ "content": "The major contributions of the paper are summarized as follows: We analyzed the necessity of introducing collaborative filtering into RAG for LLM personalization and identified the challenges: how to introduce collaborative information and how to retrieve documents that support personalized LLM generation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.612,
+ 0.915,
+ 0.681
+ ],
+ "angle": 0,
+ "content": "- We proposed a method called CFRAG, which uses contrastive learning to train user embeddings for retrieving similar users and incorporating collaborative information. It leverages LLM feedback to train the personalized retriever and reranker, enabling them to retrieve documents that support personalized LLM generation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.681,
+ 0.915,
+ 0.737
+ ],
+ "angle": 0,
+ "content": "- Experimental results on the Language Model Personalization (LaMP) [32] benchmark validate the effectiveness of CFRAG. The experimental analysis also demonstrates the importance of leveraging collaborative information."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.514,
+ 0.612,
+ 0.915,
+ 0.737
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.763,
+ 0.661,
+ 0.777
+ ],
+ "angle": 0,
+ "content": "2 Related Work"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.786,
+ 0.915,
+ 0.897
+ ],
+ "angle": 0,
+ "content": "Personalization of LLMs. Large Language Models (LLMs) [55] have demonstrated remarkable capabilities in various fields, such as text generation [22], information retrieval [56], recommender systems [5, 41], and so on. However, since LLMs are typically designed to serve all tasks with a single model and are trained on broad, domain-agnostic data, they face challenges in adapting to the personalized needs of individual users [4, 32]. Therefore, LLM personalization has attracted widespread attention [16, 31, 57]."
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.084,
+ 0.076,
+ 0.533,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.719,
+ 0.076,
+ 0.914,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.11,
+ 0.106,
+ 0.881,
+ 0.337
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.081,
+ 0.339,
+ 0.916,
+ 0.395
+ ],
+ "angle": 0,
+ "content": "Figure 2: The architecture of CFRAG. From left to right: (a) User Retrieval retrieves similar users (Section 4.1); (b) Retriever retrieves the top- \\(k\\) documents from each user's history (Section 4.2); (c) Reranker reranks the \\(m \\times k\\) documents to get the final top- \\(k\\) documents, which are then concatenated with the query and input into the LLM for personalized text generation (Section 4.3)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.405,
+ 0.484,
+ 0.599
+ ],
+ "angle": 0,
+ "content": "Existing works on LLM personalization mainly include the following types of methods: (1) Fine-tuning a personalized LLM for each user [36, 37, 42]; Tan et al. [37] fine-tuned the LLM using LoRA [12] to get personalized LoRA parameters for each user. (2) Aligning LLMs with user-specific preferences through Reinforcement Learning from Human Feedback (RLHF) [16, 23, 43]; Jang et al. [16] first trained different parameters for various objectives using RLHF, then merged these parameters based on users' personalized needs. (3) Incorporating user-specific context into the prompt [21, 27, 29, 31, 32, 57]. Richardson et al. [29] used instruction-tuned LLMs to summarize user history and then incorporated it into prompts for generation. Salemi et al. [31, 32] used RAG to retrieve relevant documents from user history based on the input query and incorporated them into the prompt."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.599,
+ 0.484,
+ 0.765
+ ],
+ "angle": 0,
+ "content": "This paper further introduces collaborative filtering for personalization based on the RAG framework. Collaborative filtering has already been applied in fields such as recommender systems [33, 34, 38, 48-52] and has been proven effective. It assumes that users who have interacted with similar items share similar preferences, and recommending items from similar users to the current user can meet their needs. Some works [11, 46] learn the collaborative information between users and items through matrix factorization [19], while others [10, 40] further explore higher-order collaborative information between users and items using graph neural networks. The application of collaborative filtering in LLM personalization remains under-explored."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.768,
+ 0.484,
+ 0.894
+ ],
+ "angle": 0,
+ "content": "Retrieval Augmented Generation. Retrieval Augmented Generation [7, 8] introduces external knowledge through document retrieval, alleviating issues such as LLM hallucinations [53], and enhancing LLMs' capabilities in knowledge-intensive tasks [17] such as open-domain question answering [14, 20]. Some works [3, 13] encode retrieved documents using separate encoders, and then fuse the results with the language model using cross-attention. A more common approach is to directly include the retrieved documents in the prompt of the LLM [2, 9, 20, 25, 35]. In recent years, this"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.405,
+ 0.918,
+ 0.462
+ ],
+ "angle": 0,
+ "content": "in-context RAG framework has also been applied to LLM personalization, which is personalized by retrieving documents from the user's history [31, 32, 57]. This paper introduces collaborative filtering by retrieving similar users' histories for better personalization."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.472,
+ 0.723,
+ 0.487
+ ],
+ "angle": 0,
+ "content": "3 Problem Formulation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.491,
+ 0.916,
+ 0.606
+ ],
+ "angle": 0,
+ "content": "Let \\(\\mathcal{U} = \\{u_1, u_2, \\ldots, u_M\\}\\) denotes the set of all users, where \\(M\\) is the number of users. Each user \\(u \\in \\mathcal{U}\\) has a chronologically ordered history \\(\\mathcal{H}_u = [d_1, d_2, \\ldots, d_N]\\) which includes all her historical documents, where \\(N\\) is the number of documents in the history. The personalized text generation dataset is \\(\\mathcal{D} = \\{(u, q, y)_i\\}_{i=1}^{|D|}\\). For each instance, \\(q\\) is the query input by the user \\(u\\) to the LLM, and \\(y\\) is the target output. Our goal is first to introduce collaborative information by retrieving the top-\\(m\\) most similar users for user \\(u\\):"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.625,
+ 0.611,
+ 0.804,
+ 0.626
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {U} _ {\\text {r e t r i e v e d}} = \\left\\{u _ {1}, u _ {2}, \\dots , u _ {m} \\right\\}.\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.632,
+ 0.914,
+ 0.659
+ ],
+ "angle": 0,
+ "content": "Then, we use a retriever to retrieve the top-\\(k\\) documents from each of the \\(m\\) users' histories, resulting in a total of \\(m \\times k\\) documents."
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.57,
+ 0.666,
+ 0.858,
+ 0.681
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {D} _ {\\text {r e t r i e v e d}} = \\{d _ {i, j} | i \\in \\{1, \\dots , m \\}, j \\in \\{1, \\dots , k \\} \\}.\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.686,
+ 0.914,
+ 0.714
+ ],
+ "angle": 0,
+ "content": "Finally, we use a reranker to rerank these \\(m \\times k\\) documents and obtain the final top-\\(k\\) documents:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.619,
+ 0.72,
+ 0.81,
+ 0.735
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {D} _ {\\text {r e r a n k e d}} = \\left\\{d _ {i} | i \\in \\{1, \\dots , k \\} \\right\\}.\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.74,
+ 0.913,
+ 0.782
+ ],
+ "angle": 0,
+ "content": "These top-\\(k\\) documents will be concatenated with the user's query \\(q\\) as a prompt and input into the LLM, enabling it to generate a response that aligns with the target output \\(y\\)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.782,
+ 0.914,
+ 0.839
+ ],
+ "angle": 0,
+ "content": "This paper primarily focuses on how to retrieve \\(\\mathcal{U}_{\\mathrm{retrieved}}\\) to introduce collaborative information, and how to train the retriever and reranker so that they can effectively retrieve documents that support the personalized LLM generation."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.85,
+ 0.664,
+ 0.866
+ ],
+ "angle": 0,
+ "content": "4 Our Approach"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.868,
+ 0.916,
+ 0.897
+ ],
+ "angle": 0,
+ "content": "This section introduces our method CFRAG. CFRAG's overall architecture is shown in Figure 2. As mentioned in Section 1, to address"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.084,
+ 0.076,
+ 0.28,
+ 0.087
+ ],
+ "angle": 0,
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.843,
+ 0.076,
+ 0.913,
+ 0.087
+ ],
+ "angle": 0,
+ "content": "Teng Shi et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.107,
+ 0.484,
+ 0.246
+ ],
+ "angle": 0,
+ "content": "Challenge 1, i.e., how to introduce collaborative information, we first train user embeddings using contrastive learning to retrieve the top-\\(m\\) most similar users (see Section 4.1). For Challenge 2, which involves retrieving documents that support personalized LLM generation, we fine-tune the personalized retriever and reranker using LLM feedback. The retriever first retrieves the top-\\(k\\) documents from the history of each of the \\(m\\) users, resulting in \\(m \\times k\\) documents (see Section 4.2). The reranker then reranks these documents to obtain the final top-\\(k\\) documents as input for the LLM (see Section 4.3)."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.084,
+ 0.259,
+ 0.245,
+ 0.273
+ ],
+ "angle": 0,
+ "content": "4.1 User Retrieval"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.278,
+ 0.484,
+ 0.418
+ ],
+ "angle": 0,
+ "content": "First, we perform user retrieval to get the top-\\(m\\) most similar users for user \\(u\\) to introduce collaborative information. However, we do not have labels indicating which users are similar to each other. To address this, we employ a contrastive learning [15, 44] approach. We apply different data augmentation methods to the user history \\(\\mathcal{H}_u\\) to obtain different views of the user's history. We treat different views of the same user as positive samples and the histories of other users as negative samples, and then we use the InfoNCE [28] loss to train user embeddings for retrieval. Figure 3 illustrates the process of training user embeddings using contrastive learning."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.425,
+ 0.484,
+ 0.567
+ ],
+ "angle": 0,
+ "content": "4.1.1 User Encoder. Specifically, we first use an embedding model (such as BERT [6], RoBERTa [26], BGE [45] etc.) \\(\\mathbf{Emb}(\\cdot)\\) to encode each document in the user's history \\(\\mathcal{H}_u\\) to obtain \\(\\mathbf{E}_u = [\\mathbf{e}_1,\\mathbf{e}_2,\\dots ,\\mathbf{e}_N]^{\\intercal}\\in \\mathbb{R}^{N\\times d}\\), where \\(\\mathbf{e}_i = \\mathbf{Emb}(d_i)\\) and \\(d\\) is the embedding dimension. To model the sequential relationships between different documents in the user's history, we introduce positional embedding \\(\\mathbf{P}\\in \\mathbb{R}^{N\\times d}\\). Afterward, the history \\(\\mathcal{H}_u\\) 's embedding becomes \\(\\widehat{\\mathbf{E}}_u = \\mathbf{E}_u + \\mathbf{P}\\). Then, we apply a transformer [39] as the user encoder to encode the user's history \\(\\widehat{\\mathbf{E}}_u\\) and average the transformer's output to obtain the user's embedding:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.146,
+ 0.575,
+ 0.482,
+ 0.593
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathbf {e} _ {u} = \\operatorname {E n c o d e r} _ {u} (u) = \\operatorname {M E A N} (\\operatorname {T r m} (\\widehat {\\mathbf {F}} _ {u})) \\in \\mathbb {R} ^ {d}, \\tag {1}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.599,
+ 0.482,
+ 0.642
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathrm{Encoder}_u(\\cdot)\\to \\mathbb{R}^d\\) denotes the user encoder, \\(\\mathrm{Trm}(\\cdot)\\) denotes a transformer encoder. Next, we train the transformer encoder using contrastive learning."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.084,
+ 0.65,
+ 0.482,
+ 0.678
+ ],
+ "angle": 0,
+ "content": "4.1.2 Data Augmentation. We generate different views of \\(\\mathcal{H}_u\\) using the following three data augmentation methods:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.084,
+ 0.678,
+ 0.485,
+ 0.72
+ ],
+ "angle": 0,
+ "content": "Document Crop. We randomly select a continuous sub-sequence of length \\( L_{c} = \\lfloor \\eta_{c}N\\rfloor \\) from \\( \\mathcal{H}_u \\), where \\( \\eta_c \\) is a hyper-parameter controlling the crop ratio. The history after cropping is as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.185,
+ 0.726,
+ 0.38,
+ 0.745
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {H} _ {u} ^ {\\mathrm {c r o p}} = [ d _ {c}, d _ {c + 1}, \\dots , d _ {c + L _ {c} - 1} ].\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.751,
+ 0.484,
+ 0.835
+ ],
+ "angle": 0,
+ "content": "Document Mask. For the history \\(\\mathcal{H}_u\\), we randomly mask out \\(L_{m} = \\lfloor \\eta_{m}N\\rfloor\\) documents \\(\\mathcal{I}_{\\mathrm{mask}} = \\{i_1,i_2,\\dots ,i_{L_m}\\}\\), where \\(\\mathcal{I}_{\\mathrm{mask}}\\) is the set of indices corresponding to the masked documents and \\(\\eta_{m}\\) is a hyper-parameter that controls the mask ratio. The masked documents are replaced with a special token [mask]. The history after masking is as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.184,
+ 0.841,
+ 0.377,
+ 0.899
+ ],
+ "angle": 0,
+ "content": "\\[\n\\begin{array}{l} \\mathcal {H} _ {u} ^ {\\text {m a s k}} = \\left[ \\hat {d} _ {1}, \\hat {d} _ {2}, \\dots , \\hat {d} _ {N} \\right], \\\\ \\hat {d} _ {i} = \\left\\{ \\begin{array}{l l} d _ {i}, & i \\notin \\mathcal {I} _ {\\text {m a s k}}, \\\\ [ \\text {m a s k} ], & i \\in \\mathcal {I} _ {\\text {m a s k}}. \\end{array} \\right. \\\\ \\end{array}\n\\]"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.524,
+ 0.105,
+ 0.905,
+ 0.245
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.515,
+ 0.249,
+ 0.913,
+ 0.264
+ ],
+ "angle": 0,
+ "content": "Figure 3: Contrastive learning for user embedding training."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.278,
+ 0.915,
+ 0.36
+ ],
+ "angle": 0,
+ "content": "Document Reorder. We randomly select a sub-sequence \\([d_r, d_{r+1}, \\ldots, d_{r+L_r-1}]\\) of length \\(L_r = \\lfloor \\eta_r N \\rfloor\\) from \\(\\mathcal{H}_u\\), where \\(\\eta_r\\) is a hyper-parameter controlling the reorder ratio, and then randomly shuffle the order of the documents within the sub-sequence to obtain \\([\\hat{d}_r, \\hat{d}_{r+1}, \\ldots, \\hat{d}_{r+L_r-1}]\\). The history after reordering is as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.571,
+ 0.366,
+ 0.856,
+ 0.384
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {H} _ {u} ^ {\\text {r e o r d e r}} = \\left[ d _ {1}, d _ {2}, \\dots , \\hat {d} _ {r}, \\dots , \\hat {d} _ {r + L _ {r} - 1}, \\dots , d _ {N} \\right].\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.389,
+ 0.915,
+ 0.528
+ ],
+ "angle": 0,
+ "content": "4.1.3 Contrastive Loss. Each time, we randomly select two data augmentation methods \\(\\mathcal{A}'\\) and \\(\\mathcal{A}''\\) to generate two different views of \\(\\mathcal{H}_u\\), denoted as \\(\\mathcal{H}_u'\\) and \\(\\mathcal{H}_u''\\). Then, using the encoder described in Section 4.1.1, we obtain the user embeddings \\(\\mathbf{e}_u'\\) and \\(\\mathbf{e}_u''\\) corresponding to the different views. Since \\(\\mathbf{e}_u'\\) and \\(\\mathbf{e}_u''\\) are obtained through data augmentation of \\(\\mathcal{H}_u\\), they are more similar to each other. Therefore, we treat them as positive samples for each other and use the views generated from the augmented histories of other users in the same batch as negative samples. We then perform contrastive learning using the InfoNCE [28] loss as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.565,
+ 0.534,
+ 0.913,
+ 0.609
+ ],
+ "angle": 0,
+ "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C L}} = - \\left[ \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {e} _ {u} ^ {\\prime} , \\mathbf {e} _ {u} ^ {\\prime \\prime}\\right) / \\tau_ {1}\\right)}{\\sum_ {u ^ {-} \\in \\mathcal {U} _ {\\mathrm {n e g}}} \\exp \\left(\\cos \\left(\\mathbf {e} _ {u} ^ {\\prime} , \\mathbf {e} _ {u ^ {-}} ^ {\\prime \\prime}\\right) / \\tau_ {1}\\right)} \\right. \\tag {2} \\\\ \\left. + \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {e} _ {u} ^ {\\prime} , \\mathbf {e} _ {u} ^ {\\prime \\prime}\\right) / \\tau_ {1}\\right)}{\\sum_ {u ^ {-} \\in \\mathcal {U} _ {\\text {n e g}}} \\exp \\left(\\cos \\left(\\mathbf {e} _ {u ^ {-}} ^ {\\prime}, \\mathbf {e} _ {u} ^ {\\prime \\prime}\\right) / \\tau_ {1}\\right)} \\right], \\\\ \\end{array}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.614,
+ 0.916,
+ 0.656
+ ],
+ "angle": 0,
+ "content": "where \\(\\tau_{1}\\) is the temperature coefficient, \\(\\mathcal{U}_{\\mathrm{neg}}\\) are the set of randomly sampled in-batch negative samples, and \\(\\cos (\\cdot)\\) denotes the cosine similarity."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.663,
+ 0.916,
+ 0.746
+ ],
+ "angle": 0,
+ "content": "4.1.4 Top-\\(m\\) User Retrieval. After training with contrastive learning, we can use the encoder from Section 4.1.1 to obtain the user embedding \\(\\mathbf{e}_u\\). We then calculate the cosine similarity between each pair of user embeddings and retrieve the top-\\(m\\) most similar users \\(\\mathcal{U}_{\\mathrm{retrieved}} = \\{u_1, u_2, \\dots, u_m\\}\\) for user \\(u\\). Subsequently, the histories of these \\(m\\) users will be used for further document retrieval."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.759,
+ 0.724,
+ 0.773
+ ],
+ "angle": 0,
+ "content": "4.2 Document Retrieval"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.777,
+ 0.916,
+ 0.862
+ ],
+ "angle": 0,
+ "content": "After retrieving the top-\\(m\\) users, we design a personalized retriever to retrieve the top-\\(k\\) documents from each user's history, resulting in a total of \\(m \\times k\\) candidate documents \\(\\mathcal{D}_{\\text{retrieved}} = \\{d_{i,j} | i \\in \\{1, \\ldots, m\\}, j \\in \\{1, \\ldots, k\\}\\}\\). This section introduces how the retriever is designed and how it's trained to retrieve documents that better align with the requirements of personalized LLM generation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.868,
+ 0.915,
+ 0.897
+ ],
+ "angle": 0,
+ "content": "4.2.1 Retriever. First, we use a pre-trained dense retrieval model (such as BGE retriever [45]) to compute the semantic relevance"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.084,
+ 0.076,
+ 0.533,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.719,
+ 0.076,
+ 0.913,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.091,
+ 0.105,
+ 0.475,
+ 0.232
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.083,
+ 0.234,
+ 0.483,
+ 0.263
+ ],
+ "angle": 0,
+ "content": "Figure 4: The method of training the retriever and reranker using LLM feedback."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.083,
+ 0.281,
+ 0.384,
+ 0.295
+ ],
+ "angle": 0,
+ "content": "between the query and the candidate documents:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.15,
+ 0.303,
+ 0.483,
+ 0.323
+ ],
+ "angle": 0,
+ "content": "\\[\nS _ {q, d} ^ {\\text {r e t r i e v e r}} = \\cos \\left(\\operatorname {E n c o d e r} _ {q} (q), \\operatorname {E n c o d e r} _ {d} (d)\\right), \\tag {3}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.331,
+ 0.485,
+ 0.463
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathrm{Encoder}_q(\\cdot)\\to \\mathbb{R}^d\\) and \\(\\mathrm{Encoder}_d(\\cdot)\\rightarrow \\mathbb{R}^d\\) are the encoders for the query and the document in the retrieval model, respectively. Pre-trained retrieval models typically use \\(S_{q,d}^{\\mathrm{retriever}}\\) directly for retrieval. However, \\(S_{q,d}^{\\mathrm{retriever}}\\) only considers the semantic relevance between the query and the document. Since different users might input the same query but expect different outputs due to their varying preferences, we further account for user personalization by calculating the preference score of the user for the document as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.158,
+ 0.471,
+ 0.483,
+ 0.49
+ ],
+ "angle": 0,
+ "content": "\\[\nS _ {u, d} ^ {\\text {r e t r i e v e r}} = \\cos \\left(\\mathrm {M L P} _ {1} \\left(\\mathbf {e} _ {u}\\right), \\operatorname {E n c o d e r} _ {d} (d)\\right), \\tag {4}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.498,
+ 0.483,
+ 0.556
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathrm{MLP}_1: \\mathbb{R}^d \\to \\mathbb{R}^d\\) is a multi-layer perceptron that maps the user embedding to the space where the cosine similarity is computed. \\(\\mathbf{e}_u\\) is the embedding obtained in Section 4.1.1. The total score for retrieval is computed as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.161,
+ 0.562,
+ 0.483,
+ 0.583
+ ],
+ "angle": 0,
+ "content": "\\[\nS _ {u, q, d} ^ {\\text {r e t r i e v e r}} = (1 - \\alpha) S _ {q, d} ^ {\\text {r e t r i e v e r}} + \\alpha S _ {u, d} ^ {\\text {r e t r i e v e r}}, \\tag {5}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.589,
+ 0.485,
+ 0.618
+ ],
+ "angle": 0,
+ "content": "where \\(\\alpha\\) is a hyper-parameter that controls the weight of personalization."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.626,
+ 0.483,
+ 0.736
+ ],
+ "angle": 0,
+ "content": "4.2.2 Training. Since the pre-trained dense retrieval model is not fine-tuned for our specific task, the retrieved results may not necessarily lead to LLM responses that better match the target output \\( y \\) [25, 35]. However, there is no ground truth indicating which documents are better. Therefore, we evaluate the difference between the LLM's output and the target output \\( y \\), using this as a label to train the retrieval model. Figure 4 shows the process of training the retriever using LLM feedback."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.737,
+ 0.483,
+ 0.822
+ ],
+ "angle": 0,
+ "content": "Specifically, we first use the pre-trained retrieval model to retrieve the top-\\(k\\) documents from each of the \\(m\\) users' histories based on \\(S_{q,d}^{\\mathrm{retriever}}\\) in Eq. (3), resulting in a total of \\(m \\times k\\) candidate documents. These documents are then concatenated with the query one by one and used as prompts for the LLM, producing \\(m \\times k\\) outputs:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.124,
+ 0.831,
+ 0.443,
+ 0.849
+ ],
+ "angle": 0,
+ "content": "\\[\n\\{O _ {q, d _ {i, j}} = \\mathrm {L L M} (q, d _ {i, j}) | i \\in \\{1, \\dots , m \\}, j \\in \\{1, \\dots , k \\} \\},\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.855,
+ 0.484,
+ 0.899
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathrm{LLM}(q, d_{i,j})\\) represents the output generated by inputting the concatenated query \\(q\\) and document \\(d_{i,j}\\) into the LLM. Then, based on the quality of these outputs, we can calculate the distribution of"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.515,
+ 0.107,
+ 0.75,
+ 0.121
+ ],
+ "angle": 0,
+ "content": "these candidate documents as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.564,
+ 0.125,
+ 0.913,
+ 0.163
+ ],
+ "angle": 0,
+ "content": "\\[\np _ {\\text {L L M}} \\left(d _ {i, j} \\mid q, y\\right) = \\frac {\\exp (\\operatorname {e v a l} \\left(y , O _ {q , d _ {i , j}}\\right))}{\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {k} \\exp (\\operatorname {e v a l} \\left(y , O _ {q , d _ {i , j}}\\right))}, \\tag {6}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.168,
+ 0.915,
+ 0.254
+ ],
+ "angle": 0,
+ "content": "where \\( \\text{eval}(\\cdot) \\) measures the difference between the target output \\( y \\) and the LLM's output, using metrics such as ROUGE [24] score. A larger value returned by \\( \\text{eval}(\\cdot) \\) indicates a better-generated result. Similarly, we can also calculate the score distribution of the candidate documents by the retrieval model based on \\( S_{u,q,d}^{\\text{retriever}} \\) in Eq. (5):"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.571,
+ 0.26,
+ 0.913,
+ 0.305
+ ],
+ "angle": 0,
+ "content": "\\[\np _ {\\text {r e t r i e v e r}} \\left(d _ {i, j} \\mid q, u\\right) = \\frac {\\exp \\left(S _ {u , q , d _ {i , j}} ^ {\\text {r e t r i e v e r}}\\right)}{\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {k} \\exp \\left(S _ {u , q , d _ {i , j}} ^ {\\text {r e t r i e v e r}}\\right)}. \\tag {7}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.309,
+ 0.915,
+ 0.379
+ ],
+ "angle": 0,
+ "content": "We aim for the retrieval model to retrieve documents that lead to better LLM-generated results, which means making the distribution \\( p_{\\mathrm{retriever}}(d|q,u) \\) in Eq. (7) closer to the distribution \\( p_{\\mathrm{LLM}}(d|q,y) \\) in Eq (6). Therefore, we compute the KL divergence between the two distributions as the loss to optimize the retriever:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.568,
+ 0.386,
+ 0.913,
+ 0.402
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {L} _ {\\text {r e t r i e v e r}} = \\mathrm {K L} \\left(p _ {\\text {r e t r i e v e r}} (d | q, u) \\mid p _ {\\text {L L M}} (d | q, y)\\right). \\tag {8}\n\\]"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.414,
+ 0.71,
+ 0.427
+ ],
+ "angle": 0,
+ "content": "4.3 Document Rerank"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.432,
+ 0.914,
+ 0.475
+ ],
+ "angle": 0,
+ "content": "After retrieving \\(\\mathcal{D}_{\\mathrm{retrieved}}\\) through the retriever, in this section, we further refine the results by reranking \\(\\mathcal{D}_{\\mathrm{retrieved}}\\) to obtain the final top-\\(k\\) ranked results \\(\\mathcal{D}_{\\mathrm{reranked}} = \\{d_i | i \\in \\{1, \\dots, k\\}\\}\\)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.482,
+ 0.913,
+ 0.536
+ ],
+ "angle": 0,
+ "content": "4.3.1 Reranker. We use a pre-trained cross-encoder (such as the BGE reranker [45]) to encode the query and document, obtaining the hidden state corresponding to the [CLS] token from the last layer:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.631,
+ 0.54,
+ 0.913,
+ 0.557
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathbf {h} _ {q, d} = \\operatorname {C r o s s E n c o d e r} (q, d), \\tag {9}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.561,
+ 0.915,
+ 0.687
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathbf{h}_{q,d} \\in \\mathbb{R}^d\\). Similarly, when reranking, in addition to considering the semantic relevance between query and document, we also take into account the user's personalized preferences. However, since the cross-encoder does not encode documents separately, it cannot compute the cosine similarity between users and documents as shown in Eq. (4) to express the user preference score. Therefore, we directly concatenate the user embeddings to the output of the cross-encoder to account for the influence of user preferences. The overall score used for reranking is calculated as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.573,
+ 0.693,
+ 0.913,
+ 0.713
+ ],
+ "angle": 0,
+ "content": "\\[\nS _ {u, q, d} ^ {\\text {r e r a n k e r}} = \\mathrm {M L P} _ {3} \\left(\\operatorname {C O N C A T} \\left(\\mathbf {h} _ {q, d}, \\operatorname {M L P} _ {2} (\\mathbf {e} _ {u})\\right)\\right), \\tag {10}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.72,
+ 0.913,
+ 0.751
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathrm{MLP}_2: \\mathbb{R}^d \\to \\mathbb{R}^d\\) and \\(\\mathrm{MLP}_3: \\mathbb{R}^{2d} \\to \\mathbb{R}\\) are two multi-layer perceptions. \\(\\mathrm{CONCAT}(\\cdot)\\) denotes the concatenation operation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.758,
+ 0.913,
+ 0.813
+ ],
+ "angle": 0,
+ "content": "4.3.2 Training. Similar to the retriever's training in Section 4.2.2, we also want the reranker to assign higher scores to the documents that lead to better LLM-generated results. Therefore, we train the reranker using a similar approach."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.813,
+ 0.915,
+ 0.897
+ ],
+ "angle": 0,
+ "content": "We use the trained retrieval model from Section 4.2.2 to retrieve top-\\(k\\) documents from the history of each of the \\(m\\) users, resulting in a total of \\(m \\times k\\) candidate documents. These documents are concatenated with the query \\(q\\) and used as prompts for the LLM, producing \\(m \\times k\\) outputs. Similar to Eq.(6), we can obtain the distribution \\(p_{\\mathrm{LLM}}(d|q,y)\\) of these candidate documents. Based on"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.085,
+ 0.076,
+ 0.28,
+ 0.087
+ ],
+ "angle": 0,
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.843,
+ 0.076,
+ 0.913,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "Teng Shi et al."
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.111,
+ 0.105,
+ 0.454,
+ 0.12
+ ],
+ "angle": 0,
+ "content": "Table 1: Statistics of the datasets used in this paper."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.09,
+ 0.125,
+ 0.478,
+ 0.2
+ ],
+ "angle": 0,
+ "content": "| Task | Input | Output | User History |
| LaMP-1 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please choose one of the following two references that is more relevant to the user's input title: [1] {reference1}; [2] {reference2}. Please just answer with “[1” or “[2” without explanation. “title”: {title}. | [1] | “title”: {title}\n“abstract”: {abstract} |
| LaMP-2 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please select the tag from [sci-fi, based on a book, comedy ... ] that is most relevant to the user's input description. Please just answer with the tag name without explanation. “description”: {description}; “tag”: | comedy | “description”: {description};\n“tag”: {tag} |
| LaMP-3 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, what is the score of the following review on a scale of 1 to 5? just answer with 1, 2, 3, 4, or 5 without further explanation. “review”: {review}; “score”: | 5 | “review”: {review}\n“score”: {score} |
| LaMP-4 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please generate a title for the given user's input text. Please generate it in the following format: {"title": "generated title"} without explanation, and use only English. “text”: {text}; “title”: | {“title”: Finding Happiness \nAfter Divorce - It Can Happen} | “text”: {text}\n“title”: {title} |
| LaMP-5 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please generate a title for the given user's input abstract. Please generate it in the following format: {"title": "generated title"} without explanation, and use only English. “abstract”: {abstract}; “title”: | {“title”: Link-Reliability Based \nTwo-Hop Routing for \nWireless Sensor Networks.} | “abstract”: {abstract}\n“title”: {title} |
| LaMP-7 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the style pattern of the historical tweets provided, please paraphrase the user's input tweet without any explanation before or after it. Please generate it in the following format: {"tweet": "generated tweet"} without explanation, and use only English. “tweet”: {tweet}. | {“tweet”:lilxcutiesworld the \ndanny picture is GOOD!! \nI really like it.} | “tweet”: {tweet} |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.632,
+ 0.483,
+ 0.771
+ ],
+ "angle": 0,
+ "content": "5.4.4 Impact of the Number of Retrieved Users. Since we enhance personalized text generation by introducing collaborative filtering, we further explored how much collaborative information to introduce, specifically the impact of the number of retrieved users on the results, as shown in Figure 8. In LaMP-1, retrieving too few or too many users leads to poorer performance, with the best results at 4 users. In LaMP-5, the performance improves as the number of users increases. This highlights the importance of introducing collaborative filtering, but it also indicates that excessive introduction can lead to decreased effectiveness."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.082,
+ 0.786,
+ 0.483,
+ 0.897
+ ],
+ "angle": 0,
+ "content": "5.4.5 Impact of the Number of Retrieved Documents. We also analyzed the impact of the number of retrieved documents, \\( k \\), on the results, as shown in Figure 9. It can be observed that as the number of retrieved documents increases, performance improves, indicating the importance of retrieving user history to reflect user preferences for enhancing LLM-generated results. Since more documents lead to longer prompts and slower LLM generation, we chose \\( k = 5 \\) for our experiments."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.631,
+ 0.64,
+ 0.645
+ ],
+ "angle": 0,
+ "content": "6 Conclusion"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.649,
+ 0.916,
+ 0.801
+ ],
+ "angle": 0,
+ "content": "In this paper, we propose CFRAG, which adapts collaborative filtering into RAG to personalize LLMs. To introduce collaborative information without explicit user labels and retrieve documents that support personalized LLM generation, we first train user embeddings through contrastive learning to retrieve similar users. Then, we design the personalized retriever and reranker that considers user preferences during retrieval and fine-tune them using LLM feedback. The results on the Language Model Personalization (LaMP) benchmark validate the effectiveness of CFRAG. The experimental analysis also confirms the effectiveness of each module within CFRAG."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.85,
+ 0.71,
+ 0.866
+ ],
+ "angle": 0,
+ "content": "A Appendix: Prompts"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.514,
+ 0.868,
+ 0.916,
+ 0.896
+ ],
+ "angle": 0,
+ "content": "We provide detailed formats for the inputs, outputs, and user histories for the LLM across different datasets, as shown in Table 4."
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.085,
+ 0.076,
+ 0.28,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.843,
+ 0.076,
+ 0.913,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "Teng Shi et al."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.085,
+ 0.106,
+ 0.178,
+ 0.12
+ ],
+ "angle": 0,
+ "content": "References"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.093,
+ 0.123,
+ 0.482,
+ 0.143
+ ],
+ "angle": 0,
+ "content": "[1] AI@Meta. 2024. Llama 3 Model Card. (2024). https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.092,
+ 0.144,
+ 0.483,
+ 0.174
+ ],
+ "angle": 0,
+ "content": "[2] Akari Asai, Zegiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. [n.d.]. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. In The Twelfth International Conference on Learning Representations."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.092,
+ 0.174,
+ 0.482,
+ 0.204
+ ],
+ "angle": 0,
+ "content": "[3] Sebastian Borgeaud, Arthur Mensch, et al. 2022. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, PMLR, 2206-2240."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.092,
+ 0.205,
+ 0.482,
+ 0.233
+ ],
+ "angle": 0,
+ "content": "[4] Jin Chen, Zheng Liu, et al. 2024. When large language models meet personalization: Perspectives of challenges and opportunities. World Wide Web 27, 4 (2024), 42."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.092,
+ 0.234,
+ 0.483,
+ 0.264
+ ],
+ "angle": 0,
+ "content": "[5] Sunhao Dai, Ninglu Shao, et al. 2023. Uncovering chatgpt's capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems. 1126-1132."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.092,
+ 0.265,
+ 0.482,
+ 0.315
+ ],
+ "angle": 0,
+ "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.092,
+ 0.316,
+ 0.482,
+ 0.355
+ ],
+ "angle": 0,
+ "content": "[7] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 6491-6501."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.092,
+ 0.356,
+ 0.482,
+ 0.385
+ ],
+ "angle": 0,
+ "content": "[8] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.092,
+ 0.386,
+ 0.482,
+ 0.415
+ ],
+ "angle": 0,
+ "content": "[9] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. In International conference on machine learning. PMLR, 3929-3938."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.088,
+ 0.416,
+ 0.482,
+ 0.455
+ ],
+ "angle": 0,
+ "content": "[10] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639-648."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.456,
+ 0.482,
+ 0.485
+ ],
+ "angle": 0,
+ "content": "[11] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web. 173-182."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.486,
+ 0.482,
+ 0.516
+ ],
+ "angle": 0,
+ "content": "[12] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. [n.d.]. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.517,
+ 0.482,
+ 0.556
+ ],
+ "angle": 0,
+ "content": "[13] Gautier Izacard and Edouard Grave. 2021. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 874-880."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.557,
+ 0.482,
+ 0.577
+ ],
+ "angle": 0,
+ "content": "[14] Gautier Izacard, Patrick Lewis, et al. 2022. Few-shot learning with retrieval augmented language models. arXiv preprint arXiv:2208.03299 1, 2 (2022), 4."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.578,
+ 0.482,
+ 0.607
+ ],
+ "angle": 0,
+ "content": "[15] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. 2020. A survey on contrastive self-supervised learning. Technologies 9, 1 (2020), 2."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.608,
+ 0.482,
+ 0.637
+ ],
+ "angle": 0,
+ "content": "[16] Joel Jang, Seungone Kim, et al. 2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv preprint arXiv:2310.11564 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.638,
+ 0.482,
+ 0.667
+ ],
+ "angle": 0,
+ "content": "[17] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In International Conference on Machine Learning. PMLR, 15696-15707."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.668,
+ 0.482,
+ 0.687
+ ],
+ "angle": 0,
+ "content": "[18] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.688,
+ 0.482,
+ 0.707
+ ],
+ "angle": 0,
+ "content": "[19] Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer 42, 8 (2009), 30-37."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.708,
+ 0.482,
+ 0.737
+ ],
+ "angle": 0,
+ "content": "[20] Patrick Lewis, Ethan Perez, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33 (2020), 9459-9474."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.738,
+ 0.482,
+ 0.768
+ ],
+ "angle": 0,
+ "content": "[21] Cheng Li, Mingyang Zhang, Qiaozhu Mei, Yaqing Wang, Spurthi Amba Hombaiah, Yi Liang, and Michael Bendersky. 2023. Teach LLMs to Personalize-An Approach inspired by Writing Education. arXiv preprint arXiv:2308.07968 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.769,
+ 0.482,
+ 0.798
+ ],
+ "angle": 0,
+ "content": "[22] Junyi Li, Tianyi Tang, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2024. Pre-trained language models for text generation: A survey. Comput. Surveys 56, 9 (2024), 1-39."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.799,
+ 0.482,
+ 0.818
+ ],
+ "angle": 0,
+ "content": "[23] Xinyu Li, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.819,
+ 0.482,
+ 0.838
+ ],
+ "angle": 0,
+ "content": "[24] Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out. 74-81."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.839,
+ 0.482,
+ 0.878
+ ],
+ "angle": 0,
+ "content": "[25] Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. [n.d.]. RA-DIT: Retrieval-Augmented Dual Instruction Tuning. In The Twelfth International Conference on Learning Representations."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.879,
+ 0.482,
+ 0.897
+ ],
+ "angle": 0,
+ "content": "[26] Yinhan Liu. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.087,
+ 0.123,
+ 0.483,
+ 0.897
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.517,
+ 0.109,
+ 0.913,
+ 0.139
+ ],
+ "angle": 0,
+ "content": "[27] Sheshera Mysore, Zhuoran Lu, et al. 2023. Pearl: Personalizing large language model writing assistants with generation-calibrated retrievers. arXiv preprint arXiv:2311.09180 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.141,
+ 0.913,
+ 0.16
+ ],
+ "angle": 0,
+ "content": "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.161,
+ 0.913,
+ 0.2
+ ],
+ "angle": 0,
+ "content": "[29] Chris Richardson, Yao Zhang, Kellen Gillespie, Sudipta Kar, Arshdeep Singh, Zeynab Raeesy, Omar Zia Khan, and Abhinav Sethy. 2023. Integrating summarization and retrieval for enhanced personalization via large language models. arXiv preprint arXiv:2310.20081 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.201,
+ 0.913,
+ 0.23
+ ],
+ "angle": 0,
+ "content": "[30] Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp 109 (1995), 109."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.231,
+ 0.913,
+ 0.27
+ ],
+ "angle": 0,
+ "content": "[31] Alireza Salemi, Surya Kallumadi, and Hamed Zamani. 2024. Optimization methods for personalizing large language models through retrieval augmentation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 752-762."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.271,
+ 0.913,
+ 0.3
+ ],
+ "angle": 0,
+ "content": "[32] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2023. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.301,
+ 0.913,
+ 0.33
+ ],
+ "angle": 0,
+ "content": "[33] Chenglei Shen, Xiao Zhang, Teng Shi, Changshuo Zhang, Guofu Xie, and Jun Xu. 2024. A survey of controllable learning: Methods and applications in information retrieval. arXiv preprint arXiv:2407.06083 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.332,
+ 0.913,
+ 0.381
+ ],
+ "angle": 0,
+ "content": "[34] Teng Shi, Zihua Si, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Dewei Leng, Yanan Niu, and Yang Song. 2024. UniSAR: Modeling User Transition Behaviors between Search and Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1029-1039."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.382,
+ 0.913,
+ 0.432
+ ],
+ "angle": 0,
+ "content": "[35] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2024. REPLUG: Retrieval-Augmented Black-Box Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 8364-8377."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.433,
+ 0.913,
+ 0.462
+ ],
+ "angle": 0,
+ "content": "[36] Zhaoxuan Tan, Zheyuan Liu, and Meng Jiang. 2024. Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts. arXiv preprint arXiv:2406.10471 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.463,
+ 0.913,
+ 0.492
+ ],
+ "angle": 0,
+ "content": "[37] Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2024. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. arXiv:2402.04401 [cs.CL] https://arxiv.org/abs/2402.04401"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.493,
+ 0.913,
+ 0.532
+ ],
+ "angle": 0,
+ "content": "[38] Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Wu Jian, and Yuning Jiang. 2025. Think Before Recommend: Unleashing the Latent Reasoning Power for Sequential Recommendation. arXiv:2503.22675 [cs.IR] https://arxiv.org/abs/2503.22675"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.533,
+ 0.913,
+ 0.552
+ ],
+ "angle": 0,
+ "content": "[39] A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.553,
+ 0.913,
+ 0.583
+ ],
+ "angle": 0,
+ "content": "[40] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval. 165-174."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.584,
+ 0.913,
+ 0.613
+ ],
+ "angle": 0,
+ "content": "[41] Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al. 2024. A survey on large language models for recommendation. World Wide Web 27, 5 (2024), 60."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.614,
+ 0.913,
+ 0.643
+ ],
+ "angle": 0,
+ "content": "[42] Xinghao Wu, Xuefeng Liu, Jianwei Niu, Haolin Wang, Shaojie Tang, and Guogang Zhu. 2024. FedLoRA: When Personalized Federated Learning Meets Low-Rank Adaptation. (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.644,
+ 0.913,
+ 0.683
+ ],
+ "angle": 0,
+ "content": "[43] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. 2024. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems 36 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.684,
+ 0.913,
+ 0.713
+ ],
+ "angle": 0,
+ "content": "[44] Zhuofeng Wu, Sinong Wang, Jiatao Gu, Madian Khabsa, Fei Sun, and Hao Ma. 2020. Clear: Contrastive learning for sentence representation. arXiv preprint arXiv:2012.15466 (2020)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.714,
+ 0.913,
+ 0.744
+ ],
+ "angle": 0,
+ "content": "[45] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597 [cs.CL]"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.745,
+ 0.913,
+ 0.773
+ ],
+ "angle": 0,
+ "content": "[46] Hong-Jian Xue, Xinyu Dai, Jianbing Zhang, Shujian Huang, and Jiajun Chen. 2017. Deep matrix factorization models for recommender systems.. In IJCAI, Vol. 17. Melbourne, Australia, 3203-3209."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.775,
+ 0.913,
+ 0.804
+ ],
+ "angle": 0,
+ "content": "[47] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng-peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.805,
+ 0.913,
+ 0.835
+ ],
+ "angle": 0,
+ "content": "[48] Changshuo Zhang, Teng Shi, Xiao Zhang, Qi Liu, Ruobing Xie, Jun Xu, and Ji-Rong Wen. 2024. Modeling Domain and Feedback Transitions for Cross-Domain Sequential Recommendation. arXiv preprint arXiv:2408.08209 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.836,
+ 0.913,
+ 0.865
+ ],
+ "angle": 0,
+ "content": "[49] Changshuo Zhang, Teng Shi, Xiao Zhang, Yanping Zheng, Ruobing Xie, Qi Liu, Jun Xu, and Ji-Rong Wen. 2024. QAGCF: Graph Collaborative Filtering for Q&A Recommendation. arXiv preprint arXiv:2406.04828 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.866,
+ 0.913,
+ 0.895
+ ],
+ "angle": 0,
+ "content": "[50] Changshuo Zhang, Xiao Zhang, Teng Shi, Jun Xu, and Ji-Rong Wen. 2025. Test-Time Alignment for Tracking User Interest Shifts in Sequential Recommendation. arXiv:2504.01489 [cs.IR] https://arxiv.org/abs/2504.01489"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.517,
+ 0.109,
+ 0.913,
+ 0.895
+ ],
+ "angle": 0,
+ "content": null
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.085,
+ 0.076,
+ 0.533,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.719,
+ 0.076,
+ 0.912,
+ 0.088
+ ],
+ "angle": 0,
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.085,
+ 0.109,
+ 0.484,
+ 0.151
+ ],
+ "angle": 0,
+ "content": "[51] Kepu Zhang, Teng Shi, Sunhao Dai, Xiao Zhang, Yinfeng Li, Jing Lu, Xiaoxue Zang, Yang Song, and Jun Xu. 2024. SAQRec: Aligning Recommender Systems to User Satisfaction via Questionnaire Feedback. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 3165-3175."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.151,
+ 0.484,
+ 0.182
+ ],
+ "angle": 0,
+ "content": "[52] Xiao Zhang, Teng Shi, Jun Xu, Zhenhua Dong, and Ji-Rong Wen. 2024. Model-Agnostic Causal Embedding Learning for Counterfactually Group-Fair Recommendation. IEEE Transactions on Knowledge and Data Engineering (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.182,
+ 0.482,
+ 0.222
+ ],
+ "angle": 0,
+ "content": "[53] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219 (2023)."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.085,
+ 0.109,
+ 0.484,
+ 0.222
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.517,
+ 0.109,
+ 0.913,
+ 0.14
+ ],
+ "angle": 0,
+ "content": "[54] Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2024. Dense text retrieval based on pretrained language models: A survey. ACM Transactions on Information Systems 42, 4 (2024), 1-60."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.518,
+ 0.141,
+ 0.914,
+ 0.171
+ ],
+ "angle": 0,
+ "content": "[55] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.517,
+ 0.171,
+ 0.914,
+ 0.201
+ ],
+ "angle": 0,
+ "content": "[56] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.517,
+ 0.201,
+ 0.915,
+ 0.231
+ ],
+ "angle": 0,
+ "content": "[57] Yuchen Zhuang, Haotian Sun, Yue Yu, Qifan Wang, Chao Zhang, and Bo Dai. 2024. HYDRA: Model Factorization Framework for Black-Box LLM Personalization. arXiv preprint arXiv:2406.02888 (2024)."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.517,
+ 0.109,
+ 0.915,
+ 0.231
+ ],
+ "angle": 0,
+ "content": null
+ }
+ ]
+]
\ No newline at end of file
diff --git a/data/2025/2504_05xxx/2504.05731/1d2e0a8a-e48a-4cb6-8ec0-35774101eda9_origin.pdf b/data/2025/2504_05xxx/2504.05731/1d2e0a8a-e48a-4cb6-8ec0-35774101eda9_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4bb5ba1fef8056c4c4d9711edfda47b980dc457c
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/1d2e0a8a-e48a-4cb6-8ec0-35774101eda9_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24cf8e274e0db1c92f380e61d9d9f76f575771d23873fefb5f116bf0c00ceb3c
+size 1203734
diff --git a/data/2025/2504_05xxx/2504.05731/full.md b/data/2025/2504_05xxx/2504.05731/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..9a01de5351d2b3d0c4960fbe0259c47f59fe8b13
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/full.md
@@ -0,0 +1,477 @@
+# Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation
+
+Teng Shi
+
+Renmin University of China
+
+Beijing, China
+
+shiteng@ruc.edu.cn
+
+Xiaoxue Zang
+
+Kai Zheng
+
+Kuaishou Technology Co., Ltd.
+
+Beijing, China
+
+xxic666@126.com
+
+zhengk92@gmail.com
+
+# Abstract
+
+Recently, the personalization of Large Language Models (LLMs) to generate content that aligns with individual user preferences has garnered widespread attention. Personalized Retrieval-Augmented Generation (RAG), which retrieves relevant documents from the user's history to reflect their preferences and enhance LLM generation, is one commonly used approach for personalization. However, existing personalized RAG methods do not consider that the histories of similar users can also assist in personalized generation for the current user, meaning that collaborative information between users can also benefit personalized generation. Inspired by the application of collaborative filtering in recommender systems, we propose a method called CFRAG, which adapts Collaborative Filtering to RAG for personalized text generation. However, this presents two challenges: (1) how to incorporate collaborative information without explicit user similarity labels? (2) how to retrieve documents that support personalized LLM generation? For Challenge 1, we use contrastive learning to train user embeddings to retrieve similar users and introduce collaborative information. For Challenge 2, we design a personalized retriever and reranker to retrieve the top- $k$ documents from these users' histories. We take into account the user's preference during retrieval and reranking. Then we leverage feedback from the LLM to fine-tune the personalized retriever and reranker, enabling them to retrieve documents that meet the personalized generation needs of the LLM. Experimental results on the Language Model Personalization (LaMP) benchmark
+
+*Corresponding authors. Work partially done at Engineering Research Center of Next-Generation Intelligent Search and Recommendation, Ministry of Education.
+Work done when Teng Shi was the intern at Kuaishou.
+
+Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org.
+
+SIGIR '25, Padua, Italy.
+
+© 2025 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 979-8-4007-1592-1/25/07
+
+https://doi.org/10.1145/XXXXXX.XXXXXXX
+
+Jun Xu*
+
+Xiao Zhang
+
+Renmin University of China
+
+Beijing, China
+
+{junxu,zhangx89}@ruc.edu.cn
+
+Yang Song
+
+Han Li
+
+Kuaishou Technology Co., Ltd.
+
+Beijing, China
+
+ys@sonyis.me
+
+lihan08@kuaishou.com
+
+validate the effectiveness of CFRAG. Further analysis confirms the importance of incorporating collaborative information.
+
+# CCS Concepts
+
+- Information systems $\rightarrow$ Personalization; - Computing methodologies $\rightarrow$ Natural language generation.
+
+# Keywords
+
+Large language model; Personalization; Retrieval augmented generation
+
+# ACM Reference Format:
+
+Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Han Li. 2025. Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '25), July 13-18, 2025, Padua, Italy. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/XXXXXX.XXXXXXX
+
+# 1 Introduction
+
+Personalizing Large Language Models (LLMs) [55] to generate personalized outputs tailored to individual user preferences has emerged as a significant and rapidly growing field [16, 23, 29, 31, 32, 36, 37, 57]. Personalized Retrieval-Augmented Generation (RAG) [8] has become a commonly used approach for personalizing LLMs [29, 31, 32, 57].
+
+The process of existing personalized RAG methods typically involves retrieving similar documents from the user's historical behaviors based on the user's input query, then concatenating these documents with the query as a prompt input to the LLM for generation. Although effective, this approach is limited to retrieving only the current user's history, neglecting collaborative information. Users with similar histories tend to be more alike, and the information from these similar users can also aid in personalizing generation for the current user. As shown in the example in Figure 1, the upper part illustrates the results of the existing RAG method, which retrieves documents from the current user's history. We can only infer from these results that "She" in the user's input refers to "Hillary Clinton". In contrast, the lower part demonstrates
+
+
+Figure 1: An example from the LaMP-4 dataset [32]. The task of LaMP-4 is to generate personalized news headlines based on user input. This example illustrates the benefit of collaborative information for LLM personalization: (a) The top shows results retrieved by the existing RAG method from the current user's history, where we can only infer that "She" in the user's input refers to "Hillary Clinton". (b) The bottom shows results retrieved by our method from similar users' histories, allowing us to infer further that "his" in the user's input refers to "Donald Trump" thus enabling the generation of a more accurate result.
+
+our method, which retrieves documents from the history of similar users. In this case, we can further infer that "his" in the user's input refers to "Donald Trump", leading to a better generation result. From this example, we can see that incorporating collaborative information allows the retrieval of more diverse documents, helping the LLM generate results that better meet the user's needs.
+
+Inspired by the application of collaborative filtering in recommender systems [11, 40, 46], we propose to adapt collaborative information into RAG to personalize LLMs. However, adapting collaborative filtering to personalized RAG presents two challenges. Challenge 1: How to incorporate collaborative information. Without explicit labels indicating which users are similar, which users' information should be selected to help personalize generation for the current user? Challenge 2: How to retrieve documents that support personalized LLM generation, rather than relying on traditional semantic relevance? Pre-trained dense retrieval models [54] only retrieve based on the semantic relevance between the query and document. Directly using these models for retrieval may not necessarily result in content that allows the LLM to generate outputs that meet the user's needs [25, 35].
+
+To address the above challenges, this paper proposes a method named CFRAG which adapts Collaborative Filtering to personalized Retrieval Augmented Generation. Firstly, to address Challenge 1, since there are no explicit user similarity labels, we use contrastive learning [15, 44] to train user embeddings for retrieving similar users to introduce collaborative information. Specifically, we apply different data augmentation methods to the user's history to obtain different views, and then treat different views of the same user's history as positive samples for each other. Then we use contrastive learning on different views to train the user embeddings. Secondly, for Challenge 2, we designed a personalized retriever and reranker to retrieve the top- $k$ documents from the histories of the retrieved users. In both retrieval and reranking, in addition to the semantic
+
+relevance between the query and documents, we also considered the user's preferences for different documents to enable personalized retrieval. Additionally, we further fine-tune the retriever and reranker based on the feedback from the LLM to ensure that the retrieved documents better support the personalized LLM generation. Finally, the top- $k$ documents are concatenated with the user's input query to form a prompt, which is then fed into the LLM for personalized generation.
+
+The major contributions of the paper are summarized as follows: We analyzed the necessity of introducing collaborative filtering into RAG for LLM personalization and identified the challenges: how to introduce collaborative information and how to retrieve documents that support personalized LLM generation.
+
+- We proposed a method called CFRAG, which uses contrastive learning to train user embeddings for retrieving similar users and incorporating collaborative information. It leverages LLM feedback to train the personalized retriever and reranker, enabling them to retrieve documents that support personalized LLM generation.
+- Experimental results on the Language Model Personalization (LaMP) [32] benchmark validate the effectiveness of CFRAG. The experimental analysis also demonstrates the importance of leveraging collaborative information.
+
+# 2 Related Work
+
+Personalization of LLMs. Large Language Models (LLMs) [55] have demonstrated remarkable capabilities in various fields, such as text generation [22], information retrieval [56], recommender systems [5, 41], and so on. However, since LLMs are typically designed to serve all tasks with a single model and are trained on broad, domain-agnostic data, they face challenges in adapting to the personalized needs of individual users [4, 32]. Therefore, LLM personalization has attracted widespread attention [16, 31, 57].
+
+
+Figure 2: The architecture of CFRAG. From left to right: (a) User Retrieval retrieves similar users (Section 4.1); (b) Retriever retrieves the top- $k$ documents from each user's history (Section 4.2); (c) Reranker reranks the $m \times k$ documents to get the final top- $k$ documents, which are then concatenated with the query and input into the LLM for personalized text generation (Section 4.3).
+
+Existing works on LLM personalization mainly include the following types of methods: (1) Fine-tuning a personalized LLM for each user [36, 37, 42]; Tan et al. [37] fine-tuned the LLM using LoRA [12] to get personalized LoRA parameters for each user. (2) Aligning LLMs with user-specific preferences through Reinforcement Learning from Human Feedback (RLHF) [16, 23, 43]; Jang et al. [16] first trained different parameters for various objectives using RLHF, then merged these parameters based on users' personalized needs. (3) Incorporating user-specific context into the prompt [21, 27, 29, 31, 32, 57]. Richardson et al. [29] used instruction-tuned LLMs to summarize user history and then incorporated it into prompts for generation. Salemi et al. [31, 32] used RAG to retrieve relevant documents from user history based on the input query and incorporated them into the prompt.
+
+This paper further introduces collaborative filtering for personalization based on the RAG framework. Collaborative filtering has already been applied in fields such as recommender systems [33, 34, 38, 48-52] and has been proven effective. It assumes that users who have interacted with similar items share similar preferences, and recommending items from similar users to the current user can meet their needs. Some works [11, 46] learn the collaborative information between users and items through matrix factorization [19], while others [10, 40] further explore higher-order collaborative information between users and items using graph neural networks. The application of collaborative filtering in LLM personalization remains under-explored.
+
+Retrieval Augmented Generation. Retrieval Augmented Generation [7, 8] introduces external knowledge through document retrieval, alleviating issues such as LLM hallucinations [53], and enhancing LLMs' capabilities in knowledge-intensive tasks [17] such as open-domain question answering [14, 20]. Some works [3, 13] encode retrieved documents using separate encoders, and then fuse the results with the language model using cross-attention. A more common approach is to directly include the retrieved documents in the prompt of the LLM [2, 9, 20, 25, 35]. In recent years, this
+
+in-context RAG framework has also been applied to LLM personalization, which is personalized by retrieving documents from the user's history [31, 32, 57]. This paper introduces collaborative filtering by retrieving similar users' histories for better personalization.
+
+# 3 Problem Formulation
+
+Let $\mathcal{U} = \{u_1, u_2, \ldots, u_M\}$ denotes the set of all users, where $M$ is the number of users. Each user $u \in \mathcal{U}$ has a chronologically ordered history $\mathcal{H}_u = [d_1, d_2, \ldots, d_N]$ which includes all her historical documents, where $N$ is the number of documents in the history. The personalized text generation dataset is $\mathcal{D} = \{(u, q, y)_i\}_{i=1}^{|D|}$ . For each instance, $q$ is the query input by the user $u$ to the LLM, and $y$ is the target output. Our goal is first to introduce collaborative information by retrieving the top- $m$ most similar users for user $u$ :
+
+$$
+\mathcal {U} _ {\text {r e t r i e v e d}} = \left\{u _ {1}, u _ {2}, \dots , u _ {m} \right\}.
+$$
+
+Then, we use a retriever to retrieve the top- $k$ documents from each of the $m$ users' histories, resulting in a total of $m \times k$ documents.
+
+$$
+\mathcal {D} _ {\text {r e t r i e v e d}} = \{d _ {i, j} | i \in \{1, \dots , m \}, j \in \{1, \dots , k \} \}.
+$$
+
+Finally, we use a reranker to rerank these $m \times k$ documents and obtain the final top- $k$ documents:
+
+$$
+\mathcal {D} _ {\text {r e r a n k e d}} = \left\{d _ {i} | i \in \{1, \dots , k \} \right\}.
+$$
+
+These top- $k$ documents will be concatenated with the user's query $q$ as a prompt and input into the LLM, enabling it to generate a response that aligns with the target output $y$ .
+
+This paper primarily focuses on how to retrieve $\mathcal{U}_{\mathrm{retrieved}}$ to introduce collaborative information, and how to train the retriever and reranker so that they can effectively retrieve documents that support the personalized LLM generation.
+
+# 4 Our Approach
+
+This section introduces our method CFRAG. CFRAG's overall architecture is shown in Figure 2. As mentioned in Section 1, to address
+
+Challenge 1, i.e., how to introduce collaborative information, we first train user embeddings using contrastive learning to retrieve the top- $m$ most similar users (see Section 4.1). For Challenge 2, which involves retrieving documents that support personalized LLM generation, we fine-tune the personalized retriever and reranker using LLM feedback. The retriever first retrieves the top- $k$ documents from the history of each of the $m$ users, resulting in $m \times k$ documents (see Section 4.2). The reranker then reranks these documents to obtain the final top- $k$ documents as input for the LLM (see Section 4.3).
+
+# 4.1 User Retrieval
+
+First, we perform user retrieval to get the top- $m$ most similar users for user $u$ to introduce collaborative information. However, we do not have labels indicating which users are similar to each other. To address this, we employ a contrastive learning [15, 44] approach. We apply different data augmentation methods to the user history $\mathcal{H}_u$ to obtain different views of the user's history. We treat different views of the same user as positive samples and the histories of other users as negative samples, and then we use the InfoNCE [28] loss to train user embeddings for retrieval. Figure 3 illustrates the process of training user embeddings using contrastive learning.
+
+4.1.1 User Encoder. Specifically, we first use an embedding model (such as BERT [6], RoBERTa [26], BGE [45] etc.) $\mathbf{Emb}(\cdot)$ to encode each document in the user's history $\mathcal{H}_u$ to obtain $\mathbf{E}_u = [\mathbf{e}_1,\mathbf{e}_2,\dots ,\mathbf{e}_N]^{\intercal}\in \mathbb{R}^{N\times d}$ , where $\mathbf{e}_i = \mathbf{Emb}(d_i)$ and $d$ is the embedding dimension. To model the sequential relationships between different documents in the user's history, we introduce positional embedding $\mathbf{P}\in \mathbb{R}^{N\times d}$ . Afterward, the history $\mathcal{H}_u$ 's embedding becomes $\widehat{\mathbf{E}}_u = \mathbf{E}_u + \mathbf{P}$ . Then, we apply a transformer [39] as the user encoder to encode the user's history $\widehat{\mathbf{E}}_u$ and average the transformer's output to obtain the user's embedding:
+
+$$
+\mathbf {e} _ {u} = \operatorname {E n c o d e r} _ {u} (u) = \operatorname {M E A N} (\operatorname {T r m} (\widehat {\mathbf {F}} _ {u})) \in \mathbb {R} ^ {d}, \tag {1}
+$$
+
+where $\mathrm{Encoder}_u(\cdot)\to \mathbb{R}^d$ denotes the user encoder, $\mathrm{Trm}(\cdot)$ denotes a transformer encoder. Next, we train the transformer encoder using contrastive learning.
+
+4.1.2 Data Augmentation. We generate different views of $\mathcal{H}_u$ using the following three data augmentation methods:
+
+Document Crop. We randomly select a continuous sub-sequence of length $L_{c} = \lfloor \eta_{c}N\rfloor$ from $\mathcal{H}_u$ , where $\eta_c$ is a hyper-parameter controlling the crop ratio. The history after cropping is as follows:
+
+$$
+\mathcal {H} _ {u} ^ {\mathrm {c r o p}} = [ d _ {c}, d _ {c + 1}, \dots , d _ {c + L _ {c} - 1} ].
+$$
+
+Document Mask. For the history $\mathcal{H}_u$ , we randomly mask out $L_{m} = \lfloor \eta_{m}N\rfloor$ documents $\mathcal{I}_{\mathrm{mask}} = \{i_1,i_2,\dots ,i_{L_m}\}$ , where $\mathcal{I}_{\mathrm{mask}}$ is the set of indices corresponding to the masked documents and $\eta_{m}$ is a hyper-parameter that controls the mask ratio. The masked documents are replaced with a special token [mask]. The history after masking is as follows:
+
+$$
+\begin{array}{l} \mathcal {H} _ {u} ^ {\text {m a s k}} = \left[ \hat {d} _ {1}, \hat {d} _ {2}, \dots , \hat {d} _ {N} \right], \\ \hat {d} _ {i} = \left\{ \begin{array}{l l} d _ {i}, & i \notin \mathcal {I} _ {\text {m a s k}}, \\ [ \text {m a s k} ], & i \in \mathcal {I} _ {\text {m a s k}}. \end{array} \right. \\ \end{array}
+$$
+
+
+Figure 3: Contrastive learning for user embedding training.
+
+Document Reorder. We randomly select a sub-sequence $[d_r, d_{r+1}, \ldots, d_{r+L_r-1}]$ of length $L_r = \lfloor \eta_r N \rfloor$ from $\mathcal{H}_u$ , where $\eta_r$ is a hyper-parameter controlling the reorder ratio, and then randomly shuffle the order of the documents within the sub-sequence to obtain $[\hat{d}_r, \hat{d}_{r+1}, \ldots, \hat{d}_{r+L_r-1}]$ . The history after reordering is as follows:
+
+$$
+\mathcal {H} _ {u} ^ {\text {r e o r d e r}} = \left[ d _ {1}, d _ {2}, \dots , \hat {d} _ {r}, \dots , \hat {d} _ {r + L _ {r} - 1}, \dots , d _ {N} \right].
+$$
+
+4.1.3 Contrastive Loss. Each time, we randomly select two data augmentation methods $\mathcal{A}'$ and $\mathcal{A}''$ to generate two different views of $\mathcal{H}_u$ , denoted as $\mathcal{H}_u'$ and $\mathcal{H}_u''$ . Then, using the encoder described in Section 4.1.1, we obtain the user embeddings $\mathbf{e}_u'$ and $\mathbf{e}_u''$ corresponding to the different views. Since $\mathbf{e}_u'$ and $\mathbf{e}_u''$ are obtained through data augmentation of $\mathcal{H}_u$ , they are more similar to each other. Therefore, we treat them as positive samples for each other and use the views generated from the augmented histories of other users in the same batch as negative samples. We then perform contrastive learning using the InfoNCE [28] loss as follows:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {C L}} = - \left[ \log \frac {\exp \left(\cos \left(\mathbf {e} _ {u} ^ {\prime} , \mathbf {e} _ {u} ^ {\prime \prime}\right) / \tau_ {1}\right)}{\sum_ {u ^ {-} \in \mathcal {U} _ {\mathrm {n e g}}} \exp \left(\cos \left(\mathbf {e} _ {u} ^ {\prime} , \mathbf {e} _ {u ^ {-}} ^ {\prime \prime}\right) / \tau_ {1}\right)} \right. \tag {2} \\ \left. + \log \frac {\exp \left(\cos \left(\mathbf {e} _ {u} ^ {\prime} , \mathbf {e} _ {u} ^ {\prime \prime}\right) / \tau_ {1}\right)}{\sum_ {u ^ {-} \in \mathcal {U} _ {\text {n e g}}} \exp \left(\cos \left(\mathbf {e} _ {u ^ {-}} ^ {\prime}, \mathbf {e} _ {u} ^ {\prime \prime}\right) / \tau_ {1}\right)} \right], \\ \end{array}
+$$
+
+where $\tau_{1}$ is the temperature coefficient, $\mathcal{U}_{\mathrm{neg}}$ are the set of randomly sampled in-batch negative samples, and $\cos (\cdot)$ denotes the cosine similarity.
+
+4.1.4 Top- $m$ User Retrieval. After training with contrastive learning, we can use the encoder from Section 4.1.1 to obtain the user embedding $\mathbf{e}_u$ . We then calculate the cosine similarity between each pair of user embeddings and retrieve the top- $m$ most similar users $\mathcal{U}_{\mathrm{retrieved}} = \{u_1, u_2, \dots, u_m\}$ for user $u$ . Subsequently, the histories of these $m$ users will be used for further document retrieval.
+
+# 4.2 Document Retrieval
+
+After retrieving the top- $m$ users, we design a personalized retriever to retrieve the top- $k$ documents from each user's history, resulting in a total of $m \times k$ candidate documents $\mathcal{D}_{\text{retrieved}} = \{d_{i,j} | i \in \{1, \ldots, m\}, j \in \{1, \ldots, k\}\}$ . This section introduces how the retriever is designed and how it's trained to retrieve documents that better align with the requirements of personalized LLM generation.
+
+4.2.1 Retriever. First, we use a pre-trained dense retrieval model (such as BGE retriever [45]) to compute the semantic relevance
+
+
+Figure 4: The method of training the retriever and reranker using LLM feedback.
+
+between the query and the candidate documents:
+
+$$
+S _ {q, d} ^ {\text {r e t r i e v e r}} = \cos \left(\operatorname {E n c o d e r} _ {q} (q), \operatorname {E n c o d e r} _ {d} (d)\right), \tag {3}
+$$
+
+where $\mathrm{Encoder}_q(\cdot)\to \mathbb{R}^d$ and $\mathrm{Encoder}_d(\cdot)\rightarrow \mathbb{R}^d$ are the encoders for the query and the document in the retrieval model, respectively. Pre-trained retrieval models typically use $S_{q,d}^{\mathrm{retriever}}$ directly for retrieval. However, $S_{q,d}^{\mathrm{retriever}}$ only considers the semantic relevance between the query and the document. Since different users might input the same query but expect different outputs due to their varying preferences, we further account for user personalization by calculating the preference score of the user for the document as follows:
+
+$$
+S _ {u, d} ^ {\text {r e t r i e v e r}} = \cos \left(\mathrm {M L P} _ {1} \left(\mathbf {e} _ {u}\right), \operatorname {E n c o d e r} _ {d} (d)\right), \tag {4}
+$$
+
+where $\mathrm{MLP}_1: \mathbb{R}^d \to \mathbb{R}^d$ is a multi-layer perceptron that maps the user embedding to the space where the cosine similarity is computed. $\mathbf{e}_u$ is the embedding obtained in Section 4.1.1. The total score for retrieval is computed as follows:
+
+$$
+S _ {u, q, d} ^ {\text {r e t r i e v e r}} = (1 - \alpha) S _ {q, d} ^ {\text {r e t r i e v e r}} + \alpha S _ {u, d} ^ {\text {r e t r i e v e r}}, \tag {5}
+$$
+
+where $\alpha$ is a hyper-parameter that controls the weight of personalization.
+
+4.2.2 Training. Since the pre-trained dense retrieval model is not fine-tuned for our specific task, the retrieved results may not necessarily lead to LLM responses that better match the target output $y$ [25, 35]. However, there is no ground truth indicating which documents are better. Therefore, we evaluate the difference between the LLM's output and the target output $y$ , using this as a label to train the retrieval model. Figure 4 shows the process of training the retriever using LLM feedback.
+
+Specifically, we first use the pre-trained retrieval model to retrieve the top- $k$ documents from each of the $m$ users' histories based on $S_{q,d}^{\mathrm{retriever}}$ in Eq. (3), resulting in a total of $m \times k$ candidate documents. These documents are then concatenated with the query one by one and used as prompts for the LLM, producing $m \times k$ outputs:
+
+$$
+\{O _ {q, d _ {i, j}} = \mathrm {L L M} (q, d _ {i, j}) | i \in \{1, \dots , m \}, j \in \{1, \dots , k \} \},
+$$
+
+where $\mathrm{LLM}(q, d_{i,j})$ represents the output generated by inputting the concatenated query $q$ and document $d_{i,j}$ into the LLM. Then, based on the quality of these outputs, we can calculate the distribution of
+
+these candidate documents as follows:
+
+$$
+p _ {\text {L L M}} \left(d _ {i, j} \mid q, y\right) = \frac {\exp (\operatorname {e v a l} \left(y , O _ {q , d _ {i , j}}\right))}{\sum_ {i = 1} ^ {m} \sum_ {j = 1} ^ {k} \exp (\operatorname {e v a l} \left(y , O _ {q , d _ {i , j}}\right))}, \tag {6}
+$$
+
+where $\text{eval}(\cdot)$ measures the difference between the target output $y$ and the LLM's output, using metrics such as ROUGE [24] score. A larger value returned by $\text{eval}(\cdot)$ indicates a better-generated result. Similarly, we can also calculate the score distribution of the candidate documents by the retrieval model based on $S_{u,q,d}^{\text{retriever}}$ in Eq. (5):
+
+$$
+p _ {\text {r e t r i e v e r}} \left(d _ {i, j} \mid q, u\right) = \frac {\exp \left(S _ {u , q , d _ {i , j}} ^ {\text {r e t r i e v e r}}\right)}{\sum_ {i = 1} ^ {m} \sum_ {j = 1} ^ {k} \exp \left(S _ {u , q , d _ {i , j}} ^ {\text {r e t r i e v e r}}\right)}. \tag {7}
+$$
+
+We aim for the retrieval model to retrieve documents that lead to better LLM-generated results, which means making the distribution $p_{\mathrm{retriever}}(d|q,u)$ in Eq. (7) closer to the distribution $p_{\mathrm{LLM}}(d|q,y)$ in Eq (6). Therefore, we compute the KL divergence between the two distributions as the loss to optimize the retriever:
+
+$$
+\mathcal {L} _ {\text {r e t r i e v e r}} = \mathrm {K L} \left(p _ {\text {r e t r i e v e r}} (d | q, u) \mid p _ {\text {L L M}} (d | q, y)\right). \tag {8}
+$$
+
+# 4.3 Document Rerank
+
+After retrieving $\mathcal{D}_{\mathrm{retrieved}}$ through the retriever, in this section, we further refine the results by reranking $\mathcal{D}_{\mathrm{retrieved}}$ to obtain the final top- $k$ ranked results $\mathcal{D}_{\mathrm{reranked}} = \{d_i | i \in \{1, \dots, k\}\}$ .
+
+4.3.1 Reranker. We use a pre-trained cross-encoder (such as the BGE reranker [45]) to encode the query and document, obtaining the hidden state corresponding to the [CLS] token from the last layer:
+
+$$
+\mathbf {h} _ {q, d} = \operatorname {C r o s s E n c o d e r} (q, d), \tag {9}
+$$
+
+where $\mathbf{h}_{q,d} \in \mathbb{R}^d$ . Similarly, when reranking, in addition to considering the semantic relevance between query and document, we also take into account the user's personalized preferences. However, since the cross-encoder does not encode documents separately, it cannot compute the cosine similarity between users and documents as shown in Eq. (4) to express the user preference score. Therefore, we directly concatenate the user embeddings to the output of the cross-encoder to account for the influence of user preferences. The overall score used for reranking is calculated as follows:
+
+$$
+S _ {u, q, d} ^ {\text {r e r a n k e r}} = \mathrm {M L P} _ {3} \left(\operatorname {C O N C A T} \left(\mathbf {h} _ {q, d}, \operatorname {M L P} _ {2} (\mathbf {e} _ {u})\right)\right), \tag {10}
+$$
+
+where $\mathrm{MLP}_2: \mathbb{R}^d \to \mathbb{R}^d$ and $\mathrm{MLP}_3: \mathbb{R}^{2d} \to \mathbb{R}$ are two multi-layer perceptions. $\mathrm{CONCAT}(\cdot)$ denotes the concatenation operation.
+
+4.3.2 Training. Similar to the retriever's training in Section 4.2.2, we also want the reranker to assign higher scores to the documents that lead to better LLM-generated results. Therefore, we train the reranker using a similar approach.
+
+We use the trained retrieval model from Section 4.2.2 to retrieve top- $k$ documents from the history of each of the $m$ users, resulting in a total of $m \times k$ candidate documents. These documents are concatenated with the query $q$ and used as prompts for the LLM, producing $m \times k$ outputs. Similar to Eq.(6), we can obtain the distribution $p_{\mathrm{LLM}}(d|q,y)$ of these candidate documents. Based on
+
+Table 1: Statistics of the datasets used in this paper.
+
+| Task | Input | Output | User History |
| LaMP-1 | The historical profiles are as follows: {history1} ... {historyk}.
+Based on the historical profiles provided, please choose one of the following two references that is more relevant to the user's input title: [1] {reference1}; [2] {reference2}. Please just answer with “[1” or “[2” without explanation. “title”: {title}. | [1] | “title”: {title}
+“abstract”: {abstract} |
| LaMP-2 | The historical profiles are as follows: {history1} ... {historyk}.
+Based on the historical profiles provided, please select the tag from [sci-fi, based on a book, comedy ... ] that is most relevant to the user's input description. Please just answer with the tag name without explanation. “description”: {description}; “tag”: | comedy | “description”: {description};
+“tag”: {tag} |
| LaMP-3 | The historical profiles are as follows: {history1} ... {historyk}.
+Based on the historical profiles provided, what is the score of the following review on a scale of 1 to 5? just answer with 1, 2, 3, 4, or 5 without further explanation. “review”: {review}; “score”: | 5 | “review”: {review}
+“score”: {score} |
| LaMP-4 | The historical profiles are as follows: {history1} ... {historyk}.
+Based on the historical profiles provided, please generate a title for the given user's input text. Please generate it in the following format: {"title": "generated title"} without explanation, and use only English. “text”: {text}; “title”: | {“title”: Finding Happiness
+After Divorce - It Can Happen} | “text”: {text}
+“title”: {title} |
| LaMP-5 | The historical profiles are as follows: {history1} ... {historyk}.
+Based on the historical profiles provided, please generate a title for the given user's input abstract. Please generate it in the following format: {"title": "generated title"} without explanation, and use only English. “abstract”: {abstract}; “title”: | {“title”: Link-Reliability Based
+Two-Hop Routing for
+Wireless Sensor Networks.} | “abstract”: {abstract}
+“title”: {title} |
| LaMP-7 | The historical profiles are as follows: {history1} ... {historyk}.
+Based on the style pattern of the historical tweets provided, please paraphrase the user's input tweet without any explanation before or after it. Please generate it in the following format: {"tweet": "generated tweet"} without explanation, and use only English. “tweet”: {tweet}. | {“tweet”:lilxcutiesworld the
+danny picture is GOOD!!
+I really like it.} | “tweet”: {tweet} |
+
+5.4.4 Impact of the Number of Retrieved Users. Since we enhance personalized text generation by introducing collaborative filtering, we further explored how much collaborative information to introduce, specifically the impact of the number of retrieved users on the results, as shown in Figure 8. In LaMP-1, retrieving too few or too many users leads to poorer performance, with the best results at 4 users. In LaMP-5, the performance improves as the number of users increases. This highlights the importance of introducing collaborative filtering, but it also indicates that excessive introduction can lead to decreased effectiveness.
+
+5.4.5 Impact of the Number of Retrieved Documents. We also analyzed the impact of the number of retrieved documents, $k$ , on the results, as shown in Figure 9. It can be observed that as the number of retrieved documents increases, performance improves, indicating the importance of retrieving user history to reflect user preferences for enhancing LLM-generated results. Since more documents lead to longer prompts and slower LLM generation, we chose $k = 5$ for our experiments.
+
+# 6 Conclusion
+
+In this paper, we propose CFRAG, which adapts collaborative filtering into RAG to personalize LLMs. To introduce collaborative information without explicit user labels and retrieve documents that support personalized LLM generation, we first train user embeddings through contrastive learning to retrieve similar users. Then, we design the personalized retriever and reranker that considers user preferences during retrieval and fine-tune them using LLM feedback. The results on the Language Model Personalization (LaMP) benchmark validate the effectiveness of CFRAG. The experimental analysis also confirms the effectiveness of each module within CFRAG.
+
+# A Appendix: Prompts
+
+We provide detailed formats for the inputs, outputs, and user histories for the LLM across different datasets, as shown in Table 4.
+
+# References
+
+[1] AI@Meta. 2024. Llama 3 Model Card. (2024). https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md
+[2] Akari Asai, Zegiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. [n.d.]. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. In The Twelfth International Conference on Learning Representations.
+[3] Sebastian Borgeaud, Arthur Mensch, et al. 2022. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, PMLR, 2206-2240.
+[4] Jin Chen, Zheng Liu, et al. 2024. When large language models meet personalization: Perspectives of challenges and opportunities. World Wide Web 27, 4 (2024), 42.
+[5] Sunhao Dai, Ninglu Shao, et al. 2023. Uncovering chatgpt's capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems. 1126-1132.
+[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers).
+[7] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 6491-6501.
+[8] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023).
+[9] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. In International conference on machine learning. PMLR, 3929-3938.
+[10] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639-648.
+[11] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web. 173-182.
+[12] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. [n.d.]. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations.
+[13] Gautier Izacard and Edouard Grave. 2021. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 874-880.
+[14] Gautier Izacard, Patrick Lewis, et al. 2022. Few-shot learning with retrieval augmented language models. arXiv preprint arXiv:2208.03299 1, 2 (2022), 4.
+[15] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. 2020. A survey on contrastive self-supervised learning. Technologies 9, 1 (2020), 2.
+[16] Joel Jang, Seungone Kim, et al. 2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv preprint arXiv:2310.11564 (2023).
+[17] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In International Conference on Machine Learning. PMLR, 15696-15707.
+[18] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).
+[19] Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer 42, 8 (2009), 30-37.
+[20] Patrick Lewis, Ethan Perez, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33 (2020), 9459-9474.
+[21] Cheng Li, Mingyang Zhang, Qiaozhu Mei, Yaqing Wang, Spurthi Amba Hombaiah, Yi Liang, and Michael Bendersky. 2023. Teach LLMs to Personalize-An Approach inspired by Writing Education. arXiv preprint arXiv:2308.07968 (2023).
+[22] Junyi Li, Tianyi Tang, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2024. Pre-trained language models for text generation: A survey. Comput. Surveys 56, 9 (2024), 1-39.
+[23] Xinyu Li, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133 (2024).
+[24] Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out. 74-81.
+[25] Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. [n.d.]. RA-DIT: Retrieval-Augmented Dual Instruction Tuning. In The Twelfth International Conference on Learning Representations.
+[26] Yinhan Liu. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019).
+
+[27] Sheshera Mysore, Zhuoran Lu, et al. 2023. Pearl: Personalizing large language model writing assistants with generation-calibrated retrievers. arXiv preprint arXiv:2311.09180 (2023).
+[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018).
+[29] Chris Richardson, Yao Zhang, Kellen Gillespie, Sudipta Kar, Arshdeep Singh, Zeynab Raeesy, Omar Zia Khan, and Abhinav Sethy. 2023. Integrating summarization and retrieval for enhanced personalization via large language models. arXiv preprint arXiv:2310.20081 (2023).
+[30] Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp 109 (1995), 109.
+[31] Alireza Salemi, Surya Kallumadi, and Hamed Zamani. 2024. Optimization methods for personalizing large language models through retrieval augmentation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 752-762.
+[32] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2023. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406 (2023).
+[33] Chenglei Shen, Xiao Zhang, Teng Shi, Changshuo Zhang, Guofu Xie, and Jun Xu. 2024. A survey of controllable learning: Methods and applications in information retrieval. arXiv preprint arXiv:2407.06083 (2024).
+[34] Teng Shi, Zihua Si, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Dewei Leng, Yanan Niu, and Yang Song. 2024. UniSAR: Modeling User Transition Behaviors between Search and Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1029-1039.
+[35] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2024. REPLUG: Retrieval-Augmented Black-Box Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 8364-8377.
+[36] Zhaoxuan Tan, Zheyuan Liu, and Meng Jiang. 2024. Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts. arXiv preprint arXiv:2406.10471 (2024).
+[37] Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2024. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. arXiv:2402.04401 [cs.CL] https://arxiv.org/abs/2402.04401
+[38] Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Wu Jian, and Yuning Jiang. 2025. Think Before Recommend: Unleashing the Latent Reasoning Power for Sequential Recommendation. arXiv:2503.22675 [cs.IR] https://arxiv.org/abs/2503.22675
+[39] A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017).
+[40] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval. 165-174.
+[41] Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al. 2024. A survey on large language models for recommendation. World Wide Web 27, 5 (2024), 60.
+[42] Xinghao Wu, Xuefeng Liu, Jianwei Niu, Haolin Wang, Shaojie Tang, and Guogang Zhu. 2024. FedLoRA: When Personalized Federated Learning Meets Low-Rank Adaptation. (2024).
+[43] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. 2024. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems 36 (2024).
+[44] Zhuofeng Wu, Sinong Wang, Jiatao Gu, Madian Khabsa, Fei Sun, and Hao Ma. 2020. Clear: Contrastive learning for sentence representation. arXiv preprint arXiv:2012.15466 (2020).
+[45] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597 [cs.CL]
+[46] Hong-Jian Xue, Xinyu Dai, Jianbing Zhang, Shujian Huang, and Jiajun Chen. 2017. Deep matrix factorization models for recommender systems.. In IJCAI, Vol. 17. Melbourne, Australia, 3203-3209.
+[47] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng-peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671 (2024).
+[48] Changshuo Zhang, Teng Shi, Xiao Zhang, Qi Liu, Ruobing Xie, Jun Xu, and Ji-Rong Wen. 2024. Modeling Domain and Feedback Transitions for Cross-Domain Sequential Recommendation. arXiv preprint arXiv:2408.08209 (2024).
+[49] Changshuo Zhang, Teng Shi, Xiao Zhang, Yanping Zheng, Ruobing Xie, Qi Liu, Jun Xu, and Ji-Rong Wen. 2024. QAGCF: Graph Collaborative Filtering for Q&A Recommendation. arXiv preprint arXiv:2406.04828 (2024).
+[50] Changshuo Zhang, Xiao Zhang, Teng Shi, Jun Xu, and Ji-Rong Wen. 2025. Test-Time Alignment for Tracking User Interest Shifts in Sequential Recommendation. arXiv:2504.01489 [cs.IR] https://arxiv.org/abs/2504.01489
+
+[51] Kepu Zhang, Teng Shi, Sunhao Dai, Xiao Zhang, Yinfeng Li, Jing Lu, Xiaoxue Zang, Yang Song, and Jun Xu. 2024. SAQRec: Aligning Recommender Systems to User Satisfaction via Questionnaire Feedback. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 3165-3175.
+[52] Xiao Zhang, Teng Shi, Jun Xu, Zhenhua Dong, and Ji-Rong Wen. 2024. Model-Agnostic Causal Embedding Learning for Counterfactually Group-Fair Recommendation. IEEE Transactions on Knowledge and Data Engineering (2024).
+[53] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219 (2023).
+
+[54] Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2024. Dense text retrieval based on pretrained language models: A survey. ACM Transactions on Information Systems 42, 4 (2024), 1-60.
+[55] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223 (2023).
+[56] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107 (2023).
+[57] Yuchen Zhuang, Haotian Sun, Yue Yu, Qifan Wang, Chao Zhang, and Bo Dai. 2024. HYDRA: Model Factorization Framework for Black-Box LLM Personalization. arXiv preprint arXiv:2406.02888 (2024).
\ No newline at end of file
diff --git a/data/2025/2504_05xxx/2504.05731/images/0fe4c17f7d3b0064ee98d9689518a2b9f9a84baea0031c9d564a3783a1b2fd6b.jpg b/data/2025/2504_05xxx/2504.05731/images/0fe4c17f7d3b0064ee98d9689518a2b9f9a84baea0031c9d564a3783a1b2fd6b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..bb2f3e04396201d1ac31af717fe94c35933b9d28
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/0fe4c17f7d3b0064ee98d9689518a2b9f9a84baea0031c9d564a3783a1b2fd6b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fd1c5b5f9da3fc6c8ae41307334f5baadffb7c95871df63a15b9536f90a707c
+size 11210
diff --git a/data/2025/2504_05xxx/2504.05731/images/1770e817342ffc39a1149bbb6148e83683024c61a30a0f540c0cbacac3546bc7.jpg b/data/2025/2504_05xxx/2504.05731/images/1770e817342ffc39a1149bbb6148e83683024c61a30a0f540c0cbacac3546bc7.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ea794584557d7a9c6f6996da28dbc566140f10d2
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/1770e817342ffc39a1149bbb6148e83683024c61a30a0f540c0cbacac3546bc7.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a03578fd08cf32a9f4ed4084a7ca529350c6fea845137d2f1bc40058f08401c
+size 5609
diff --git a/data/2025/2504_05xxx/2504.05731/images/1b4b2debe6472ddf2afda8745e85c61d5617ebc984a7385407045378024caa4a.jpg b/data/2025/2504_05xxx/2504.05731/images/1b4b2debe6472ddf2afda8745e85c61d5617ebc984a7385407045378024caa4a.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..cc7346afaf9d25b739762efa4e6fdff712f84801
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/1b4b2debe6472ddf2afda8745e85c61d5617ebc984a7385407045378024caa4a.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78879fab6889e7b21cfc3a338d5cfcf462416c54db67a4e223c46b89febb49e7
+size 5653
diff --git a/data/2025/2504_05xxx/2504.05731/images/20182a7db51c3fa3024198281989434352d72122a26392fd7dce39da64ecec9d.jpg b/data/2025/2504_05xxx/2504.05731/images/20182a7db51c3fa3024198281989434352d72122a26392fd7dce39da64ecec9d.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..395cf1242e5cb21f5f7dff5354970b1c444fbbf0
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/20182a7db51c3fa3024198281989434352d72122a26392fd7dce39da64ecec9d.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29fc94098fd80f6c0a07f74478a1c730d6d141b217d71adf954cabde8826183b
+size 11735
diff --git a/data/2025/2504_05xxx/2504.05731/images/208072a2dd51268962130922f22ad7a74e36f3ec865e34dc65a393a56cbe2a6c.jpg b/data/2025/2504_05xxx/2504.05731/images/208072a2dd51268962130922f22ad7a74e36f3ec865e34dc65a393a56cbe2a6c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..44c8861b38b9b777077cfcc145b7fc72084527de
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/208072a2dd51268962130922f22ad7a74e36f3ec865e34dc65a393a56cbe2a6c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a242e2ca9efcb6d8e85a2f36b5e11b72ad1b88b97ec64ce37b92ba5ceab02594
+size 11029
diff --git a/data/2025/2504_05xxx/2504.05731/images/286390de4f0085e9bc40dde8cd842fc967aed23ac33c9c8b2fef3cd5ce5de750.jpg b/data/2025/2504_05xxx/2504.05731/images/286390de4f0085e9bc40dde8cd842fc967aed23ac33c9c8b2fef3cd5ce5de750.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..586cb9a156b1d0a2b7f068975bccb75361df9888
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/286390de4f0085e9bc40dde8cd842fc967aed23ac33c9c8b2fef3cd5ce5de750.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:496b8936eab3301919f825ff23f57bd1a4f6d5223333b83ce5ade23c4eaac7a6
+size 4956
diff --git a/data/2025/2504_05xxx/2504.05731/images/2949233b012ca5dc6bde7e95580b620544ae0db053562b8a4fa23243c5566927.jpg b/data/2025/2504_05xxx/2504.05731/images/2949233b012ca5dc6bde7e95580b620544ae0db053562b8a4fa23243c5566927.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..b5bda3507dca67b6287bb54f485f799679279fb5
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/2949233b012ca5dc6bde7e95580b620544ae0db053562b8a4fa23243c5566927.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ff7a02847a7bc5b28d9f31b8ce5ae55daefaca7ffff36ab098dacac636f5b32
+size 24835
diff --git a/data/2025/2504_05xxx/2504.05731/images/2b38bacf7f2f45a0608c219d86ab323366512e80026a8f5314c0af3b68f98d9f.jpg b/data/2025/2504_05xxx/2504.05731/images/2b38bacf7f2f45a0608c219d86ab323366512e80026a8f5314c0af3b68f98d9f.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..7b0ef8d086914ec827de631ced0c687809d5255a
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/2b38bacf7f2f45a0608c219d86ab323366512e80026a8f5314c0af3b68f98d9f.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e05a538b789c69de93fa504dcf1c3bc9bfbcad9e772d4d60df5a1c4a9e6af73
+size 6257
diff --git a/data/2025/2504_05xxx/2504.05731/images/321ed29b3fe1bd2bf8b65455f1b9c4e37c82123640abf0223a5bca357a1a3ec4.jpg b/data/2025/2504_05xxx/2504.05731/images/321ed29b3fe1bd2bf8b65455f1b9c4e37c82123640abf0223a5bca357a1a3ec4.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..4ef9c5949195064055e808bc9cc8053b6ef60aad
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/321ed29b3fe1bd2bf8b65455f1b9c4e37c82123640abf0223a5bca357a1a3ec4.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ad6740d32a796934aab711abb45d44f24bcf66adf17aed0bd0f30afc6c3a85f
+size 3412
diff --git a/data/2025/2504_05xxx/2504.05731/images/34f15c2d425a0ba3e0b25f67abb59aecc0f185aae42a557093eae9d488fd5f26.jpg b/data/2025/2504_05xxx/2504.05731/images/34f15c2d425a0ba3e0b25f67abb59aecc0f185aae42a557093eae9d488fd5f26.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..9cc06e41e5f6b3a351e6699dba767419362dabfc
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/34f15c2d425a0ba3e0b25f67abb59aecc0f185aae42a557093eae9d488fd5f26.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:412ba66e5c5f8b126bea2ecf73a01e103f9ae70d6afa417de2271869fe77fdfc
+size 78571
diff --git a/data/2025/2504_05xxx/2504.05731/images/370a5924ecdacae48795859736686a99e261ba862019de8a358e02e9a7b90220.jpg b/data/2025/2504_05xxx/2504.05731/images/370a5924ecdacae48795859736686a99e261ba862019de8a358e02e9a7b90220.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ff12f7fc7ccd5a177ec45437e0c61602afc06739
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/370a5924ecdacae48795859736686a99e261ba862019de8a358e02e9a7b90220.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7604a9b6435be28a7f6962658e1a553bc9745ea823151c6587f76df2cb5a3443
+size 5494
diff --git a/data/2025/2504_05xxx/2504.05731/images/45032d0b3f912fe033f347f966ca5cdbdf24f9320b31a6df192191757e19f649.jpg b/data/2025/2504_05xxx/2504.05731/images/45032d0b3f912fe033f347f966ca5cdbdf24f9320b31a6df192191757e19f649.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c8f923a4a9df8a6231f110e32c68d0662dccf550
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/45032d0b3f912fe033f347f966ca5cdbdf24f9320b31a6df192191757e19f649.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02e9036ff8fded88eb193539aa0dc28431fc796d05d4d0fcd84e1f5970674c8b
+size 5846
diff --git a/data/2025/2504_05xxx/2504.05731/images/4d1923b82576cd1772eb0abb7c69d9b2c56470581b24e15f43d202ce4b7dd758.jpg b/data/2025/2504_05xxx/2504.05731/images/4d1923b82576cd1772eb0abb7c69d9b2c56470581b24e15f43d202ce4b7dd758.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..0a6d189ecb9add7fd80f90bcf16c939d2dc07cdb
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/4d1923b82576cd1772eb0abb7c69d9b2c56470581b24e15f43d202ce4b7dd758.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3b87f97aee29aeb8d9de5924b3f2ea3f60d04d331780301cdbaec9e3e11113e
+size 84307
diff --git a/data/2025/2504_05xxx/2504.05731/images/4de643f3cc500517a2b74f1c2f8233ac753bd2f716de8d3e58deb45f124dd930.jpg b/data/2025/2504_05xxx/2504.05731/images/4de643f3cc500517a2b74f1c2f8233ac753bd2f716de8d3e58deb45f124dd930.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..5e809eba8b6def8be6e1907e7a22a8e550600718
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/4de643f3cc500517a2b74f1c2f8233ac753bd2f716de8d3e58deb45f124dd930.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f064563940284d55aa6c8916bab75e9989e46531fc15271414653411ed586a7
+size 5529
diff --git a/data/2025/2504_05xxx/2504.05731/images/553b67f5416e9bf7a6397eb8a5e9f5ff9e90f77643a2e2ab2977210184904014.jpg b/data/2025/2504_05xxx/2504.05731/images/553b67f5416e9bf7a6397eb8a5e9f5ff9e90f77643a2e2ab2977210184904014.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..4abeca23e1ed56360a581f9bba8f31f9a80cb081
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/553b67f5416e9bf7a6397eb8a5e9f5ff9e90f77643a2e2ab2977210184904014.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e58dc77f628dc0b0ffd906a5e2e90d441159b51fe177804782d73bbbdd4513b9
+size 5646
diff --git a/data/2025/2504_05xxx/2504.05731/images/5bb368b797bf6324680c0dab35687b3a755828cbb10a15bd627324662e645b92.jpg b/data/2025/2504_05xxx/2504.05731/images/5bb368b797bf6324680c0dab35687b3a755828cbb10a15bd627324662e645b92.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..99b31713f487385e91738a8052afbf8bfa5b7d39
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/5bb368b797bf6324680c0dab35687b3a755828cbb10a15bd627324662e645b92.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b67b13f94cfe85643f3b4aeb68b97df3b3fb159524f8d6399edd2ba3849ca839
+size 9148
diff --git a/data/2025/2504_05xxx/2504.05731/images/625fdaef24bf247580e641a34943670ea15cb8f8da28784bc68a9e6a62dbe508.jpg b/data/2025/2504_05xxx/2504.05731/images/625fdaef24bf247580e641a34943670ea15cb8f8da28784bc68a9e6a62dbe508.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f295b1883e402a9575335472ff113e1973ddef6b
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/625fdaef24bf247580e641a34943670ea15cb8f8da28784bc68a9e6a62dbe508.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36a21a75a5e8bfcb2c92ff8caf4b33eeacc3b533c8a75bb77067d50a604ca3a7
+size 21303
diff --git a/data/2025/2504_05xxx/2504.05731/images/6cea494ce69dd19e4155bbac2bdb81b8052ffdd33a3f5e4661fcc323a21cbc4f.jpg b/data/2025/2504_05xxx/2504.05731/images/6cea494ce69dd19e4155bbac2bdb81b8052ffdd33a3f5e4661fcc323a21cbc4f.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..82d08f5a31135144ca3670eb8f97cfbd7404a9c7
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/6cea494ce69dd19e4155bbac2bdb81b8052ffdd33a3f5e4661fcc323a21cbc4f.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff12ddfe82c96f02d909a0d2bf7d8385bf894da6ce1f654176270dffa3e393a7
+size 235053
diff --git a/data/2025/2504_05xxx/2504.05731/images/6edc98bed93f6b43480ec1d654b337ca409da377fc6ab6198e2c9c1f23130ce8.jpg b/data/2025/2504_05xxx/2504.05731/images/6edc98bed93f6b43480ec1d654b337ca409da377fc6ab6198e2c9c1f23130ce8.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..df9170c2e8c5fefc555d766a28025bd49c496752
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/6edc98bed93f6b43480ec1d654b337ca409da377fc6ab6198e2c9c1f23130ce8.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f391d6482d1ea5b7f0a4e043995ed4e37e2bf033b778b9f3e45c8d4f23d334ca
+size 4712
diff --git a/data/2025/2504_05xxx/2504.05731/images/7082bea53d532e5d94a2453a6bded23a4f6f523ae049809dfc0d039c880c149f.jpg b/data/2025/2504_05xxx/2504.05731/images/7082bea53d532e5d94a2453a6bded23a4f6f523ae049809dfc0d039c880c149f.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..8d014ecfdd9e2920ea5c57dcbfe44e5361a0055c
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/7082bea53d532e5d94a2453a6bded23a4f6f523ae049809dfc0d039c880c149f.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b561e83a200ec6a14c599c14841a2ba132bdb49957d800e53d97811657280cff
+size 123785
diff --git a/data/2025/2504_05xxx/2504.05731/images/79735191f069bd3fe577f6f68843cfd442708060861370a8d9a6cbcee6d995b3.jpg b/data/2025/2504_05xxx/2504.05731/images/79735191f069bd3fe577f6f68843cfd442708060861370a8d9a6cbcee6d995b3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..50c1e9fd19f25616a063372d0e5aff8408654a1e
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/79735191f069bd3fe577f6f68843cfd442708060861370a8d9a6cbcee6d995b3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5299a359cbaea3a498ceaaddac1c64feee5d88e7e490f629c57b257b6aebd364
+size 11428
diff --git a/data/2025/2504_05xxx/2504.05731/images/81368817db3e34bbb1db188fa57860badccf1c91f22e28f4651cf117e8756cd9.jpg b/data/2025/2504_05xxx/2504.05731/images/81368817db3e34bbb1db188fa57860badccf1c91f22e28f4651cf117e8756cd9.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f3756c63e602124c4510a329c5a4749a2394f2ea
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/81368817db3e34bbb1db188fa57860badccf1c91f22e28f4651cf117e8756cd9.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57d31dc4afd505c86d6a68821f74ba5fed53e82b0be1cc000258a29ea450159f
+size 9595
diff --git a/data/2025/2504_05xxx/2504.05731/images/83c00f18c2b907f8f9980e476359f97f132ed1a1ca8fa078b88d50e7d5d7f922.jpg b/data/2025/2504_05xxx/2504.05731/images/83c00f18c2b907f8f9980e476359f97f132ed1a1ca8fa078b88d50e7d5d7f922.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f24cb4e2de81a2b7e6dddbccd0b5753d53749686
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/83c00f18c2b907f8f9980e476359f97f132ed1a1ca8fa078b88d50e7d5d7f922.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61b5ec95253218d57033e26f6f43bdd5e93f488010064fa6350419950aeae8e3
+size 9603
diff --git a/data/2025/2504_05xxx/2504.05731/images/83cde8aa2ad4ce4f601933f54a812e13703cb17f9d80156a716ccbf53ac36f6e.jpg b/data/2025/2504_05xxx/2504.05731/images/83cde8aa2ad4ce4f601933f54a812e13703cb17f9d80156a716ccbf53ac36f6e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3686fe3bb832a3bbec7ddba6acc0106ce7f545ec
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/83cde8aa2ad4ce4f601933f54a812e13703cb17f9d80156a716ccbf53ac36f6e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6373c04d5a3a2d29649a7774d57dacc1eecb216c8adee51a2a6487607a95b15e
+size 10362
diff --git a/data/2025/2504_05xxx/2504.05731/images/86faea8a327cef5e2b9ac66cf7e276f3ee4ebff4fccc9a5f5db5d20515b61ee6.jpg b/data/2025/2504_05xxx/2504.05731/images/86faea8a327cef5e2b9ac66cf7e276f3ee4ebff4fccc9a5f5db5d20515b61ee6.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..23aeb184f531085413c49715bb5f4186c143e7df
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/86faea8a327cef5e2b9ac66cf7e276f3ee4ebff4fccc9a5f5db5d20515b61ee6.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5cabda29cf34800bcc8c1370922a1aadeed515b26ec0fa1ea37c6135aebd8f73
+size 10450
diff --git a/data/2025/2504_05xxx/2504.05731/images/92815eb6c2f51b4c9ef6ca63b04dd6a8f815dc0d32224bff35a685479dc7ee07.jpg b/data/2025/2504_05xxx/2504.05731/images/92815eb6c2f51b4c9ef6ca63b04dd6a8f815dc0d32224bff35a685479dc7ee07.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..bf5385e47aaf97db05baa8b73d896277a98959d7
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/92815eb6c2f51b4c9ef6ca63b04dd6a8f815dc0d32224bff35a685479dc7ee07.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0651bdfdc890f5c4ea9da45f37cfb78cd4af746d13ba0733a7ed41413f1a8cd9
+size 25603
diff --git a/data/2025/2504_05xxx/2504.05731/images/94633987414c00f738c3423cbe8c6e65847b434df8ed4ae5565aeb3a005e42e8.jpg b/data/2025/2504_05xxx/2504.05731/images/94633987414c00f738c3423cbe8c6e65847b434df8ed4ae5565aeb3a005e42e8.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c7ab35375c80b6b7acb55ac5a6d9ee9018257812
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/94633987414c00f738c3423cbe8c6e65847b434df8ed4ae5565aeb3a005e42e8.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96957d044ed61f92b6e4be5e8e90b85ba386b7caf9507e08d1e70ba65cbc6310
+size 10759
diff --git a/data/2025/2504_05xxx/2504.05731/images/9976451dd3eec8608f9caf953988cee5f27caa57ccb7db4f2daace3708385103.jpg b/data/2025/2504_05xxx/2504.05731/images/9976451dd3eec8608f9caf953988cee5f27caa57ccb7db4f2daace3708385103.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c83851ed968ec225caa86418536271422662554d
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/9976451dd3eec8608f9caf953988cee5f27caa57ccb7db4f2daace3708385103.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:542435d6328e6f412f6a5a3d6d9289f74f464d87275b683595303afc79ef6c4a
+size 15813
diff --git a/data/2025/2504_05xxx/2504.05731/images/9c481f693eb54eb4efbf0d5c0f2af5aa8f4d150f073794b47f3ad1f2fb237cf5.jpg b/data/2025/2504_05xxx/2504.05731/images/9c481f693eb54eb4efbf0d5c0f2af5aa8f4d150f073794b47f3ad1f2fb237cf5.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..20baa4da6277ea23f8ee591e025252beb737b004
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/9c481f693eb54eb4efbf0d5c0f2af5aa8f4d150f073794b47f3ad1f2fb237cf5.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a8c66c86fe572718a97102b832be77e76e8133ac53e603c9ebf5ee7f9d9c873
+size 10890
diff --git a/data/2025/2504_05xxx/2504.05731/images/9ff4e06204e6c48c44c4467eca5a18d03e07c6da14c61020e7c1603ef065ab96.jpg b/data/2025/2504_05xxx/2504.05731/images/9ff4e06204e6c48c44c4467eca5a18d03e07c6da14c61020e7c1603ef065ab96.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..10ba3cd44ab2343fe84fdb0b6e3ea8aca63edfec
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/9ff4e06204e6c48c44c4467eca5a18d03e07c6da14c61020e7c1603ef065ab96.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca4a568a3f89a371867f4ac5be8918b4f93b8bfb915865129509e7fd51c7e986
+size 2923
diff --git a/data/2025/2504_05xxx/2504.05731/images/a6b102852714ef7badacf8034fddfa94795403b0b6339e25278197262aa0cdab.jpg b/data/2025/2504_05xxx/2504.05731/images/a6b102852714ef7badacf8034fddfa94795403b0b6339e25278197262aa0cdab.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..9ab2dce0cab9f8db31e446ebfbd42df67458d643
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/a6b102852714ef7badacf8034fddfa94795403b0b6339e25278197262aa0cdab.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb13b24718d1ded6f7fd05d12266e5d49cb502a4561f94523d7a81e9f99c1d78
+size 3524
diff --git a/data/2025/2504_05xxx/2504.05731/images/acb69f3db4e448b0e25bc42ebbf3669d4e4b97072a06599b923f91b4cb5074dc.jpg b/data/2025/2504_05xxx/2504.05731/images/acb69f3db4e448b0e25bc42ebbf3669d4e4b97072a06599b923f91b4cb5074dc.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..5ced79633e3bafec3dcd448db11008608c9a32ec
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/acb69f3db4e448b0e25bc42ebbf3669d4e4b97072a06599b923f91b4cb5074dc.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07acca60c3b530601f62c5295c0dd574f9ff6da271e6d1b31509deb6f29f65d4
+size 11365
diff --git a/data/2025/2504_05xxx/2504.05731/images/c145d4b73d48d432adc24495509ff5b7dc143be4525fd4b84a31137fa6d78e46.jpg b/data/2025/2504_05xxx/2504.05731/images/c145d4b73d48d432adc24495509ff5b7dc143be4525fd4b84a31137fa6d78e46.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..6cf8fd1c3de5a4ca0ee975155be878709afda870
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/c145d4b73d48d432adc24495509ff5b7dc143be4525fd4b84a31137fa6d78e46.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89896b153f6f692775376a0510cf0b23c0e729d9afdf6accf5e4fa52b92089aa
+size 97148
diff --git a/data/2025/2504_05xxx/2504.05731/images/e096674f0f80370de56434a59eb14a7c664e0ab43e4a24769d974004128917ef.jpg b/data/2025/2504_05xxx/2504.05731/images/e096674f0f80370de56434a59eb14a7c664e0ab43e4a24769d974004128917ef.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a3bf4042c65ae1a82e3f0690e3a794579e4416d6
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/e096674f0f80370de56434a59eb14a7c664e0ab43e4a24769d974004128917ef.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d677ffbd065529fcc9187358746ed9bd8c7afc8feaf899b37dda22a8707c1c2c
+size 10977
diff --git a/data/2025/2504_05xxx/2504.05731/images/ec0d097f9eaf257d62428318090b463fd41bd34756cf25a311d0f4e5debaae76.jpg b/data/2025/2504_05xxx/2504.05731/images/ec0d097f9eaf257d62428318090b463fd41bd34756cf25a311d0f4e5debaae76.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..fc1f67c241e35f33fdddde307ab6ff470ed479c0
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/ec0d097f9eaf257d62428318090b463fd41bd34756cf25a311d0f4e5debaae76.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b78df4ade1437ce8984c9881bb6ab2ce3e7e7748b471c0770c89cea4ed809918
+size 3839
diff --git a/data/2025/2504_05xxx/2504.05731/images/f519fd19f5d090c02b1805b4837891bd2fdbf2eba9b80e44d9e885596ab1fdf6.jpg b/data/2025/2504_05xxx/2504.05731/images/f519fd19f5d090c02b1805b4837891bd2fdbf2eba9b80e44d9e885596ab1fdf6.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c58f48c006f045ef39dfb63b3b40b3e3f8c121ca
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/f519fd19f5d090c02b1805b4837891bd2fdbf2eba9b80e44d9e885596ab1fdf6.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d742c736e0290d5e92c2725aef4de16c98cd0481977b9924b116f0a8aee878ef
+size 5404
diff --git a/data/2025/2504_05xxx/2504.05731/images/fa1ad7982cf1c2b2fc0d929344eb06b13bcbc11bbbb85e0c5d98093ed5b060b1.jpg b/data/2025/2504_05xxx/2504.05731/images/fa1ad7982cf1c2b2fc0d929344eb06b13bcbc11bbbb85e0c5d98093ed5b060b1.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..73f22037d4cbc5154693bd1b489cfebab47e245c
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/images/fa1ad7982cf1c2b2fc0d929344eb06b13bcbc11bbbb85e0c5d98093ed5b060b1.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18289802d5238cc87c76ac1a0aa1edf20961e2e34fb18a652b24310d5ba5d6da
+size 7264
diff --git a/data/2025/2504_05xxx/2504.05731/layout.json b/data/2025/2504_05xxx/2504.05731/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..1e2672f2c94cea762f2e2fb57d69767e10227749
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05731/layout.json
@@ -0,0 +1,12949 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 52,
+ 79,
+ 559,
+ 118
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 79,
+ 559,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 79,
+ 559,
+ 118
+ ],
+ "type": "text",
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 159,
+ 125,
+ 205,
+ 137
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 159,
+ 125,
+ 205,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 125,
+ 205,
+ 137
+ ],
+ "type": "text",
+ "content": "Teng Shi"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 124,
+ 139,
+ 241,
+ 150
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 124,
+ 139,
+ 241,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 139,
+ 241,
+ 150
+ ],
+ "type": "text",
+ "content": "Renmin University of China"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 153,
+ 151,
+ 212,
+ 162
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 153,
+ 151,
+ 212,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 153,
+ 151,
+ 212,
+ 162
+ ],
+ "type": "text",
+ "content": "Beijing, China"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 141,
+ 163,
+ 223,
+ 174
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 163,
+ 223,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 163,
+ 223,
+ 174
+ ],
+ "type": "text",
+ "content": "shiteng@ruc.edu.cn"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 146,
+ 197,
+ 217,
+ 210
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 197,
+ 217,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 197,
+ 217,
+ 210
+ ],
+ "type": "text",
+ "content": "Xiaoxue Zang"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 156,
+ 211,
+ 208,
+ 224
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 156,
+ 211,
+ 208,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 156,
+ 211,
+ 208,
+ 224
+ ],
+ "type": "text",
+ "content": "Kai Zheng"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 119,
+ 224,
+ 246,
+ 236
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 224,
+ 246,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 224,
+ 246,
+ 236
+ ],
+ "type": "text",
+ "content": "Kuaishou Technology Co., Ltd."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 152,
+ 236,
+ 211,
+ 247
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 152,
+ 236,
+ 211,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 152,
+ 236,
+ 211,
+ 247
+ ],
+ "type": "text",
+ "content": "Beijing, China"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 145,
+ 248,
+ 219,
+ 258
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 248,
+ 219,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 248,
+ 219,
+ 258
+ ],
+ "type": "text",
+ "content": "xxic666@126.com"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 136,
+ 260,
+ 228,
+ 272
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 136,
+ 260,
+ 228,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 136,
+ 260,
+ 228,
+ 272
+ ],
+ "type": "text",
+ "content": "zhengk92@gmail.com"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 51,
+ 279,
+ 96,
+ 290
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 279,
+ 96,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 279,
+ 96,
+ 290
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 50,
+ 293,
+ 295,
+ 568
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 293,
+ 295,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 293,
+ 295,
+ 568
+ ],
+ "type": "text",
+ "content": "Recently, the personalization of Large Language Models (LLMs) to generate content that aligns with individual user preferences has garnered widespread attention. Personalized Retrieval-Augmented Generation (RAG), which retrieves relevant documents from the user's history to reflect their preferences and enhance LLM generation, is one commonly used approach for personalization. However, existing personalized RAG methods do not consider that the histories of similar users can also assist in personalized generation for the current user, meaning that collaborative information between users can also benefit personalized generation. Inspired by the application of collaborative filtering in recommender systems, we propose a method called CFRAG, which adapts Collaborative Filtering to RAG for personalized text generation. However, this presents two challenges: (1) how to incorporate collaborative information without explicit user similarity labels? (2) how to retrieve documents that support personalized LLM generation? For Challenge 1, we use contrastive learning to train user embeddings to retrieve similar users and introduce collaborative information. For Challenge 2, we design a personalized retriever and reranker to retrieve the top-"
+ },
+ {
+ "bbox": [
+ 50,
+ 293,
+ 295,
+ 568
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 50,
+ 293,
+ 295,
+ 568
+ ],
+ "type": "text",
+ "content": " documents from these users' histories. We take into account the user's preference during retrieval and reranking. Then we leverage feedback from the LLM to fine-tune the personalized retriever and reranker, enabling them to retrieve documents that meet the personalized generation needs of the LLM. Experimental results on the Language Model Personalization (LaMP) benchmark"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 50,
+ 579,
+ 295,
+ 605
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 579,
+ 295,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 579,
+ 295,
+ 605
+ ],
+ "type": "text",
+ "content": "*Corresponding authors. Work partially done at Engineering Research Center of Next-Generation Intelligent Search and Recommendation, Ministry of Education. \nWork done when Teng Shi was the intern at Kuaishou."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 50,
+ 617,
+ 295,
+ 675
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 617,
+ 295,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 617,
+ 295,
+ 675
+ ],
+ "type": "text",
+ "content": "Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 52,
+ 676,
+ 117,
+ 684
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 676,
+ 117,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 676,
+ 117,
+ 684
+ ],
+ "type": "text",
+ "content": "SIGIR '25, Padua, Italy."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 52,
+ 685,
+ 289,
+ 700
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 685,
+ 289,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 685,
+ 289,
+ 700
+ ],
+ "type": "text",
+ "content": "© 2025 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 979-8-4007-1592-1/25/07"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 52,
+ 700,
+ 177,
+ 709
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 700,
+ 177,
+ 709
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 700,
+ 177,
+ 709
+ ],
+ "type": "text",
+ "content": "https://doi.org/10.1145/XXXXXX.XXXXXXX"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 408,
+ 125,
+ 447,
+ 137
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 408,
+ 125,
+ 447,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 408,
+ 125,
+ 447,
+ 137
+ ],
+ "type": "text",
+ "content": "Jun Xu*"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 399,
+ 139,
+ 458,
+ 152
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 399,
+ 139,
+ 458,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 399,
+ 139,
+ 458,
+ 152
+ ],
+ "type": "text",
+ "content": "Xiao Zhang"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 371,
+ 152,
+ 487,
+ 163
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 371,
+ 152,
+ 487,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 371,
+ 152,
+ 487,
+ 163
+ ],
+ "type": "text",
+ "content": "Renmin University of China"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 399,
+ 164,
+ 459,
+ 175
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 399,
+ 164,
+ 459,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 399,
+ 164,
+ 459,
+ 175
+ ],
+ "type": "text",
+ "content": "Beijing, China"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 367,
+ 176,
+ 490,
+ 187
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 367,
+ 176,
+ 490,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 367,
+ 176,
+ 490,
+ 187
+ ],
+ "type": "text",
+ "content": "{junxu,zhangx89}@ruc.edu.cn"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 402,
+ 197,
+ 455,
+ 210
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 402,
+ 197,
+ 455,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 402,
+ 197,
+ 455,
+ 210
+ ],
+ "type": "text",
+ "content": "Yang Song"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 410,
+ 212,
+ 446,
+ 222
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 410,
+ 212,
+ 446,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 410,
+ 212,
+ 446,
+ 222
+ ],
+ "type": "text",
+ "content": "Han Li"
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 365,
+ 224,
+ 492,
+ 236
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 365,
+ 224,
+ 492,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 365,
+ 224,
+ 492,
+ 236
+ ],
+ "type": "text",
+ "content": "Kuaishou Technology Co., Ltd."
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 399,
+ 236,
+ 458,
+ 247
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 399,
+ 236,
+ 458,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 399,
+ 236,
+ 458,
+ 247
+ ],
+ "type": "text",
+ "content": "Beijing, China"
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 399,
+ 248,
+ 458,
+ 259
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 399,
+ 248,
+ 458,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 399,
+ 248,
+ 458,
+ 259
+ ],
+ "type": "text",
+ "content": "ys@sonyis.me"
+ }
+ ]
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 380,
+ 260,
+ 477,
+ 270
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 260,
+ 477,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 260,
+ 477,
+ 270
+ ],
+ "type": "text",
+ "content": "lihan08@kuaishou.com"
+ }
+ ]
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 314,
+ 280,
+ 559,
+ 303
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 280,
+ 559,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 280,
+ 559,
+ 303
+ ],
+ "type": "text",
+ "content": "validate the effectiveness of CFRAG. Further analysis confirms the importance of incorporating collaborative information."
+ }
+ ]
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 315,
+ 312,
+ 388,
+ 325
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 312,
+ 388,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 312,
+ 388,
+ 325
+ ],
+ "type": "text",
+ "content": "CCS Concepts"
+ }
+ ]
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 314,
+ 327,
+ 567,
+ 350
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 327,
+ 567,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 327,
+ 567,
+ 350
+ ],
+ "type": "text",
+ "content": "- Information systems "
+ },
+ {
+ "bbox": [
+ 314,
+ 327,
+ 567,
+ 350
+ ],
+ "type": "inline_equation",
+ "content": "\\rightarrow"
+ },
+ {
+ "bbox": [
+ 314,
+ 327,
+ 567,
+ 350
+ ],
+ "type": "text",
+ "content": " Personalization; - Computing methodologies "
+ },
+ {
+ "bbox": [
+ 314,
+ 327,
+ 567,
+ 350
+ ],
+ "type": "inline_equation",
+ "content": "\\rightarrow"
+ },
+ {
+ "bbox": [
+ 314,
+ 327,
+ 567,
+ 350
+ ],
+ "type": "text",
+ "content": " Natural language generation."
+ }
+ ]
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 315,
+ 360,
+ 367,
+ 372
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 360,
+ 367,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 360,
+ 367,
+ 372
+ ],
+ "type": "text",
+ "content": "Keywords"
+ }
+ ]
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 314,
+ 374,
+ 559,
+ 396
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 374,
+ 559,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 374,
+ 559,
+ 396
+ ],
+ "type": "text",
+ "content": "Large language model; Personalization; Retrieval augmented generation"
+ }
+ ]
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 315,
+ 403,
+ 405,
+ 411
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 403,
+ 405,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 403,
+ 405,
+ 411
+ ],
+ "type": "text",
+ "content": "ACM Reference Format:"
+ }
+ ]
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 314,
+ 413,
+ 559,
+ 472
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 413,
+ 559,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 413,
+ 559,
+ 472
+ ],
+ "type": "text",
+ "content": "Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Han Li. 2025. Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '25), July 13-18, 2025, Padua, Italy. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/XXXXXX.XXXXXXX"
+ }
+ ]
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 315,
+ 487,
+ 398,
+ 498
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 487,
+ 398,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 487,
+ 398,
+ 498
+ ],
+ "type": "text",
+ "content": "1 Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 313,
+ 501,
+ 559,
+ 567
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 501,
+ 559,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 501,
+ 559,
+ 567
+ ],
+ "type": "text",
+ "content": "Personalizing Large Language Models (LLMs) [55] to generate personalized outputs tailored to individual user preferences has emerged as a significant and rapidly growing field [16, 23, 29, 31, 32, 36, 37, 57]. Personalized Retrieval-Augmented Generation (RAG) [8] has become a commonly used approach for personalizing LLMs [29, 31, 32, 57]."
+ }
+ ]
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 313,
+ 567,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 567,
+ 559,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 567,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": "The process of existing personalized RAG methods typically involves retrieving similar documents from the user's historical behaviors based on the user's input query, then concatenating these documents with the query as a prompt input to the LLM for generation. Although effective, this approach is limited to retrieving only the current user's history, neglecting collaborative information. Users with similar histories tend to be more alike, and the information from these similar users can also aid in personalizing generation for the current user. As shown in the example in Figure 1, the upper part illustrates the results of the existing RAG method, which retrieves documents from the current user's history. We can only infer from these results that \"She\" in the user's input refers to \"Hillary Clinton\". In contrast, the lower part demonstrates"
+ }
+ ]
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 14,
+ 221,
+ 37,
+ 555
+ ],
+ "type": "aside_text",
+ "angle": 270,
+ "lines": [
+ {
+ "bbox": [
+ 14,
+ 221,
+ 37,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 14,
+ 221,
+ 37,
+ 555
+ ],
+ "type": "text",
+ "content": "arXiv:2504.05731v1 [cs.IR] 8 Apr 2025"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 65,
+ 82,
+ 541,
+ 276
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 65,
+ 82,
+ 541,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 65,
+ 82,
+ 541,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 65,
+ 82,
+ 541,
+ 276
+ ],
+ "type": "image",
+ "image_path": "c145d4b73d48d432adc24495509ff5b7dc143be4525fd4b84a31137fa6d78e46.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 50,
+ 278,
+ 559,
+ 334
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 278,
+ 559,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 278,
+ 559,
+ 334
+ ],
+ "type": "text",
+ "content": "Figure 1: An example from the LaMP-4 dataset [32]. The task of LaMP-4 is to generate personalized news headlines based on user input. This example illustrates the benefit of collaborative information for LLM personalization: (a) The top shows results retrieved by the existing RAG method from the current user's history, where we can only infer that \"She\" in the user's input refers to \"Hillary Clinton\". (b) The bottom shows results retrieved by our method from similar users' histories, allowing us to infer further that \"his\" in the user's input refers to \"Donald Trump\" thus enabling the generation of a more accurate result."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 50,
+ 342,
+ 295,
+ 407
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 342,
+ 295,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 342,
+ 295,
+ 407
+ ],
+ "type": "text",
+ "content": "our method, which retrieves documents from the history of similar users. In this case, we can further infer that \"his\" in the user's input refers to \"Donald Trump\", leading to a better generation result. From this example, we can see that incorporating collaborative information allows the retrieval of more diverse documents, helping the LLM generate results that better meet the user's needs."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 50,
+ 407,
+ 295,
+ 561
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 407,
+ 295,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 407,
+ 295,
+ 561
+ ],
+ "type": "text",
+ "content": "Inspired by the application of collaborative filtering in recommender systems [11, 40, 46], we propose to adapt collaborative information into RAG to personalize LLMs. However, adapting collaborative filtering to personalized RAG presents two challenges. Challenge 1: How to incorporate collaborative information. Without explicit labels indicating which users are similar, which users' information should be selected to help personalize generation for the current user? Challenge 2: How to retrieve documents that support personalized LLM generation, rather than relying on traditional semantic relevance? Pre-trained dense retrieval models [54] only retrieve based on the semantic relevance between the query and document. Directly using these models for retrieval may not necessarily result in content that allows the LLM to generate outputs that meet the user's needs [25, 35]."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 50,
+ 561,
+ 295,
+ 704
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 561,
+ 295,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 561,
+ 295,
+ 704
+ ],
+ "type": "text",
+ "content": "To address the above challenges, this paper proposes a method named CFRAG which adapts Collaborative Filtering to personalized Retrieval Augmented Generation. Firstly, to address Challenge 1, since there are no explicit user similarity labels, we use contrastive learning [15, 44] to train user embeddings for retrieving similar users to introduce collaborative information. Specifically, we apply different data augmentation methods to the user's history to obtain different views, and then treat different views of the same user's history as positive samples for each other. Then we use contrastive learning on different views to train the user embeddings. Secondly, for Challenge 2, we designed a personalized retriever and reranker to retrieve the top-"
+ },
+ {
+ "bbox": [
+ 50,
+ 561,
+ 295,
+ 704
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 50,
+ 561,
+ 295,
+ 704
+ ],
+ "type": "text",
+ "content": " documents from the histories of the retrieved users. In both retrieval and reranking, in addition to the semantic"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 313,
+ 342,
+ 559,
+ 430
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 342,
+ 559,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 342,
+ 559,
+ 430
+ ],
+ "type": "text",
+ "content": "relevance between the query and documents, we also considered the user's preferences for different documents to enable personalized retrieval. Additionally, we further fine-tune the retriever and reranker based on the feedback from the LLM to ensure that the retrieved documents better support the personalized LLM generation. Finally, the top-"
+ },
+ {
+ "bbox": [
+ 313,
+ 342,
+ 559,
+ 430
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 313,
+ 342,
+ 559,
+ 430
+ ],
+ "type": "text",
+ "content": " documents are concatenated with the user's input query to form a prompt, which is then fed into the LLM for personalized generation."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 314,
+ 430,
+ 559,
+ 485
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 430,
+ 559,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 430,
+ 559,
+ 485
+ ],
+ "type": "text",
+ "content": "The major contributions of the paper are summarized as follows: We analyzed the necessity of introducing collaborative filtering into RAG for LLM personalization and identified the challenges: how to introduce collaborative information and how to retrieve documents that support personalized LLM generation."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 314,
+ 484,
+ 559,
+ 583
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 11,
+ "blocks": [
+ {
+ "bbox": [
+ 314,
+ 484,
+ 559,
+ 539
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 484,
+ 559,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 484,
+ 559,
+ 539
+ ],
+ "type": "text",
+ "content": "- We proposed a method called CFRAG, which uses contrastive learning to train user embeddings for retrieving similar users and incorporating collaborative information. It leverages LLM feedback to train the personalized retriever and reranker, enabling them to retrieve documents that support personalized LLM generation."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 314,
+ 539,
+ 559,
+ 583
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 539,
+ 559,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 539,
+ 559,
+ 583
+ ],
+ "type": "text",
+ "content": "- Experimental results on the Language Model Personalization (LaMP) [32] benchmark validate the effectiveness of CFRAG. The experimental analysis also demonstrates the importance of leveraging collaborative information."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 604,
+ 404,
+ 615
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 604,
+ 404,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 604,
+ 404,
+ 615
+ ],
+ "type": "text",
+ "content": "2 Related Work"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 313,
+ 622,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 622,
+ 559,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 622,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": "Personalization of LLMs. Large Language Models (LLMs) [55] have demonstrated remarkable capabilities in various fields, such as text generation [22], information retrieval [56], recommender systems [5, 41], and so on. However, since LLMs are typically designed to serve all tasks with a single model and are trained on broad, domain-agnostic data, they face challenges in adapting to the personalized needs of individual users [4, 32]. Therefore, LLM personalization has attracted widespread attention [16, 31, 57]."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 171,
+ 68
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 171,
+ 68
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 171,
+ 68
+ ],
+ "type": "text",
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 69
+ ],
+ "type": "text",
+ "content": "Teng Shi et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 1
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 67,
+ 83,
+ 539,
+ 266
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 67,
+ 83,
+ 539,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 83,
+ 539,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 83,
+ 539,
+ 266
+ ],
+ "type": "image",
+ "image_path": "34f15c2d425a0ba3e0b25f67abb59aecc0f185aae42a557093eae9d488fd5f26.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "type": "text",
+ "content": "Figure 2: The architecture of CFRAG. From left to right: (a) User Retrieval retrieves similar users (Section 4.1); (b) Retriever retrieves the top- "
+ },
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "type": "text",
+ "content": " documents from each user's history (Section 4.2); (c) Reranker reranks the "
+ },
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "type": "text",
+ "content": " documents to get the final top- "
+ },
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 49,
+ 268,
+ 560,
+ 312
+ ],
+ "type": "text",
+ "content": " documents, which are then concatenated with the query and input into the LLM for personalized text generation (Section 4.3)."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 50,
+ 320,
+ 296,
+ 474
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 320,
+ 296,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 320,
+ 296,
+ 474
+ ],
+ "type": "text",
+ "content": "Existing works on LLM personalization mainly include the following types of methods: (1) Fine-tuning a personalized LLM for each user [36, 37, 42]; Tan et al. [37] fine-tuned the LLM using LoRA [12] to get personalized LoRA parameters for each user. (2) Aligning LLMs with user-specific preferences through Reinforcement Learning from Human Feedback (RLHF) [16, 23, 43]; Jang et al. [16] first trained different parameters for various objectives using RLHF, then merged these parameters based on users' personalized needs. (3) Incorporating user-specific context into the prompt [21, 27, 29, 31, 32, 57]. Richardson et al. [29] used instruction-tuned LLMs to summarize user history and then incorporated it into prompts for generation. Salemi et al. [31, 32] used RAG to retrieve relevant documents from user history based on the input query and incorporated them into the prompt."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 50,
+ 474,
+ 296,
+ 605
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 474,
+ 296,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 474,
+ 296,
+ 605
+ ],
+ "type": "text",
+ "content": "This paper further introduces collaborative filtering for personalization based on the RAG framework. Collaborative filtering has already been applied in fields such as recommender systems [33, 34, 38, 48-52] and has been proven effective. It assumes that users who have interacted with similar items share similar preferences, and recommending items from similar users to the current user can meet their needs. Some works [11, 46] learn the collaborative information between users and items through matrix factorization [19], while others [10, 40] further explore higher-order collaborative information between users and items using graph neural networks. The application of collaborative filtering in LLM personalization remains under-explored."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 50,
+ 608,
+ 296,
+ 708
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 608,
+ 296,
+ 708
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 608,
+ 296,
+ 708
+ ],
+ "type": "text",
+ "content": "Retrieval Augmented Generation. Retrieval Augmented Generation [7, 8] introduces external knowledge through document retrieval, alleviating issues such as LLM hallucinations [53], and enhancing LLMs' capabilities in knowledge-intensive tasks [17] such as open-domain question answering [14, 20]. Some works [3, 13] encode retrieved documents using separate encoders, and then fuse the results with the language model using cross-attention. A more common approach is to directly include the retrieved documents in the prompt of the LLM [2, 9, 20, 25, 35]. In recent years, this"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 314,
+ 320,
+ 561,
+ 365
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 320,
+ 561,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 320,
+ 561,
+ 365
+ ],
+ "type": "text",
+ "content": "in-context RAG framework has also been applied to LLM personalization, which is personalized by retrieving documents from the user's history [31, 32, 57]. This paper introduces collaborative filtering by retrieving similar users' histories for better personalization."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 315,
+ 373,
+ 442,
+ 385
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 373,
+ 442,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 373,
+ 442,
+ 385
+ ],
+ "type": "text",
+ "content": "3 Problem Formulation"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": "Let "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{U} = \\{u_1, u_2, \\ldots, u_M\\}"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " denotes the set of all users, where "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "M"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " is the number of users. Each user "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "u \\in \\mathcal{U}"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " has a chronologically ordered history "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u = [d_1, d_2, \\ldots, d_N]"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " which includes all her historical documents, where "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "N"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " is the number of documents in the history. The personalized text generation dataset is "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{D} = \\{(u, q, y)_i\\}_{i=1}^{|D|}"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": ". For each instance, "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "q"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " is the query input by the user "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "u"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " to the LLM, and "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "y"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " is the target output. Our goal is first to introduce collaborative information by retrieving the top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": " most similar users for user "
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "u"
+ },
+ {
+ "bbox": [
+ 314,
+ 388,
+ 560,
+ 479
+ ],
+ "type": "text",
+ "content": ":"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 382,
+ 483,
+ 492,
+ 495
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 382,
+ 483,
+ 492,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 483,
+ 492,
+ 495
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {U} _ {\\text {r e t r i e v e d}} = \\left\\{u _ {1}, u _ {2}, \\dots , u _ {m} \\right\\}.",
+ "image_path": "9ff4e06204e6c48c44c4467eca5a18d03e07c6da14c61020e7c1603ef065ab96.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "type": "text",
+ "content": "Then, we use a retriever to retrieve the top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "type": "text",
+ "content": " documents from each of the "
+ },
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "type": "text",
+ "content": " users' histories, resulting in a total of "
+ },
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 314,
+ 500,
+ 559,
+ 521
+ ],
+ "type": "text",
+ "content": " documents."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 348,
+ 527,
+ 525,
+ 539
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 348,
+ 527,
+ 525,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 348,
+ 527,
+ 525,
+ 539
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {D} _ {\\text {r e t r i e v e d}} = \\{d _ {i, j} | i \\in \\{1, \\dots , m \\}, j \\in \\{1, \\dots , k \\} \\}.",
+ "image_path": "286390de4f0085e9bc40dde8cd842fc967aed23ac33c9c8b2fef3cd5ce5de750.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 314,
+ 543,
+ 559,
+ 565
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 543,
+ 559,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 543,
+ 559,
+ 565
+ ],
+ "type": "text",
+ "content": "Finally, we use a reranker to rerank these "
+ },
+ {
+ "bbox": [
+ 314,
+ 543,
+ 559,
+ 565
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 314,
+ 543,
+ 559,
+ 565
+ ],
+ "type": "text",
+ "content": " documents and obtain the final top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 543,
+ 559,
+ 565
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 314,
+ 543,
+ 559,
+ 565
+ ],
+ "type": "text",
+ "content": " documents:"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 378,
+ 570,
+ 495,
+ 582
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 378,
+ 570,
+ 495,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 378,
+ 570,
+ 495,
+ 582
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {D} _ {\\text {r e r a n k e d}} = \\left\\{d _ {i} | i \\in \\{1, \\dots , k \\} \\right\\}.",
+ "image_path": "a6b102852714ef7badacf8034fddfa94795403b0b6339e25278197262aa0cdab.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "type": "text",
+ "content": "These top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "type": "text",
+ "content": " documents will be concatenated with the user's query "
+ },
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "type": "inline_equation",
+ "content": "q"
+ },
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "type": "text",
+ "content": " as a prompt and input into the LLM, enabling it to generate a response that aligns with the target output "
+ },
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "type": "inline_equation",
+ "content": "y"
+ },
+ {
+ "bbox": [
+ 314,
+ 586,
+ 558,
+ 619
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 314,
+ 619,
+ 559,
+ 664
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 619,
+ 559,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 619,
+ 559,
+ 664
+ ],
+ "type": "text",
+ "content": "This paper primarily focuses on how to retrieve "
+ },
+ {
+ "bbox": [
+ 314,
+ 619,
+ 559,
+ 664
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{U}_{\\mathrm{retrieved}}"
+ },
+ {
+ "bbox": [
+ 314,
+ 619,
+ 559,
+ 664
+ ],
+ "type": "text",
+ "content": " to introduce collaborative information, and how to train the retriever and reranker so that they can effectively retrieve documents that support the personalized LLM generation."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 315,
+ 673,
+ 406,
+ 685
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 673,
+ 406,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 673,
+ 406,
+ 685
+ ],
+ "type": "text",
+ "content": "4 Our Approach"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 314,
+ 687,
+ 560,
+ 710
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 687,
+ 560,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 687,
+ 560,
+ 710
+ ],
+ "type": "text",
+ "content": "This section introduces our method CFRAG. CFRAG's overall architecture is shown in Figure 2. As mentioned in Section 1, to address"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "type": "text",
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 440,
+ 60,
+ 559,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 440,
+ 60,
+ 559,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 440,
+ 60,
+ 559,
+ 69
+ ],
+ "type": "text",
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 2
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "text",
+ "content": "Challenge 1, i.e., how to introduce collaborative information, we first train user embeddings using contrastive learning to retrieve the top-"
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "text",
+ "content": " most similar users (see Section 4.1). For Challenge 2, which involves retrieving documents that support personalized LLM generation, we fine-tune the personalized retriever and reranker using LLM feedback. The retriever first retrieves the top-"
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "text",
+ "content": " documents from the history of each of the "
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "text",
+ "content": " users, resulting in "
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "text",
+ "content": " documents (see Section 4.2). The reranker then reranks these documents to obtain the final top-"
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 50,
+ 84,
+ 296,
+ 194
+ ],
+ "type": "text",
+ "content": " documents as input for the LLM (see Section 4.3)."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 51,
+ 205,
+ 149,
+ 216
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 205,
+ 149,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 205,
+ 149,
+ 216
+ ],
+ "type": "text",
+ "content": "4.1 User Retrieval"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "type": "text",
+ "content": "First, we perform user retrieval to get the top-"
+ },
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "type": "text",
+ "content": " most similar users for user "
+ },
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "type": "inline_equation",
+ "content": "u"
+ },
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "type": "text",
+ "content": " to introduce collaborative information. However, we do not have labels indicating which users are similar to each other. To address this, we employ a contrastive learning [15, 44] approach. We apply different data augmentation methods to the user history "
+ },
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 50,
+ 220,
+ 296,
+ 331
+ ],
+ "type": "text",
+ "content": " to obtain different views of the user's history. We treat different views of the same user as positive samples and the histories of other users as negative samples, and then we use the InfoNCE [28] loss to train user embeddings for retrieval. Figure 3 illustrates the process of training user embeddings using contrastive learning."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": "4.1.1 User Encoder. Specifically, we first use an embedding model (such as BERT [6], RoBERTa [26], BGE [45] etc.) "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{Emb}(\\cdot)"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": " to encode each document in the user's history "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": " to obtain "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{E}_u = [\\mathbf{e}_1,\\mathbf{e}_2,\\dots ,\\mathbf{e}_N]^{\\intercal}\\in \\mathbb{R}^{N\\times d}"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{e}_i = \\mathbf{Emb}(d_i)"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "d"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": " is the embedding dimension. To model the sequential relationships between different documents in the user's history, we introduce positional embedding "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{P}\\in \\mathbb{R}^{N\\times d}"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": ". Afterward, the history "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": " 's embedding becomes "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "\\widehat{\\mathbf{E}}_u = \\mathbf{E}_u + \\mathbf{P}"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": ". Then, we apply a transformer [39] as the user encoder to encode the user's history "
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "inline_equation",
+ "content": "\\widehat{\\mathbf{E}}_u"
+ },
+ {
+ "bbox": [
+ 50,
+ 336,
+ 296,
+ 449
+ ],
+ "type": "text",
+ "content": " and average the transformer's output to obtain the user's embedding:"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 89,
+ 455,
+ 294,
+ 469
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 89,
+ 455,
+ 294,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 89,
+ 455,
+ 294,
+ 469
+ ],
+ "type": "interline_equation",
+ "content": "\\mathbf {e} _ {u} = \\operatorname {E n c o d e r} _ {u} (u) = \\operatorname {M E A N} (\\operatorname {T r m} (\\widehat {\\mathbf {F}} _ {u})) \\in \\mathbb {R} ^ {d}, \\tag {1}",
+ "image_path": "1770e817342ffc39a1149bbb6148e83683024c61a30a0f540c0cbacac3546bc7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 50,
+ 474,
+ 294,
+ 508
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 474,
+ 294,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 474,
+ 294,
+ 508
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 50,
+ 474,
+ 294,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{Encoder}_u(\\cdot)\\to \\mathbb{R}^d"
+ },
+ {
+ "bbox": [
+ 50,
+ 474,
+ 294,
+ 508
+ ],
+ "type": "text",
+ "content": " denotes the user encoder, "
+ },
+ {
+ "bbox": [
+ 50,
+ 474,
+ 294,
+ 508
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{Trm}(\\cdot)"
+ },
+ {
+ "bbox": [
+ 50,
+ 474,
+ 294,
+ 508
+ ],
+ "type": "text",
+ "content": " denotes a transformer encoder. Next, we train the transformer encoder using contrastive learning."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 51,
+ 514,
+ 294,
+ 536
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 514,
+ 294,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 514,
+ 294,
+ 536
+ ],
+ "type": "text",
+ "content": "4.1.2 Data Augmentation. We generate different views of "
+ },
+ {
+ "bbox": [
+ 51,
+ 514,
+ 294,
+ 536
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 51,
+ 514,
+ 294,
+ 536
+ ],
+ "type": "text",
+ "content": " using the following three data augmentation methods:"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "type": "text",
+ "content": "Document Crop. We randomly select a continuous sub-sequence of length "
+ },
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "type": "inline_equation",
+ "content": "L_{c} = \\lfloor \\eta_{c}N\\rfloor"
+ },
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "type": "text",
+ "content": " from "
+ },
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "type": "inline_equation",
+ "content": "\\eta_c"
+ },
+ {
+ "bbox": [
+ 51,
+ 536,
+ 296,
+ 570
+ ],
+ "type": "text",
+ "content": " is a hyper-parameter controlling the crop ratio. The history after cropping is as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 113,
+ 574,
+ 232,
+ 590
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 574,
+ 232,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 574,
+ 232,
+ 590
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {H} _ {u} ^ {\\mathrm {c r o p}} = [ d _ {c}, d _ {c + 1}, \\dots , d _ {c + L _ {c} - 1} ].",
+ "image_path": "321ed29b3fe1bd2bf8b65455f1b9c4e37c82123640abf0223a5bca357a1a3ec4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "text",
+ "content": "Document Mask. For the history "
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "text",
+ "content": ", we randomly mask out "
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "inline_equation",
+ "content": "L_{m} = \\lfloor \\eta_{m}N\\rfloor"
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "text",
+ "content": " documents "
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{I}_{\\mathrm{mask}} = \\{i_1,i_2,\\dots ,i_{L_m}\\}"
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{I}_{\\mathrm{mask}}"
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "text",
+ "content": " is the set of indices corresponding to the masked documents and "
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "inline_equation",
+ "content": "\\eta_{m}"
+ },
+ {
+ "bbox": [
+ 50,
+ 594,
+ 296,
+ 661
+ ],
+ "type": "text",
+ "content": " is a hyper-parameter that controls the mask ratio. The masked documents are replaced with a special token [mask]. The history after masking is as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 112,
+ 666,
+ 230,
+ 712
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 666,
+ 230,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 666,
+ 230,
+ 712
+ ],
+ "type": "interline_equation",
+ "content": "\\begin{array}{l} \\mathcal {H} _ {u} ^ {\\text {m a s k}} = \\left[ \\hat {d} _ {1}, \\hat {d} _ {2}, \\dots , \\hat {d} _ {N} \\right], \\\\ \\hat {d} _ {i} = \\left\\{ \\begin{array}{l l} d _ {i}, & i \\notin \\mathcal {I} _ {\\text {m a s k}}, \\\\ [ \\text {m a s k} ], & i \\in \\mathcal {I} _ {\\text {m a s k}}. \\end{array} \\right. \\\\ \\end{array}",
+ "image_path": "fa1ad7982cf1c2b2fc0d929344eb06b13bcbc11bbbb85e0c5d98093ed5b060b1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 320,
+ 83,
+ 553,
+ 194
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 320,
+ 83,
+ 553,
+ 194
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 83,
+ 553,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 83,
+ 553,
+ 194
+ ],
+ "type": "image",
+ "image_path": "625fdaef24bf247580e641a34943670ea15cb8f8da28784bc68a9e6a62dbe508.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 315,
+ 197,
+ 558,
+ 209
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 197,
+ 558,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 197,
+ 558,
+ 209
+ ],
+ "type": "text",
+ "content": "Figure 3: Contrastive learning for user embedding training."
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "text",
+ "content": "Document Reorder. We randomly select a sub-sequence "
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "inline_equation",
+ "content": "[d_r, d_{r+1}, \\ldots, d_{r+L_r-1}]"
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "text",
+ "content": " of length "
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "inline_equation",
+ "content": "L_r = \\lfloor \\eta_r N \\rfloor"
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "text",
+ "content": " from "
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "inline_equation",
+ "content": "\\eta_r"
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "text",
+ "content": " is a hyper-parameter controlling the reorder ratio, and then randomly shuffle the order of the documents within the sub-sequence to obtain "
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "inline_equation",
+ "content": "[\\hat{d}_r, \\hat{d}_{r+1}, \\ldots, \\hat{d}_{r+L_r-1}]"
+ },
+ {
+ "bbox": [
+ 314,
+ 220,
+ 559,
+ 285
+ ],
+ "type": "text",
+ "content": ". The history after reordering is as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 349,
+ 289,
+ 523,
+ 304
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 349,
+ 289,
+ 523,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 349,
+ 289,
+ 523,
+ 304
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {H} _ {u} ^ {\\text {r e o r d e r}} = \\left[ d _ {1}, d _ {2}, \\dots , \\hat {d} _ {r}, \\dots , \\hat {d} _ {r + L _ {r} - 1}, \\dots , d _ {N} \\right].",
+ "image_path": "6edc98bed93f6b43480ec1d654b337ca409da377fc6ab6198e2c9c1f23130ce8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": "4.1.3 Contrastive Loss. Each time, we randomly select two data augmentation methods "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{A}'"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{A}''"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": " to generate two different views of "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": ", denoted as "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u'"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u''"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": ". Then, using the encoder described in Section 4.1.1, we obtain the user embeddings "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{e}_u'"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{e}_u''"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": " corresponding to the different views. Since "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{e}_u'"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{e}_u''"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": " are obtained through data augmentation of "
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_u"
+ },
+ {
+ "bbox": [
+ 314,
+ 308,
+ 559,
+ 418
+ ],
+ "type": "text",
+ "content": ", they are more similar to each other. Therefore, we treat them as positive samples for each other and use the views generated from the augmented histories of other users in the same batch as negative samples. We then perform contrastive learning using the InfoNCE [28] loss as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 345,
+ 422,
+ 558,
+ 482
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 345,
+ 422,
+ 558,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 422,
+ 558,
+ 482
+ ],
+ "type": "interline_equation",
+ "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C L}} = - \\left[ \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {e} _ {u} ^ {\\prime} , \\mathbf {e} _ {u} ^ {\\prime \\prime}\\right) / \\tau_ {1}\\right)}{\\sum_ {u ^ {-} \\in \\mathcal {U} _ {\\mathrm {n e g}}} \\exp \\left(\\cos \\left(\\mathbf {e} _ {u} ^ {\\prime} , \\mathbf {e} _ {u ^ {-}} ^ {\\prime \\prime}\\right) / \\tau_ {1}\\right)} \\right. \\tag {2} \\\\ \\left. + \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {e} _ {u} ^ {\\prime} , \\mathbf {e} _ {u} ^ {\\prime \\prime}\\right) / \\tau_ {1}\\right)}{\\sum_ {u ^ {-} \\in \\mathcal {U} _ {\\text {n e g}}} \\exp \\left(\\cos \\left(\\mathbf {e} _ {u ^ {-}} ^ {\\prime}, \\mathbf {e} _ {u} ^ {\\prime \\prime}\\right) / \\tau_ {1}\\right)} \\right], \\\\ \\end{array}",
+ "image_path": "9976451dd3eec8608f9caf953988cee5f27caa57ccb7db4f2daace3708385103.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{1}"
+ },
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "type": "text",
+ "content": " is the temperature coefficient, "
+ },
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{U}_{\\mathrm{neg}}"
+ },
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "type": "text",
+ "content": " are the set of randomly sampled in-batch negative samples, and "
+ },
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "type": "inline_equation",
+ "content": "\\cos (\\cdot)"
+ },
+ {
+ "bbox": [
+ 314,
+ 486,
+ 560,
+ 519
+ ],
+ "type": "text",
+ "content": " denotes the cosine similarity."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "text",
+ "content": "4.1.4 Top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "text",
+ "content": " User Retrieval. After training with contrastive learning, we can use the encoder from Section 4.1.1 to obtain the user embedding "
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{e}_u"
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "text",
+ "content": ". We then calculate the cosine similarity between each pair of user embeddings and retrieve the top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "text",
+ "content": " most similar users "
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{U}_{\\mathrm{retrieved}} = \\{u_1, u_2, \\dots, u_m\\}"
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "text",
+ "content": " for user "
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "inline_equation",
+ "content": "u"
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "text",
+ "content": ". Subsequently, the histories of these "
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 314,
+ 525,
+ 560,
+ 590
+ ],
+ "type": "text",
+ "content": " users will be used for further document retrieval."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 315,
+ 601,
+ 443,
+ 612
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 601,
+ 443,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 601,
+ 443,
+ 612
+ ],
+ "type": "text",
+ "content": "4.2 Document Retrieval"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "text",
+ "content": "After retrieving the top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "text",
+ "content": " users, we design a personalized retriever to retrieve the top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "text",
+ "content": " documents from each user's history, resulting in a total of "
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "text",
+ "content": " candidate documents "
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{D}_{\\text{retrieved}} = \\{d_{i,j} | i \\in \\{1, \\ldots, m\\}, j \\in \\{1, \\ldots, k\\}\\}"
+ },
+ {
+ "bbox": [
+ 314,
+ 615,
+ 560,
+ 682
+ ],
+ "type": "text",
+ "content": ". This section introduces how the retriever is designed and how it's trained to retrieve documents that better align with the requirements of personalized LLM generation."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 314,
+ 687,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 687,
+ 559,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 687,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": "4.2.1 Retriever. First, we use a pre-trained dense retrieval model (such as BGE retriever [45]) to compute the semantic relevance"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 171,
+ 68
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 171,
+ 68
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 171,
+ 68
+ ],
+ "type": "text",
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 68
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 68
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 68
+ ],
+ "type": "text",
+ "content": "Teng Shi et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 55,
+ 83,
+ 290,
+ 183
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 55,
+ 83,
+ 290,
+ 183
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 83,
+ 290,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 83,
+ 290,
+ 183
+ ],
+ "type": "image",
+ "image_path": "92815eb6c2f51b4c9ef6ca63b04dd6a8f815dc0d32224bff35a685479dc7ee07.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 50,
+ 185,
+ 295,
+ 208
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 185,
+ 295,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 185,
+ 295,
+ 208
+ ],
+ "type": "text",
+ "content": "Figure 4: The method of training the retriever and reranker using LLM feedback."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 50,
+ 222,
+ 235,
+ 233
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 222,
+ 235,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 222,
+ 235,
+ 233
+ ],
+ "type": "text",
+ "content": "between the query and the candidate documents:"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 91,
+ 239,
+ 295,
+ 255
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 91,
+ 239,
+ 295,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 91,
+ 239,
+ 295,
+ 255
+ ],
+ "type": "interline_equation",
+ "content": "S _ {q, d} ^ {\\text {r e t r i e v e r}} = \\cos \\left(\\operatorname {E n c o d e r} _ {q} (q), \\operatorname {E n c o d e r} _ {d} (d)\\right), \\tag {3}",
+ "image_path": "45032d0b3f912fe033f347f966ca5cdbdf24f9320b31a6df192191757e19f649.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{Encoder}_q(\\cdot)\\to \\mathbb{R}^d"
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{Encoder}_d(\\cdot)\\rightarrow \\mathbb{R}^d"
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "text",
+ "content": " are the encoders for the query and the document in the retrieval model, respectively. Pre-trained retrieval models typically use "
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "inline_equation",
+ "content": "S_{q,d}^{\\mathrm{retriever}}"
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "text",
+ "content": " directly for retrieval. However, "
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "inline_equation",
+ "content": "S_{q,d}^{\\mathrm{retriever}}"
+ },
+ {
+ "bbox": [
+ 50,
+ 262,
+ 296,
+ 366
+ ],
+ "type": "text",
+ "content": " only considers the semantic relevance between the query and the document. Since different users might input the same query but expect different outputs due to their varying preferences, we further account for user personalization by calculating the preference score of the user for the document as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 96,
+ 373,
+ 295,
+ 388
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 96,
+ 373,
+ 295,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 96,
+ 373,
+ 295,
+ 388
+ ],
+ "type": "interline_equation",
+ "content": "S _ {u, d} ^ {\\text {r e t r i e v e r}} = \\cos \\left(\\mathrm {M L P} _ {1} \\left(\\mathbf {e} _ {u}\\right), \\operatorname {E n c o d e r} _ {d} (d)\\right), \\tag {4}",
+ "image_path": "f519fd19f5d090c02b1805b4837891bd2fdbf2eba9b80e44d9e885596ab1fdf6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 50,
+ 394,
+ 295,
+ 440
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 394,
+ 295,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 394,
+ 295,
+ 440
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 50,
+ 394,
+ 295,
+ 440
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{MLP}_1: \\mathbb{R}^d \\to \\mathbb{R}^d"
+ },
+ {
+ "bbox": [
+ 50,
+ 394,
+ 295,
+ 440
+ ],
+ "type": "text",
+ "content": " is a multi-layer perceptron that maps the user embedding to the space where the cosine similarity is computed. "
+ },
+ {
+ "bbox": [
+ 50,
+ 394,
+ 295,
+ 440
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{e}_u"
+ },
+ {
+ "bbox": [
+ 50,
+ 394,
+ 295,
+ 440
+ ],
+ "type": "text",
+ "content": " is the embedding obtained in Section 4.1.1. The total score for retrieval is computed as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 98,
+ 445,
+ 295,
+ 461
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 98,
+ 445,
+ 295,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 98,
+ 445,
+ 295,
+ 461
+ ],
+ "type": "interline_equation",
+ "content": "S _ {u, q, d} ^ {\\text {r e t r i e v e r}} = (1 - \\alpha) S _ {q, d} ^ {\\text {r e t r i e v e r}} + \\alpha S _ {u, d} ^ {\\text {r e t r i e v e r}}, \\tag {5}",
+ "image_path": "553b67f5416e9bf7a6397eb8a5e9f5ff9e90f77643a2e2ab2977210184904014.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 50,
+ 466,
+ 296,
+ 489
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 466,
+ 296,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 466,
+ 296,
+ 489
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 50,
+ 466,
+ 296,
+ 489
+ ],
+ "type": "inline_equation",
+ "content": "\\alpha"
+ },
+ {
+ "bbox": [
+ 50,
+ 466,
+ 296,
+ 489
+ ],
+ "type": "text",
+ "content": " is a hyper-parameter that controls the weight of personalization."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 50,
+ 495,
+ 295,
+ 582
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 495,
+ 295,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 495,
+ 295,
+ 582
+ ],
+ "type": "text",
+ "content": "4.2.2 Training. Since the pre-trained dense retrieval model is not fine-tuned for our specific task, the retrieved results may not necessarily lead to LLM responses that better match the target output "
+ },
+ {
+ "bbox": [
+ 50,
+ 495,
+ 295,
+ 582
+ ],
+ "type": "inline_equation",
+ "content": "y"
+ },
+ {
+ "bbox": [
+ 50,
+ 495,
+ 295,
+ 582
+ ],
+ "type": "text",
+ "content": " [25, 35]. However, there is no ground truth indicating which documents are better. Therefore, we evaluate the difference between the LLM's output and the target output "
+ },
+ {
+ "bbox": [
+ 50,
+ 495,
+ 295,
+ 582
+ ],
+ "type": "inline_equation",
+ "content": "y"
+ },
+ {
+ "bbox": [
+ 50,
+ 495,
+ 295,
+ 582
+ ],
+ "type": "text",
+ "content": ", using this as a label to train the retrieval model. Figure 4 shows the process of training the retriever using LLM feedback."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "text",
+ "content": "Specifically, we first use the pre-trained retrieval model to retrieve the top-"
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "text",
+ "content": " documents from each of the "
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "text",
+ "content": " users' histories based on "
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "inline_equation",
+ "content": "S_{q,d}^{\\mathrm{retriever}}"
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "text",
+ "content": " in Eq. (3), resulting in a total of "
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "text",
+ "content": " candidate documents. These documents are then concatenated with the query one by one and used as prompts for the LLM, producing "
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 50,
+ 583,
+ 295,
+ 651
+ ],
+ "type": "text",
+ "content": " outputs:"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 75,
+ 658,
+ 271,
+ 672
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 658,
+ 271,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 658,
+ 271,
+ 672
+ ],
+ "type": "interline_equation",
+ "content": "\\{O _ {q, d _ {i, j}} = \\mathrm {L L M} (q, d _ {i, j}) | i \\in \\{1, \\dots , m \\}, j \\in \\{1, \\dots , k \\} \\},",
+ "image_path": "370a5924ecdacae48795859736686a99e261ba862019de8a358e02e9a7b90220.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{LLM}(q, d_{i,j})"
+ },
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "type": "text",
+ "content": " represents the output generated by inputting the concatenated query "
+ },
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "type": "inline_equation",
+ "content": "q"
+ },
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "type": "text",
+ "content": " and document "
+ },
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "type": "inline_equation",
+ "content": "d_{i,j}"
+ },
+ {
+ "bbox": [
+ 50,
+ 677,
+ 296,
+ 712
+ ],
+ "type": "text",
+ "content": " into the LLM. Then, based on the quality of these outputs, we can calculate the distribution of"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 315,
+ 84,
+ 459,
+ 95
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 84,
+ 459,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 84,
+ 459,
+ 95
+ ],
+ "type": "text",
+ "content": "these candidate documents as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 345,
+ 99,
+ 558,
+ 129
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 345,
+ 99,
+ 558,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 99,
+ 558,
+ 129
+ ],
+ "type": "interline_equation",
+ "content": "p _ {\\text {L L M}} \\left(d _ {i, j} \\mid q, y\\right) = \\frac {\\exp (\\operatorname {e v a l} \\left(y , O _ {q , d _ {i , j}}\\right))}{\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {k} \\exp (\\operatorname {e v a l} \\left(y , O _ {q , d _ {i , j}}\\right))}, \\tag {6}",
+ "image_path": "81368817db3e34bbb1db188fa57860badccf1c91f22e28f4651cf117e8756cd9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "inline_equation",
+ "content": "\\text{eval}(\\cdot)"
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "text",
+ "content": " measures the difference between the target output "
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "inline_equation",
+ "content": "y"
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "text",
+ "content": " and the LLM's output, using metrics such as ROUGE [24] score. A larger value returned by "
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "inline_equation",
+ "content": "\\text{eval}(\\cdot)"
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "text",
+ "content": " indicates a better-generated result. Similarly, we can also calculate the score distribution of the candidate documents by the retrieval model based on "
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "inline_equation",
+ "content": "S_{u,q,d}^{\\text{retriever}}"
+ },
+ {
+ "bbox": [
+ 314,
+ 133,
+ 559,
+ 201
+ ],
+ "type": "text",
+ "content": " in Eq. (5):"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 349,
+ 205,
+ 558,
+ 241
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 349,
+ 205,
+ 558,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 349,
+ 205,
+ 558,
+ 241
+ ],
+ "type": "interline_equation",
+ "content": "p _ {\\text {r e t r i e v e r}} \\left(d _ {i, j} \\mid q, u\\right) = \\frac {\\exp \\left(S _ {u , q , d _ {i , j}} ^ {\\text {r e t r i e v e r}}\\right)}{\\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {k} \\exp \\left(S _ {u , q , d _ {i , j}} ^ {\\text {r e t r i e v e r}}\\right)}. \\tag {7}",
+ "image_path": "5bb368b797bf6324680c0dab35687b3a755828cbb10a15bd627324662e645b92.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 314,
+ 244,
+ 559,
+ 300
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 244,
+ 559,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 244,
+ 559,
+ 300
+ ],
+ "type": "text",
+ "content": "We aim for the retrieval model to retrieve documents that lead to better LLM-generated results, which means making the distribution "
+ },
+ {
+ "bbox": [
+ 314,
+ 244,
+ 559,
+ 300
+ ],
+ "type": "inline_equation",
+ "content": "p_{\\mathrm{retriever}}(d|q,u)"
+ },
+ {
+ "bbox": [
+ 314,
+ 244,
+ 559,
+ 300
+ ],
+ "type": "text",
+ "content": " in Eq. (7) closer to the distribution "
+ },
+ {
+ "bbox": [
+ 314,
+ 244,
+ 559,
+ 300
+ ],
+ "type": "inline_equation",
+ "content": "p_{\\mathrm{LLM}}(d|q,y)"
+ },
+ {
+ "bbox": [
+ 314,
+ 244,
+ 559,
+ 300
+ ],
+ "type": "text",
+ "content": " in Eq (6). Therefore, we compute the KL divergence between the two distributions as the loss to optimize the retriever:"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 347,
+ 305,
+ 558,
+ 318
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 347,
+ 305,
+ 558,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 347,
+ 305,
+ 558,
+ 318
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} _ {\\text {r e t r i e v e r}} = \\mathrm {K L} \\left(p _ {\\text {r e t r i e v e r}} (d | q, u) \\mid p _ {\\text {L L M}} (d | q, y)\\right). \\tag {8}",
+ "image_path": "4de643f3cc500517a2b74f1c2f8233ac753bd2f716de8d3e58deb45f124dd930.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 315,
+ 327,
+ 434,
+ 338
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 327,
+ 434,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 327,
+ 434,
+ 338
+ ],
+ "type": "text",
+ "content": "4.3 Document Rerank"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "text",
+ "content": "After retrieving "
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{D}_{\\mathrm{retrieved}}"
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "text",
+ "content": " through the retriever, in this section, we further refine the results by reranking "
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{D}_{\\mathrm{retrieved}}"
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "text",
+ "content": " to obtain the final top-"
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "text",
+ "content": " ranked results "
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{D}_{\\mathrm{reranked}} = \\{d_i | i \\in \\{1, \\dots, k\\}\\}"
+ },
+ {
+ "bbox": [
+ 314,
+ 342,
+ 559,
+ 376
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 314,
+ 381,
+ 558,
+ 424
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 381,
+ 558,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 381,
+ 558,
+ 424
+ ],
+ "type": "text",
+ "content": "4.3.1 Reranker. We use a pre-trained cross-encoder (such as the BGE reranker [45]) to encode the query and document, obtaining the hidden state corresponding to the [CLS] token from the last layer:"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 386,
+ 427,
+ 558,
+ 441
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 386,
+ 427,
+ 558,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 427,
+ 558,
+ 441
+ ],
+ "type": "interline_equation",
+ "content": "\\mathbf {h} _ {q, d} = \\operatorname {C r o s s E n c o d e r} (q, d), \\tag {9}",
+ "image_path": "ec0d097f9eaf257d62428318090b463fd41bd34756cf25a311d0f4e5debaae76.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 313,
+ 444,
+ 559,
+ 544
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 444,
+ 559,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 444,
+ 559,
+ 544
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 313,
+ 444,
+ 559,
+ 544
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{h}_{q,d} \\in \\mathbb{R}^d"
+ },
+ {
+ "bbox": [
+ 313,
+ 444,
+ 559,
+ 544
+ ],
+ "type": "text",
+ "content": ". Similarly, when reranking, in addition to considering the semantic relevance between query and document, we also take into account the user's personalized preferences. However, since the cross-encoder does not encode documents separately, it cannot compute the cosine similarity between users and documents as shown in Eq. (4) to express the user preference score. Therefore, we directly concatenate the user embeddings to the output of the cross-encoder to account for the influence of user preferences. The overall score used for reranking is calculated as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 350,
+ 548,
+ 558,
+ 564
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 350,
+ 548,
+ 558,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 350,
+ 548,
+ 558,
+ 564
+ ],
+ "type": "interline_equation",
+ "content": "S _ {u, q, d} ^ {\\text {r e r a n k e r}} = \\mathrm {M L P} _ {3} \\left(\\operatorname {C O N C A T} \\left(\\mathbf {h} _ {q, d}, \\operatorname {M L P} _ {2} (\\mathbf {e} _ {u})\\right)\\right), \\tag {10}",
+ "image_path": "2b38bacf7f2f45a0608c219d86ab323366512e80026a8f5314c0af3b68f98d9f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{MLP}_2: \\mathbb{R}^d \\to \\mathbb{R}^d"
+ },
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{MLP}_3: \\mathbb{R}^{2d} \\to \\mathbb{R}"
+ },
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "type": "text",
+ "content": " are two multi-layer perceptions. "
+ },
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{CONCAT}(\\cdot)"
+ },
+ {
+ "bbox": [
+ 313,
+ 570,
+ 558,
+ 594
+ ],
+ "type": "text",
+ "content": " denotes the concatenation operation."
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 314,
+ 600,
+ 558,
+ 643
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 600,
+ 558,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 600,
+ 558,
+ 643
+ ],
+ "type": "text",
+ "content": "4.3.2 Training. Similar to the retriever's training in Section 4.2.2, we also want the reranker to assign higher scores to the documents that lead to better LLM-generated results. Therefore, we train the reranker using a similar approach."
+ }
+ ]
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": "We use the trained retrieval model from Section 4.2.2 to retrieve top-"
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": " documents from the history of each of the "
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "inline_equation",
+ "content": "m"
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": " users, resulting in a total of "
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": " candidate documents. These documents are concatenated with the query "
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "inline_equation",
+ "content": "q"
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": " and used as prompts for the LLM, producing "
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "inline_equation",
+ "content": "m \\times k"
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": " outputs. Similar to Eq.(6), we can obtain the distribution "
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "inline_equation",
+ "content": "p_{\\mathrm{LLM}}(d|q,y)"
+ },
+ {
+ "bbox": [
+ 313,
+ 643,
+ 559,
+ 710
+ ],
+ "type": "text",
+ "content": " of these candidate documents. Based on"
+ }
+ ]
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "type": "text",
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 440,
+ 60,
+ 558,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 440,
+ 60,
+ 558,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 440,
+ 60,
+ 558,
+ 69
+ ],
+ "type": "text",
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 55,
+ 99,
+ 292,
+ 158
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 67,
+ 83,
+ 277,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 83,
+ 277,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 83,
+ 277,
+ 95
+ ],
+ "type": "text",
+ "content": "Table 1: Statistics of the datasets used in this paper."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 55,
+ 99,
+ 292,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 99,
+ 292,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 99,
+ 292,
+ 158
+ ],
+ "type": "table",
+ "html": "| Task | Input | Output | User History |
| LaMP-1 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please choose one of the following two references that is more relevant to the user's input title: [1] {reference1}; [2] {reference2}. Please just answer with “[1” or “[2” without explanation. “title”: {title}. | [1] | “title”: {title}\n“abstract”: {abstract} |
| LaMP-2 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please select the tag from [sci-fi, based on a book, comedy ... ] that is most relevant to the user's input description. Please just answer with the tag name without explanation. “description”: {description}; “tag”: | comedy | “description”: {description};\n“tag”: {tag} |
| LaMP-3 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, what is the score of the following review on a scale of 1 to 5? just answer with 1, 2, 3, 4, or 5 without further explanation. “review”: {review}; “score”: | 5 | “review”: {review}\n“score”: {score} |
| LaMP-4 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please generate a title for the given user's input text. Please generate it in the following format: {"title": "generated title"} without explanation, and use only English. “text”: {text}; “title”: | {“title”: Finding Happiness \nAfter Divorce - It Can Happen} | “text”: {text}\n“title”: {title} |
| LaMP-5 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the historical profiles provided, please generate a title for the given user's input abstract. Please generate it in the following format: {"title": "generated title"} without explanation, and use only English. “abstract”: {abstract}; “title”: | {“title”: Link-Reliability Based \nTwo-Hop Routing for \nWireless Sensor Networks.} | “abstract”: {abstract}\n“title”: {title} |
| LaMP-7 | The historical profiles are as follows: {history1} ... {historyk}. \nBased on the style pattern of the historical tweets provided, please paraphrase the user's input tweet without any explanation before or after it. Please generate it in the following format: {"tweet": "generated tweet"} without explanation, and use only English. “tweet”: {tweet}. | {“tweet”:lilxcutiesworld the \ndanny picture is GOOD!! \nI really like it.} | “tweet”: {tweet} |
",
+ "image_path": "6cea494ce69dd19e4155bbac2bdb81b8052ffdd33a3f5e4661fcc323a21cbc4f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 50,
+ 500,
+ 295,
+ 610
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 500,
+ 295,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 500,
+ 295,
+ 610
+ ],
+ "type": "text",
+ "content": "5.4.4 Impact of the Number of Retrieved Users. Since we enhance personalized text generation by introducing collaborative filtering, we further explored how much collaborative information to introduce, specifically the impact of the number of retrieved users on the results, as shown in Figure 8. In LaMP-1, retrieving too few or too many users leads to poorer performance, with the best results at 4 users. In LaMP-5, the performance improves as the number of users increases. This highlights the importance of introducing collaborative filtering, but it also indicates that excessive introduction can lead to decreased effectiveness."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 50,
+ 622,
+ 295,
+ 710
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 622,
+ 295,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 622,
+ 295,
+ 710
+ ],
+ "type": "text",
+ "content": "5.4.5 Impact of the Number of Retrieved Documents. We also analyzed the impact of the number of retrieved documents, "
+ },
+ {
+ "bbox": [
+ 50,
+ 622,
+ 295,
+ 710
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 50,
+ 622,
+ 295,
+ 710
+ ],
+ "type": "text",
+ "content": ", on the results, as shown in Figure 9. It can be observed that as the number of retrieved documents increases, performance improves, indicating the importance of retrieving user history to reflect user preferences for enhancing LLM-generated results. Since more documents lead to longer prompts and slower LLM generation, we chose "
+ },
+ {
+ "bbox": [
+ 50,
+ 622,
+ 295,
+ 710
+ ],
+ "type": "inline_equation",
+ "content": "k = 5"
+ },
+ {
+ "bbox": [
+ 50,
+ 622,
+ 295,
+ 710
+ ],
+ "type": "text",
+ "content": " for our experiments."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 315,
+ 499,
+ 391,
+ 510
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 499,
+ 391,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 499,
+ 391,
+ 510
+ ],
+ "type": "text",
+ "content": "6 Conclusion"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 313,
+ 514,
+ 560,
+ 634
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 514,
+ 560,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 514,
+ 560,
+ 634
+ ],
+ "type": "text",
+ "content": "In this paper, we propose CFRAG, which adapts collaborative filtering into RAG to personalize LLMs. To introduce collaborative information without explicit user labels and retrieve documents that support personalized LLM generation, we first train user embeddings through contrastive learning to retrieve similar users. Then, we design the personalized retriever and reranker that considers user preferences during retrieval and fine-tune them using LLM feedback. The results on the Language Model Personalization (LaMP) benchmark validate the effectiveness of CFRAG. The experimental analysis also confirms the effectiveness of each module within CFRAG."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 315,
+ 673,
+ 434,
+ 685
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 673,
+ 434,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 673,
+ 434,
+ 685
+ ],
+ "type": "text",
+ "content": "A Appendix: Prompts"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 314,
+ 687,
+ 560,
+ 709
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 687,
+ 560,
+ 709
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 687,
+ 560,
+ 709
+ ],
+ "type": "text",
+ "content": "We provide detailed formats for the inputs, outputs, and user histories for the LLM across different datasets, as shown in Table 4."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 326,
+ 69
+ ],
+ "type": "text",
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 440,
+ 60,
+ 559,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 440,
+ 60,
+ 559,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 440,
+ 60,
+ 559,
+ 69
+ ],
+ "type": "text",
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 8
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 52,
+ 83,
+ 108,
+ 95
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 83,
+ 108,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 83,
+ 108,
+ 95
+ ],
+ "type": "text",
+ "content": "References"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 53,
+ 97,
+ 295,
+ 710
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 29,
+ "blocks": [
+ {
+ "bbox": [
+ 56,
+ 97,
+ 294,
+ 113
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 97,
+ 294,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 97,
+ 294,
+ 113
+ ],
+ "type": "text",
+ "content": "[1] AI@Meta. 2024. Llama 3 Model Card. (2024). https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 56,
+ 114,
+ 295,
+ 137
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 114,
+ 295,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 114,
+ 295,
+ 137
+ ],
+ "type": "text",
+ "content": "[2] Akari Asai, Zegiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. [n.d.]. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. In The Twelfth International Conference on Learning Representations."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 56,
+ 137,
+ 294,
+ 161
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 137,
+ 294,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 137,
+ 294,
+ 161
+ ],
+ "type": "text",
+ "content": "[3] Sebastian Borgeaud, Arthur Mensch, et al. 2022. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, PMLR, 2206-2240."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 56,
+ 162,
+ 294,
+ 184
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 162,
+ 294,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 162,
+ 294,
+ 184
+ ],
+ "type": "text",
+ "content": "[4] Jin Chen, Zheng Liu, et al. 2024. When large language models meet personalization: Perspectives of challenges and opportunities. World Wide Web 27, 4 (2024), 42."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 56,
+ 185,
+ 295,
+ 209
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 185,
+ 295,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 185,
+ 295,
+ 209
+ ],
+ "type": "text",
+ "content": "[5] Sunhao Dai, Ninglu Shao, et al. 2023. Uncovering chatgpt's capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems. 1126-1132."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 56,
+ 209,
+ 294,
+ 249
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 209,
+ 294,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 209,
+ 294,
+ 249
+ ],
+ "type": "text",
+ "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 56,
+ 250,
+ 294,
+ 281
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 250,
+ 294,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 250,
+ 294,
+ 281
+ ],
+ "type": "text",
+ "content": "[7] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 6491-6501."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 56,
+ 281,
+ 294,
+ 304
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 281,
+ 294,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 281,
+ 294,
+ 304
+ ],
+ "type": "text",
+ "content": "[8] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 56,
+ 305,
+ 294,
+ 328
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 56,
+ 305,
+ 294,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 56,
+ 305,
+ 294,
+ 328
+ ],
+ "type": "text",
+ "content": "[9] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. In International conference on machine learning. PMLR, 3929-3938."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 53,
+ 329,
+ 294,
+ 360
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 329,
+ 294,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 329,
+ 294,
+ 360
+ ],
+ "type": "text",
+ "content": "[10] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639-648."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 53,
+ 361,
+ 294,
+ 384
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 361,
+ 294,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 361,
+ 294,
+ 384
+ ],
+ "type": "text",
+ "content": "[11] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web. 173-182."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 53,
+ 384,
+ 294,
+ 408
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 384,
+ 294,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 384,
+ 294,
+ 408
+ ],
+ "type": "text",
+ "content": "[12] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. [n.d.]. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 53,
+ 409,
+ 294,
+ 440
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 409,
+ 294,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 409,
+ 294,
+ 440
+ ],
+ "type": "text",
+ "content": "[13] Gautier Izacard and Edouard Grave. 2021. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 874-880."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 53,
+ 441,
+ 294,
+ 456
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 441,
+ 294,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 441,
+ 294,
+ 456
+ ],
+ "type": "text",
+ "content": "[14] Gautier Izacard, Patrick Lewis, et al. 2022. Few-shot learning with retrieval augmented language models. arXiv preprint arXiv:2208.03299 1, 2 (2022), 4."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 53,
+ 457,
+ 294,
+ 480
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 457,
+ 294,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 457,
+ 294,
+ 480
+ ],
+ "type": "text",
+ "content": "[15] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. 2020. A survey on contrastive self-supervised learning. Technologies 9, 1 (2020), 2."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 53,
+ 481,
+ 294,
+ 504
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 481,
+ 294,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 481,
+ 294,
+ 504
+ ],
+ "type": "text",
+ "content": "[16] Joel Jang, Seungone Kim, et al. 2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv preprint arXiv:2310.11564 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 53,
+ 505,
+ 294,
+ 528
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 505,
+ 294,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 505,
+ 294,
+ 528
+ ],
+ "type": "text",
+ "content": "[17] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In International Conference on Machine Learning. PMLR, 15696-15707."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 53,
+ 529,
+ 294,
+ 544
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 529,
+ 294,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 529,
+ 294,
+ 544
+ ],
+ "type": "text",
+ "content": "[18] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 53,
+ 544,
+ 294,
+ 559
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 544,
+ 294,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 544,
+ 294,
+ 559
+ ],
+ "type": "text",
+ "content": "[19] Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer 42, 8 (2009), 30-37."
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 53,
+ 560,
+ 294,
+ 583
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 560,
+ 294,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 560,
+ 294,
+ 583
+ ],
+ "type": "text",
+ "content": "[20] Patrick Lewis, Ethan Perez, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33 (2020), 9459-9474."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 53,
+ 584,
+ 294,
+ 608
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 584,
+ 294,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 584,
+ 294,
+ 608
+ ],
+ "type": "text",
+ "content": "[21] Cheng Li, Mingyang Zhang, Qiaozhu Mei, Yaqing Wang, Spurthi Amba Hombaiah, Yi Liang, and Michael Bendersky. 2023. Teach LLMs to Personalize-An Approach inspired by Writing Education. arXiv preprint arXiv:2308.07968 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 53,
+ 609,
+ 294,
+ 632
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 609,
+ 294,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 609,
+ 294,
+ 632
+ ],
+ "type": "text",
+ "content": "[22] Junyi Li, Tianyi Tang, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2024. Pre-trained language models for text generation: A survey. Comput. Surveys 56, 9 (2024), 1-39."
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 53,
+ 632,
+ 294,
+ 647
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 632,
+ 294,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 632,
+ 294,
+ 647
+ ],
+ "type": "text",
+ "content": "[23] Xinyu Li, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 53,
+ 648,
+ 294,
+ 663
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 648,
+ 294,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 648,
+ 294,
+ 663
+ ],
+ "type": "text",
+ "content": "[24] Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out. 74-81."
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 53,
+ 664,
+ 294,
+ 695
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 664,
+ 294,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 664,
+ 294,
+ 695
+ ],
+ "type": "text",
+ "content": "[25] Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. [n.d.]. RA-DIT: Retrieval-Augmented Dual Instruction Tuning. In The Twelfth International Conference on Learning Representations."
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 53,
+ 696,
+ 294,
+ 710
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 696,
+ 294,
+ 710
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 696,
+ 294,
+ 710
+ ],
+ "type": "text",
+ "content": "[26] Yinhan Liu. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)."
+ }
+ ]
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 316,
+ 86,
+ 558,
+ 708
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 54,
+ "blocks": [
+ {
+ "bbox": [
+ 316,
+ 86,
+ 558,
+ 110
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 316,
+ 86,
+ 558,
+ 110
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 316,
+ 86,
+ 558,
+ 110
+ ],
+ "type": "text",
+ "content": "[27] Sheshera Mysore, Zhuoran Lu, et al. 2023. Pearl: Personalizing large language model writing assistants with generation-calibrated retrievers. arXiv preprint arXiv:2311.09180 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 317,
+ 111,
+ 558,
+ 126
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 111,
+ 558,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 111,
+ 558,
+ 126
+ ],
+ "type": "text",
+ "content": "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)."
+ }
+ ]
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 317,
+ 127,
+ 558,
+ 158
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 127,
+ 558,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 127,
+ 558,
+ 158
+ ],
+ "type": "text",
+ "content": "[29] Chris Richardson, Yao Zhang, Kellen Gillespie, Sudipta Kar, Arshdeep Singh, Zeynab Raeesy, Omar Zia Khan, and Abhinav Sethy. 2023. Integrating summarization and retrieval for enhanced personalization via large language models. arXiv preprint arXiv:2310.20081 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 317,
+ 159,
+ 558,
+ 182
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 159,
+ 558,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 159,
+ 558,
+ 182
+ ],
+ "type": "text",
+ "content": "[30] Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp 109 (1995), 109."
+ }
+ ]
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 317,
+ 182,
+ 558,
+ 213
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 182,
+ 558,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 182,
+ 558,
+ 213
+ ],
+ "type": "text",
+ "content": "[31] Alireza Salemi, Surya Kallumadi, and Hamed Zamani. 2024. Optimization methods for personalizing large language models through retrieval augmentation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 752-762."
+ }
+ ]
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 317,
+ 214,
+ 558,
+ 237
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 214,
+ 558,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 214,
+ 558,
+ 237
+ ],
+ "type": "text",
+ "content": "[32] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2023. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 317,
+ 238,
+ 558,
+ 261
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 238,
+ 558,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 238,
+ 558,
+ 261
+ ],
+ "type": "text",
+ "content": "[33] Chenglei Shen, Xiao Zhang, Teng Shi, Changshuo Zhang, Guofu Xie, and Jun Xu. 2024. A survey of controllable learning: Methods and applications in information retrieval. arXiv preprint arXiv:2407.06083 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 317,
+ 262,
+ 558,
+ 301
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 262,
+ 558,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 262,
+ 558,
+ 301
+ ],
+ "type": "text",
+ "content": "[34] Teng Shi, Zihua Si, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Dewei Leng, Yanan Niu, and Yang Song. 2024. UniSAR: Modeling User Transition Behaviors between Search and Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1029-1039."
+ }
+ ]
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 317,
+ 302,
+ 558,
+ 342
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 302,
+ 558,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 302,
+ 558,
+ 342
+ ],
+ "type": "text",
+ "content": "[35] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2024. REPLUG: Retrieval-Augmented Black-Box Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 8364-8377."
+ }
+ ]
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 317,
+ 342,
+ 558,
+ 365
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 342,
+ 558,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 342,
+ 558,
+ 365
+ ],
+ "type": "text",
+ "content": "[36] Zhaoxuan Tan, Zheyuan Liu, and Meng Jiang. 2024. Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts. arXiv preprint arXiv:2406.10471 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 317,
+ 366,
+ 558,
+ 389
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 366,
+ 558,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 366,
+ 558,
+ 389
+ ],
+ "type": "text",
+ "content": "[37] Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2024. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. arXiv:2402.04401 [cs.CL] https://arxiv.org/abs/2402.04401"
+ }
+ ]
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 317,
+ 390,
+ 558,
+ 421
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 390,
+ 558,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 390,
+ 558,
+ 421
+ ],
+ "type": "text",
+ "content": "[38] Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Wu Jian, and Yuning Jiang. 2025. Think Before Recommend: Unleashing the Latent Reasoning Power for Sequential Recommendation. arXiv:2503.22675 [cs.IR] https://arxiv.org/abs/2503.22675"
+ }
+ ]
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 317,
+ 422,
+ 558,
+ 437
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 422,
+ 558,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 422,
+ 558,
+ 437
+ ],
+ "type": "text",
+ "content": "[39] A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)."
+ }
+ ]
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 317,
+ 437,
+ 558,
+ 461
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 437,
+ 558,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 437,
+ 558,
+ 461
+ ],
+ "type": "text",
+ "content": "[40] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval. 165-174."
+ }
+ ]
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 317,
+ 462,
+ 558,
+ 485
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 462,
+ 558,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 462,
+ 558,
+ 485
+ ],
+ "type": "text",
+ "content": "[41] Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al. 2024. A survey on large language models for recommendation. World Wide Web 27, 5 (2024), 60."
+ }
+ ]
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 317,
+ 486,
+ 558,
+ 509
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 486,
+ 558,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 486,
+ 558,
+ 509
+ ],
+ "type": "text",
+ "content": "[42] Xinghao Wu, Xuefeng Liu, Jianwei Niu, Haolin Wang, Shaojie Tang, and Guogang Zhu. 2024. FedLoRA: When Personalized Federated Learning Meets Low-Rank Adaptation. (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 317,
+ 510,
+ 558,
+ 540
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 510,
+ 558,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 510,
+ 558,
+ 540
+ ],
+ "type": "text",
+ "content": "[43] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. 2024. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems 36 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 317,
+ 541,
+ 558,
+ 564
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 541,
+ 558,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 541,
+ 558,
+ 564
+ ],
+ "type": "text",
+ "content": "[44] Zhuofeng Wu, Sinong Wang, Jiatao Gu, Madian Khabsa, Fei Sun, and Hao Ma. 2020. Clear: Contrastive learning for sentence representation. arXiv preprint arXiv:2012.15466 (2020)."
+ }
+ ]
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 317,
+ 565,
+ 558,
+ 589
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 565,
+ 558,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 565,
+ 558,
+ 589
+ ],
+ "type": "text",
+ "content": "[45] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597 [cs.CL]"
+ }
+ ]
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 317,
+ 590,
+ 558,
+ 612
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 590,
+ 558,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 590,
+ 558,
+ 612
+ ],
+ "type": "text",
+ "content": "[46] Hong-Jian Xue, Xinyu Dai, Jianbing Zhang, Shujian Huang, and Jiajun Chen. 2017. Deep matrix factorization models for recommender systems.. In IJCAI, Vol. 17. Melbourne, Australia, 3203-3209."
+ }
+ ]
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 317,
+ 613,
+ 558,
+ 636
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 613,
+ 558,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 613,
+ 558,
+ 636
+ ],
+ "type": "text",
+ "content": "[47] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng-peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 317,
+ 637,
+ 558,
+ 661
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 637,
+ 558,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 637,
+ 558,
+ 661
+ ],
+ "type": "text",
+ "content": "[48] Changshuo Zhang, Teng Shi, Xiao Zhang, Qi Liu, Ruobing Xie, Jun Xu, and Ji-Rong Wen. 2024. Modeling Domain and Feedback Transitions for Cross-Domain Sequential Recommendation. arXiv preprint arXiv:2408.08209 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 317,
+ 662,
+ 558,
+ 685
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 662,
+ 558,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 662,
+ 558,
+ 685
+ ],
+ "type": "text",
+ "content": "[49] Changshuo Zhang, Teng Shi, Xiao Zhang, Yanping Zheng, Ruobing Xie, Qi Liu, Jun Xu, and Ji-Rong Wen. 2024. QAGCF: Graph Collaborative Filtering for Q&A Recommendation. arXiv preprint arXiv:2406.04828 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 317,
+ 685,
+ 558,
+ 708
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 685,
+ 558,
+ 708
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 685,
+ 558,
+ 708
+ ],
+ "type": "text",
+ "content": "[50] Changshuo Zhang, Xiao Zhang, Teng Shi, Jun Xu, and Ji-Rong Wen. 2025. Test-Time Alignment for Tracking User Interest Shifts in Sequential Recommendation. arXiv:2504.01489 [cs.IR] https://arxiv.org/abs/2504.01489"
+ }
+ ]
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 171,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 171,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 171,
+ 69
+ ],
+ "type": "text",
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 515,
+ 60,
+ 558,
+ 69
+ ],
+ "type": "text",
+ "content": "Teng Shi et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 9
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 52,
+ 86,
+ 296,
+ 175
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 5,
+ "blocks": [
+ {
+ "bbox": [
+ 52,
+ 86,
+ 296,
+ 119
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 86,
+ 296,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 86,
+ 296,
+ 119
+ ],
+ "type": "text",
+ "content": "[51] Kepu Zhang, Teng Shi, Sunhao Dai, Xiao Zhang, Yinfeng Li, Jing Lu, Xiaoxue Zang, Yang Song, and Jun Xu. 2024. SAQRec: Aligning Recommender Systems to User Satisfaction via Questionnaire Feedback. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 3165-3175."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 53,
+ 119,
+ 296,
+ 144
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 119,
+ 296,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 119,
+ 296,
+ 144
+ ],
+ "type": "text",
+ "content": "[52] Xiao Zhang, Teng Shi, Jun Xu, Zhenhua Dong, and Ji-Rong Wen. 2024. Model-Agnostic Causal Embedding Learning for Counterfactually Group-Fair Recommendation. IEEE Transactions on Knowledge and Data Engineering (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 53,
+ 144,
+ 294,
+ 175
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 144,
+ 294,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 144,
+ 294,
+ 175
+ ],
+ "type": "text",
+ "content": "[53] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 316,
+ 86,
+ 559,
+ 182
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 10,
+ "blocks": [
+ {
+ "bbox": [
+ 316,
+ 86,
+ 558,
+ 110
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 316,
+ 86,
+ 558,
+ 110
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 316,
+ 86,
+ 558,
+ 110
+ ],
+ "type": "text",
+ "content": "[54] Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2024. Dense text retrieval based on pretrained language models: A survey. ACM Transactions on Information Systems 42, 4 (2024), 1-60."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 317,
+ 111,
+ 559,
+ 135
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 111,
+ 559,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 111,
+ 559,
+ 135
+ ],
+ "type": "text",
+ "content": "[55] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 316,
+ 135,
+ 559,
+ 159
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 316,
+ 135,
+ 559,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 316,
+ 135,
+ 559,
+ 159
+ ],
+ "type": "text",
+ "content": "[56] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 316,
+ 159,
+ 559,
+ 182
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 316,
+ 159,
+ 559,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 316,
+ 159,
+ 559,
+ 182
+ ],
+ "type": "text",
+ "content": "[57] Yuchen Zhuang, Haotian Sun, Yue Yu, Qifan Wang, Chao Zhang, and Bo Dai. 2024. HYDRA: Model Factorization Framework for Black-Box LLM Personalization. arXiv preprint arXiv:2406.02888 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 326,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 326,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 60,
+ 326,
+ 69
+ ],
+ "type": "text",
+ "content": "Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 440,
+ 60,
+ 558,
+ 69
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 440,
+ 60,
+ 558,
+ 69
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 440,
+ 60,
+ 558,
+ 69
+ ],
+ "type": "text",
+ "content": "SIGIR '25, July 13-18, 2025, Padua, Italy."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 10
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/data/2025/2504_05xxx/2504.05741/97d4b9fa-4704-4c9a-9165-90317d2d4443_content_list.json b/data/2025/2504_05xxx/2504.05741/97d4b9fa-4704-4c9a-9165-90317d2d4443_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..67ffa72f7c5f86746c49c8383d85e117c4d4851e
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05741/97d4b9fa-4704-4c9a-9165-90317d2d4443_content_list.json
@@ -0,0 +1,2564 @@
+[
+ {
+ "type": "text",
+ "text": "DDT: Decoupled Diffusion Transformer",
+ "text_level": 1,
+ "bbox": [
+ 294,
+ 130,
+ 702,
+ 152
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Shuai Wang| Source (HuggingFace Datasets) | Languages | Size |
| BAAI/Infinity-Instruct^* | EN | 1,456,927 |
| HuggingFaceTB/smoltalk | EN | 409,537 |
| allenai/tulu-3-sft-personas-math | EN | 149,960 |
| parinzee/seed-free-synthetic-instruct-thai-v1 | TH | 118,898 |
| HuggingFaceTB/smoltalk | EN | 96,356 |
| HuggingFaceTB/smoltalk | EN | 83,144 |
| arcee-ai/EvolKit-75K | EN | 74,174 |
| AI-MO/NuminaMath-TIR | EN | 72,441 |
| Post-training-Data-Flywheel/AutoIF-instruct-61k | EN | 61,492 |
| argilla/ifeval-like-data | EN | 56,339 |
| HuggingFaceTB/smoltalk | EN | 53,342 |
| ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k | EN | 50,000 |
| ai2-adapt-dev/tulu_v3.9_synthetic_finalresp_wildguardmixtrain_decontaminated_50k | EN | 50,000 |
| allenai/tulu-3-sft-personas-math-grade | EN | 49,980 |
| allenai/tulu-3-sft-personas-code | EN | 34,999 |
| HuggingFaceTB/smoltalk | EN | 34,424 |
| allenai/tulu-3-sft-personas-instruction-following | EN | 29,980 |
| airesearch/WangchanThaiInstruct | TH | 25,014 |
| allenai/tulu-3-sft-personas-algebra | EN | 20,000 |
| arcee-ai/EvolKit-20k-vi | VI | 15,378 |
| allenai/coconot | EN | 10,983 |
| ai2-adapt-dev/tulu_v3.9_scirff_10k | EN | 10,000 |
| Source (Generated) | Languages | Size |
| qwen_gemma_synthetic_tamil | TA | 480,000 |
| qwen_gemma_synthetic_thai | TH | 480,000 |
| qwen_gemma_synthetic_indonesian | ID | 465,019 |
| qwen_gemma_synthetic_vietnamese | VI | 465,019 |
| gemma_synthetic_indonesian | ID | 458,149 |
| gemma_synthetic_filipino | TL | 455,093 |
| gemma_synthetic_viet | VI | 291,576 |
| gemma_synthetic_tamil | TA | 276,314 |
| gemma_synthetic_thai | TH | 186,339 |
| gemma_synthetic_javanese | JV | 110,000 |
| gemma_synthetic_sudanese | SU | 110,000 |
| llama_gemma_synthetic_thai | TH | 88,920 |
| llama_gemma_synthetic_tamil | TA | 88,920 |
| llama_gemma_synthetic_vietnamese | VI | 88,920 |
| llama_gemma_synthetic_javanese | JV | 88,920 |
| llama_gemma_synthetic_indonesian | ID | 88,920 |
| llama_gemma_synthetic_filipino | TL | 80,000 |
| enrich_27k | SEA | 27,463 |
| seaMultilingual_systemchat | SEA | 1,903 |
| rewritten_oasst | SEA | 841 |
| rewritten_helpsteer | SEA | 838 |
",
+ "bbox": [
+ 201,
+ 332,
+ 796,
+ 843
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Table 9: List of datasets for Stage-2-IFT.",
+ "bbox": [
+ 359,
+ 853,
+ 636,
+ 866
+ ],
+ "page_idx": 14
+ }
+]
\ No newline at end of file
diff --git a/data/2025/2504_05xxx/2504.05747/ec2428c8-34ac-40c3-9d6a-d16ed6c46317_model.json b/data/2025/2504_05xxx/2504.05747/ec2428c8-34ac-40c3-9d6a-d16ed6c46317_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..406c6942e81eebaef161dbebed1111ed13b908df
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/ec2428c8-34ac-40c3-9d6a-d16ed6c46317_model.json
@@ -0,0 +1,2507 @@
+[
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.165,
+ 0.083,
+ 0.231,
+ 0.128
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.239,
+ 0.096,
+ 0.829,
+ 0.117
+ ],
+ "angle": 0,
+ "content": "SEA-LION: Southeast Asian Languages in One Network"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.13,
+ 0.128,
+ 0.871,
+ 0.294
+ ],
+ "angle": 0,
+ "content": "Raymond Ng\\*, Thanh Ngan Nguyen\\*, Yuli Huang\\*, Ngee Chia Tai\\*, Wai Yi Leong\\*, Wei Qi Leong\\*, Xianbin Yong\\*, Jian Gang Ngui\\*, Yosephine Susanto\\*, Nicholas Cheng\\*, Hamsawardhini Rengarajan\\*, Peerat Limkonchotiwat\\*, Adithya Venkatadri Hulagadri\\*, Kok Wai Teng\\*, Yeo Yeow Tong\\*, Bryan Siow\\*, Wei Yi Teo\\*, Wayne Lau\\*, Choon Meng Tan\\*, Brandon Ong\\*, Zhi Hao Ong\\*, Jann Railey Montalan\\*, Adwin Chan\\*, Sajeban Antonyrex\\*, Ren Lee\\*, Esther Choa\\*, David Ong Tat-Wee\\*, Bing Jie Darius Liu\\*, William Chandra Tjhi\\*, Erik Cambria\\*, Leslie Teo\\* AI Singapore, National University of Singapore \\*Nanyang Technological University https://sea-lion.ai"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.261,
+ 0.31,
+ 0.341,
+ 0.324
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.142,
+ 0.334,
+ 0.461,
+ 0.69
+ ],
+ "angle": 0,
+ "content": "Recently, Large Language Models (LLMs) have dominated much of the artificial intelligence scene with their ability to process and generate natural languages. However, the majority of LLM research and development remains English-centric, leaving low-resource languages such as those in the Southeast Asian (SEA) region underrepresented. To address this representation gap, we introduce Llama-SEA-LION-8B-IT and Gemma-SEA-LION-9B-IT, two cutting-edge multilingual LLMs designed for SEA languages. The SEA-LION family of LLMs supports 11 SEA languages, namely English, Chinese, Indonesian, Vietnamese, Malay, Thai, Burmese, Lao, Filipino, Tamil, and Khmer. Our work leverages large-scale multilingual continued pre-training with a comprehensive post-training regime involving multiple stages of instruction fine-tuning, alignment, and model merging. Evaluation results on multilingual benchmarks show that our models achieve state-of-the-art performance across LLMs supporting SEA languages. We open-source the models to benefit the wider SEA community."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.698,
+ 0.26,
+ 0.713
+ ],
+ "angle": 0,
+ "content": "1 Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.723,
+ 0.49,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Large language models (LLMs) have significantly transformed the field of natural language processing, achieving remarkable performance in text generation, summarization and sentiment analysis (Brown et al., 2020; OpenAI, 2023; Dubey et al., 2024; Rivière et al., 2024; Zhang et al., 2024b; Yeo et al., 2024). Despite their impressive capabilities, most LLMs remain heavily English-centric (Wendler et al., 2024; Zhong et al., 2024). Unfortunately, this situation has led LLMs in regions with many under-represented languages such"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.31,
+ 0.885,
+ 0.422
+ ],
+ "angle": 0,
+ "content": "as Southeast Asia (SEA) to suffer. Languages with lower resources, such as Filipino, Lao, Burmese and Khmer in the SEA region, are not supported by many open-source English-centric LLMs. This underscores the need to bridge the resource and representation gap between English and SEA languages."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.426,
+ 0.885,
+ 0.763
+ ],
+ "angle": 0,
+ "content": "Recently, there have been many attempts to create multilingual LLMs in an open-source manner, e.g., BLOOM (Scao et al., 2022), a project aimed at increasing multilingual presence in opensource LLMs by supporting 46 languages. Popular LLM families such as Llama (Dubey et al., 2024), Gemma (Rivière et al., 2024) and Qwen (Yang et al., 2024a) have also introduced multilingual LLMs for their latest iteration. During our evaluations, we found that the performance of these models is acceptable in the general case, i.e., when considering evaluation benchmarks formulated from English datasets. However, we observe that the performance degrades on SEA-specific benchmarks. Moreover, researchers have also introduced LLMs such as SeaLLMs (Nguyen et al., 2024; Zhang et al., 2024a) and Sailor (Dou et al., 2024) to specifically address the LLM gap in SEA languages. However, the performance of these models is less than ideal for languages such as Thai or Tamil\\(^{2}\\) (10X et al., 2024; AI Products Team, 2024)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.508,
+ 0.767,
+ 0.885,
+ 0.879
+ ],
+ "angle": 0,
+ "content": "In this paper, we address the issues by proposing a robust open-source Southeast Asian model with data transparency for reproducibility, namely SEA-LION - a family of LLMs continued pretrained (CPT) and fine-tuned on Llama-3.1-8B-Instruct for Llama-SEA-LION-8B-IT and Gemma2-9B for Gemma-SEA-LION-9B-IT with a focus"
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.137,
+ 0.907,
+ 0.332,
+ 0.92
+ ],
+ "angle": 0,
+ "content": "\\(^{1}\\)SEA-LION Models Collection"
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.509,
+ 0.895,
+ 0.883,
+ 0.922
+ ],
+ "angle": 0,
+ "content": "2Tamil is one of the official languages in Singapore. It is also spoken in other areas in the SEA region, such as Malaysia."
+ },
+ {
+ "type": "aside_text",
+ "bbox": [
+ 0.023,
+ 0.291,
+ 0.061,
+ 0.706
+ ],
+ "angle": 270,
+ "content": "arXiv:2504.05747v4 [cs.CL] 30 Oct 2025"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.113,
+ 0.085,
+ 0.49,
+ 0.325
+ ],
+ "angle": 0,
+ "content": "on SEA languages. To tackle the performance problem, we utilize 200 billion English, code, and SEA languages tokens as well as 16.8 million English and SEA languages instruction and answer pairs for CPT and post-training steps, respectively, to achieve a significant improvement in SEA languages. In order to allow our models to be used by everyone without restrictions, we release our models under the fully open MIT license. We benchmark our models against the SEA-HELM(Susanto et al., 2025) and Open LLM Leaderboard| Source (HuggingFace Datasets) | Languages | Size |
| BAAI/Infinity-Instruct^* | EN | 1,456,927 |
| HuggingFaceTB/smoltalk | EN | 409,537 |
| allenai/tulu-3-sft-personas-math | EN | 149,960 |
| parinzee/seed-free-synthetic-instruct-thai-v1 | TH | 118,898 |
| HuggingFaceTB/smoltalk | EN | 96,356 |
| HuggingFaceTB/smoltalk | EN | 83,144 |
| arcee-ai/EvolKit-75K | EN | 74,174 |
| AI-MO/NuminaMath-TIR | EN | 72,441 |
| Post-training-Data-Flywheel/AutoIF-instruct-61k | EN | 61,492 |
| argilla/ifeval-like-data | EN | 56,339 |
| HuggingFaceTB/smoltalk | EN | 53,342 |
| ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k | EN | 50,000 |
| ai2-adapt-dev/tulu_v3.9_synthetic_finalresp_wildguardmixtrain_decontaminated_50k | EN | 50,000 |
| allenai/tulu-3-sft-personas-math-grade | EN | 49,980 |
| allenai/tulu-3-sft-personas-code | EN | 34,999 |
| HuggingFaceTB/smoltalk | EN | 34,424 |
| allenai/tulu-3-sft-personas-instruction-following | EN | 29,980 |
| airesearch/WangchanThaiInstruct | TH | 25,014 |
| allenai/tulu-3-sft-personas-algebra | EN | 20,000 |
| arcee-ai/EvolKit-20k-vi | VI | 15,378 |
| allenai/coconot | EN | 10,983 |
| ai2-adapt-dev/tulu_v3.9_scirff_10k | EN | 10,000 |
| Source (Generated) | Languages | Size |
| qwen_gemma_synthetic_tamil | TA | 480,000 |
| qwen_gemma_synthetic_thai | TH | 480,000 |
| qwen_gemma_synthetic_indonesian | ID | 465,019 |
| qwen_gemma_synthetic_vietnamese | VI | 465,019 |
| gemma_synthetic_indonesian | ID | 458,149 |
| gemma_synthetic_filipino | TL | 455,093 |
| gemma_synthetic_viet | VI | 291,576 |
| gemma_synthetic_tamil | TA | 276,314 |
| gemma_synthetic_thai | TH | 186,339 |
| gemma_synthetic_javanese | JV | 110,000 |
| gemma_synthetic_sudanese | SU | 110,000 |
| llama_gemma_synthetic_thai | TH | 88,920 |
| llama_gemma_synthetic_tamil | TA | 88,920 |
| llama_gemma_synthetic_vietnamese | VI | 88,920 |
| llama_gemma_synthetic_javanese | JV | 88,920 |
| llama_gemma_synthetic_indonesian | ID | 88,920 |
| llama_gemma_synthetic_filipino | TL | 80,000 |
| enrich_27k | SEA | 27,463 |
| seaMultilingual_systemchat | SEA | 1,903 |
| rewritten_oasst | SEA | 841 |
| rewritten_helpsteer | SEA | 838 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.36,
+ 0.854,
+ 0.637,
+ 0.868
+ ],
+ "angle": 0,
+ "content": "Table 9: List of datasets for Stage-2-IFT."
+ }
+ ]
+]
\ No newline at end of file
diff --git a/data/2025/2504_05xxx/2504.05747/ec2428c8-34ac-40c3-9d6a-d16ed6c46317_origin.pdf b/data/2025/2504_05xxx/2504.05747/ec2428c8-34ac-40c3-9d6a-d16ed6c46317_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..20d6a57cd2f16c6e0d5a9a005a1b862648a14970
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/ec2428c8-34ac-40c3-9d6a-d16ed6c46317_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a9ab244e0d795e5a10e2b89a20072a80f0fabbf9f64431a181dbc153a00d533
+size 2395052
diff --git a/data/2025/2504_05xxx/2504.05747/full.md b/data/2025/2504_05xxx/2504.05747/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..6292a968ab0d78f8c1e9a38b672444f29115ec9d
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/full.md
@@ -0,0 +1,371 @@
+
+
+# SEA-LION: Southeast Asian Languages in One Network
+
+Raymond Ng\*, Thanh Ngan Nguyen\*, Yuli Huang\*, Ngee Chia Tai\*, Wai Yi Leong\*, Wei Qi Leong\*, Xianbin Yong\*, Jian Gang Ngui\*, Yosephine Susanto\*, Nicholas Cheng\*, Hamsawardhini Rengarajan\*, Peerat Limkonchotiwat\*, Adithya Venkatadri Hulagadri\*, Kok Wai Teng\*, Yeo Yeow Tong\*, Bryan Siow\*, Wei Yi Teo\*, Wayne Lau\*, Choon Meng Tan\*, Brandon Ong\*, Zhi Hao Ong\*, Jann Railey Montalan\*, Adwin Chan\*, Sajeban Antonyrex\*, Ren Lee\*, Esther Choa\*, David Ong Tat-Wee\*, Bing Jie Darius Liu\*, William Chandra Tjhi\*, Erik Cambria\*, Leslie Teo\* AI Singapore, National University of Singapore \*Nanyang Technological University https://sea-lion.ai
+
+# Abstract
+
+Recently, Large Language Models (LLMs) have dominated much of the artificial intelligence scene with their ability to process and generate natural languages. However, the majority of LLM research and development remains English-centric, leaving low-resource languages such as those in the Southeast Asian (SEA) region underrepresented. To address this representation gap, we introduce Llama-SEA-LION-8B-IT and Gemma-SEA-LION-9B-IT, two cutting-edge multilingual LLMs designed for SEA languages. The SEA-LION family of LLMs supports 11 SEA languages, namely English, Chinese, Indonesian, Vietnamese, Malay, Thai, Burmese, Lao, Filipino, Tamil, and Khmer. Our work leverages large-scale multilingual continued pre-training with a comprehensive post-training regime involving multiple stages of instruction fine-tuning, alignment, and model merging. Evaluation results on multilingual benchmarks show that our models achieve state-of-the-art performance across LLMs supporting SEA languages. We open-source the models to benefit the wider SEA community.
+
+# 1 Introduction
+
+Large language models (LLMs) have significantly transformed the field of natural language processing, achieving remarkable performance in text generation, summarization and sentiment analysis (Brown et al., 2020; OpenAI, 2023; Dubey et al., 2024; Rivière et al., 2024; Zhang et al., 2024b; Yeo et al., 2024). Despite their impressive capabilities, most LLMs remain heavily English-centric (Wendler et al., 2024; Zhong et al., 2024). Unfortunately, this situation has led LLMs in regions with many under-represented languages such
+
+as Southeast Asia (SEA) to suffer. Languages with lower resources, such as Filipino, Lao, Burmese and Khmer in the SEA region, are not supported by many open-source English-centric LLMs. This underscores the need to bridge the resource and representation gap between English and SEA languages.
+
+Recently, there have been many attempts to create multilingual LLMs in an open-source manner, e.g., BLOOM (Scao et al., 2022), a project aimed at increasing multilingual presence in opensource LLMs by supporting 46 languages. Popular LLM families such as Llama (Dubey et al., 2024), Gemma (Rivière et al., 2024) and Qwen (Yang et al., 2024a) have also introduced multilingual LLMs for their latest iteration. During our evaluations, we found that the performance of these models is acceptable in the general case, i.e., when considering evaluation benchmarks formulated from English datasets. However, we observe that the performance degrades on SEA-specific benchmarks. Moreover, researchers have also introduced LLMs such as SeaLLMs (Nguyen et al., 2024; Zhang et al., 2024a) and Sailor (Dou et al., 2024) to specifically address the LLM gap in SEA languages. However, the performance of these models is less than ideal for languages such as Thai or Tamil $^{2}$ (10X et al., 2024; AI Products Team, 2024).
+
+In this paper, we address the issues by proposing a robust open-source Southeast Asian model with data transparency for reproducibility, namely SEA-LION - a family of LLMs continued pretrained (CPT) and fine-tuned on Llama-3.1-8B-Instruct for Llama-SEA-LION-8B-IT and Gemma2-9B for Gemma-SEA-LION-9B-IT with a focus
+
+on SEA languages. To tackle the performance problem, we utilize 200 billion English, code, and SEA languages tokens as well as 16.8 million English and SEA languages instruction and answer pairs for CPT and post-training steps, respectively, to achieve a significant improvement in SEA languages. In order to allow our models to be used by everyone without restrictions, we release our models under the fully open MIT license. We benchmark our models against the SEA-HELM(Susanto et al., 2025) and Open LLM Leaderboard| Source (HuggingFace Datasets) | Languages | Size |
| BAAI/Infinity-Instruct^* | EN | 1,456,927 |
| HuggingFaceTB/smoltalk | EN | 409,537 |
| allenai/tulu-3-sft-personas-math | EN | 149,960 |
| parinzee/seed-free-synthetic-instruct-thai-v1 | TH | 118,898 |
| HuggingFaceTB/smoltalk | EN | 96,356 |
| HuggingFaceTB/smoltalk | EN | 83,144 |
| arcee-ai/EvolKit-75K | EN | 74,174 |
| AI-MO/NuminaMath-TIR | EN | 72,441 |
| Post-training-Data-Flywheel/AutoIF-instruct-61k | EN | 61,492 |
| argilla/ifeval-like-data | EN | 56,339 |
| HuggingFaceTB/smoltalk | EN | 53,342 |
| ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k | EN | 50,000 |
| ai2-adapt-dev/tulu_v3.9_synthetic_finalresp_wildguardmixtrain_decontaminated_50k | EN | 50,000 |
| allenai/tulu-3-sft-personas-math-grade | EN | 49,980 |
| allenai/tulu-3-sft-personas-code | EN | 34,999 |
| HuggingFaceTB/smoltalk | EN | 34,424 |
| allenai/tulu-3-sft-personas-instruction-following | EN | 29,980 |
| airesearch/WangchanThaiInstruct | TH | 25,014 |
| allenai/tulu-3-sft-personas-algebra | EN | 20,000 |
| arcee-ai/EvolKit-20k-vi | VI | 15,378 |
| allenai/coconot | EN | 10,983 |
| ai2-adapt-dev/tulu_v3.9_scirff_10k | EN | 10,000 |
| Source (Generated) | Languages | Size |
| qwen_gemma_synthetic_tamil | TA | 480,000 |
| qwen_gemma_synthetic_thai | TH | 480,000 |
| qwen_gemma_synthetic_indonesian | ID | 465,019 |
| qwen_gemma_synthetic_vietnamese | VI | 465,019 |
| gemma_synthetic_indonesian | ID | 458,149 |
| gemma_synthetic_filipino | TL | 455,093 |
| gemma_synthetic_viet | VI | 291,576 |
| gemma_synthetic_tamil | TA | 276,314 |
| gemma_synthetic_thai | TH | 186,339 |
| gemma_synthetic_javanese | JV | 110,000 |
| gemma_synthetic_sudanese | SU | 110,000 |
| llama_gemma_synthetic_thai | TH | 88,920 |
| llama_gemma_synthetic_tamil | TA | 88,920 |
| llama_gemma_synthetic_vietnamese | VI | 88,920 |
| llama_gemma_synthetic_javanese | JV | 88,920 |
| llama_gemma_synthetic_indonesian | ID | 88,920 |
| llama_gemma_synthetic_filipino | TL | 80,000 |
| enrich_27k | SEA | 27,463 |
| seaMultilingual_systemchat | SEA | 1,903 |
| rewritten_oasst | SEA | 841 |
| rewritten_helpsteer | SEA | 838 |
+
+Table 9: List of datasets for Stage-2-IFT.
\ No newline at end of file
diff --git a/data/2025/2504_05xxx/2504.05747/images/0f75f24e6df40fe78b1dd64776ad1d19b8df275d0fba107911bc92e1b8175379.jpg b/data/2025/2504_05xxx/2504.05747/images/0f75f24e6df40fe78b1dd64776ad1d19b8df275d0fba107911bc92e1b8175379.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..69d19ebe8fdc3b3f6008d53b745e86e6e7bf7e26
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/0f75f24e6df40fe78b1dd64776ad1d19b8df275d0fba107911bc92e1b8175379.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5c24052879c9fce04fe27ab50f90e4c583a024e28e17cce353dc80c57c05fe6
+size 182451
diff --git a/data/2025/2504_05xxx/2504.05747/images/4867daa85c3c9dd6e65602d241de43dd14c0427a828b2219e89a5413008cb6c3.jpg b/data/2025/2504_05xxx/2504.05747/images/4867daa85c3c9dd6e65602d241de43dd14c0427a828b2219e89a5413008cb6c3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f592767a24e1c6b64a2bb98f3d9e32f3ca39a752
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/4867daa85c3c9dd6e65602d241de43dd14c0427a828b2219e89a5413008cb6c3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95c074a9759b545cf9a173ecd53c9f969f6677fdca1bdb49b038a6ef66a78839
+size 66671
diff --git a/data/2025/2504_05xxx/2504.05747/images/4c339bc0c829686d5690eeb36036737c9fc9fc6bbb2f0e5a2fc19dd66e119f36.jpg b/data/2025/2504_05xxx/2504.05747/images/4c339bc0c829686d5690eeb36036737c9fc9fc6bbb2f0e5a2fc19dd66e119f36.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2ae1bcac8d737aa9ae1c6f80f950074b14b2d214
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/4c339bc0c829686d5690eeb36036737c9fc9fc6bbb2f0e5a2fc19dd66e119f36.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b790cd98a8ef3c20ce3733d32f71cee330ac199803751066d717093a1b765620
+size 42082
diff --git a/data/2025/2504_05xxx/2504.05747/images/4e16748c2d5d2a0090d54b396c70b170858f6680defc5b52a25cfe87b65aacd2.jpg b/data/2025/2504_05xxx/2504.05747/images/4e16748c2d5d2a0090d54b396c70b170858f6680defc5b52a25cfe87b65aacd2.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..b4f1be368424e7ce5eeb604ce7661830a3ab4f0a
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/4e16748c2d5d2a0090d54b396c70b170858f6680defc5b52a25cfe87b65aacd2.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6714721883376e703eecb1f7c06603efd8a8284f381177e9e93711ae2c50cd8b
+size 15753
diff --git a/data/2025/2504_05xxx/2504.05747/images/55f574138c852425fa5e8074e6ad64c6e4c0ac613ddfd8d9c24174ce9df33553.jpg b/data/2025/2504_05xxx/2504.05747/images/55f574138c852425fa5e8074e6ad64c6e4c0ac613ddfd8d9c24174ce9df33553.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..839fb32ded519fcd36fdc7e4cd98041effad773a
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/55f574138c852425fa5e8074e6ad64c6e4c0ac613ddfd8d9c24174ce9df33553.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d494cb5d3b8f85cabdec1daaa3fe5582602187877195a72f87e051105b205b8
+size 13496
diff --git a/data/2025/2504_05xxx/2504.05747/images/78429de77740e6d1e87d54a49d5ac0a30a1623513b766b33f3be80741d5f7606.jpg b/data/2025/2504_05xxx/2504.05747/images/78429de77740e6d1e87d54a49d5ac0a30a1623513b766b33f3be80741d5f7606.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..1941bf6e7f7edaeb6b411322096995d44b8c62f7
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/78429de77740e6d1e87d54a49d5ac0a30a1623513b766b33f3be80741d5f7606.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb30f7e7ba76b5876b24c12121ffbf93f6541ba1048aedc4b8d3b427cc1f9a00
+size 136182
diff --git a/data/2025/2504_05xxx/2504.05747/images/9138939f8b29aa04aaea835f7a6d240c1c7202541e014a22e29a7fadfe0d65fb.jpg b/data/2025/2504_05xxx/2504.05747/images/9138939f8b29aa04aaea835f7a6d240c1c7202541e014a22e29a7fadfe0d65fb.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2f2b38f90b2bc40c40a47a749e63312adff2103f
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/9138939f8b29aa04aaea835f7a6d240c1c7202541e014a22e29a7fadfe0d65fb.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bcb6039b27ad375b63742f55bf92cf9c3c9aaaa2f3ac38f90f115a7eaa0e0ca
+size 38674
diff --git a/data/2025/2504_05xxx/2504.05747/images/95807f122e753274efc4c43da49e9d40d32d3ff051642f5e33ffc65fc2dd8d5a.jpg b/data/2025/2504_05xxx/2504.05747/images/95807f122e753274efc4c43da49e9d40d32d3ff051642f5e33ffc65fc2dd8d5a.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..11b549fb3dbeffda4ed35094335d9137f4763280
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/95807f122e753274efc4c43da49e9d40d32d3ff051642f5e33ffc65fc2dd8d5a.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f72dd780fbbebe34e62cbed2c58aeb0d14397e6feaefcc4b630f88cdcc1a7b9
+size 40420
diff --git a/data/2025/2504_05xxx/2504.05747/images/9f6b9f3761ad7213c0ce54d33552c028e28aae4e28023f2ef7dc3583c49097e7.jpg b/data/2025/2504_05xxx/2504.05747/images/9f6b9f3761ad7213c0ce54d33552c028e28aae4e28023f2ef7dc3583c49097e7.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..94b909b9da022cb708c525fb5fe7ef31b0bcbfea
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/9f6b9f3761ad7213c0ce54d33552c028e28aae4e28023f2ef7dc3583c49097e7.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c3bd2776f5d79890ce859518117788c8ee9b06ba8b1e4e165a8a9199cadf174
+size 4948
diff --git a/data/2025/2504_05xxx/2504.05747/images/d1fb51f7d5203701b1c1be87c57a86727b27c69725ee0cf2689b1ea2cdbdd506.jpg b/data/2025/2504_05xxx/2504.05747/images/d1fb51f7d5203701b1c1be87c57a86727b27c69725ee0cf2689b1ea2cdbdd506.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..cfd5eab31db91021ebc454dd6df97dbff99fd652
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/d1fb51f7d5203701b1c1be87c57a86727b27c69725ee0cf2689b1ea2cdbdd506.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb2739dc25d5729703d4211fd0a609acca60e728fa85059f4241f9f0e5465a71
+size 42352
diff --git a/data/2025/2504_05xxx/2504.05747/images/db3a7dde14ce0b341033213c154d6a9aa45d69bffbec2d1963b8cbfc70cff39a.jpg b/data/2025/2504_05xxx/2504.05747/images/db3a7dde14ce0b341033213c154d6a9aa45d69bffbec2d1963b8cbfc70cff39a.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..749082af956588573e3836ff1afe5b2275c974b0
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/db3a7dde14ce0b341033213c154d6a9aa45d69bffbec2d1963b8cbfc70cff39a.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:397cf7cf957640763d35fd5bb49dfc5f77bf20a5ba6912a02a943e30be7e560c
+size 173130
diff --git a/data/2025/2504_05xxx/2504.05747/images/dd8f45e05294968ff296d594ea86737670808790006a8a27a017ff220a05095b.jpg b/data/2025/2504_05xxx/2504.05747/images/dd8f45e05294968ff296d594ea86737670808790006a8a27a017ff220a05095b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..818963349ff32b631ec6da0e0844f335dc8706ab
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/dd8f45e05294968ff296d594ea86737670808790006a8a27a017ff220a05095b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3aec199588df527d75fbc098e18879f13ed58ea7a4a0e62f3f5eedcca59ba92
+size 23003
diff --git a/data/2025/2504_05xxx/2504.05747/images/e406902fdae58a202c5ead85c91057043dad6359c378e31f2bf590aaf4e7548a.jpg b/data/2025/2504_05xxx/2504.05747/images/e406902fdae58a202c5ead85c91057043dad6359c378e31f2bf590aaf4e7548a.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..69d3453a6513aa46a2ad47240c37a2fc1bfa4321
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/images/e406902fdae58a202c5ead85c91057043dad6359c378e31f2bf590aaf4e7548a.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ddecc5341cb76e67c8a403e327e5a79ed76aa24ae2b6a3c9272cf18828ae21e
+size 99875
diff --git a/data/2025/2504_05xxx/2504.05747/layout.json b/data/2025/2504_05xxx/2504.05747/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..68f4b79af15ed0b6f1a6ff0fda10bb7b40789c0d
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05747/layout.json
@@ -0,0 +1,7833 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 98,
+ 69,
+ 137,
+ 107
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 98,
+ 69,
+ 137,
+ 107
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 98,
+ 69,
+ 137,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 98,
+ 69,
+ 137,
+ 107
+ ],
+ "type": "image",
+ "image_path": "9f6b9f3761ad7213c0ce54d33552c028e28aae4e28023f2ef7dc3583c49097e7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 142,
+ 80,
+ 493,
+ 98
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 80,
+ 493,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 80,
+ 493,
+ 98
+ ],
+ "type": "text",
+ "content": "SEA-LION: Southeast Asian Languages in One Network"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 77,
+ 107,
+ 518,
+ 247
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 107,
+ 518,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 107,
+ 518,
+ 247
+ ],
+ "type": "text",
+ "content": "Raymond Ng\\*, Thanh Ngan Nguyen\\*, Yuli Huang\\*, Ngee Chia Tai\\*, Wai Yi Leong\\*, Wei Qi Leong\\*, Xianbin Yong\\*, Jian Gang Ngui\\*, Yosephine Susanto\\*, Nicholas Cheng\\*, Hamsawardhini Rengarajan\\*, Peerat Limkonchotiwat\\*, Adithya Venkatadri Hulagadri\\*, Kok Wai Teng\\*, Yeo Yeow Tong\\*, Bryan Siow\\*, Wei Yi Teo\\*, Wayne Lau\\*, Choon Meng Tan\\*, Brandon Ong\\*, Zhi Hao Ong\\*, Jann Railey Montalan\\*, Adwin Chan\\*, Sajeban Antonyrex\\*, Ren Lee\\*, Esther Choa\\*, David Ong Tat-Wee\\*, Bing Jie Darius Liu\\*, William Chandra Tjhi\\*, Erik Cambria\\*, Leslie Teo\\* AI Singapore, National University of Singapore \\*Nanyang Technological University https://sea-lion.ai"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 155,
+ 260,
+ 202,
+ 272
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 155,
+ 260,
+ 202,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 155,
+ 260,
+ 202,
+ 272
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 84,
+ 280,
+ 274,
+ 580
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 84,
+ 280,
+ 274,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 84,
+ 280,
+ 274,
+ 580
+ ],
+ "type": "text",
+ "content": "Recently, Large Language Models (LLMs) have dominated much of the artificial intelligence scene with their ability to process and generate natural languages. However, the majority of LLM research and development remains English-centric, leaving low-resource languages such as those in the Southeast Asian (SEA) region underrepresented. To address this representation gap, we introduce Llama-SEA-LION-8B-IT and Gemma-SEA-LION-9B-IT, two cutting-edge multilingual LLMs designed for SEA languages. The SEA-LION family of LLMs supports 11 SEA languages, namely English, Chinese, Indonesian, Vietnamese, Malay, Thai, Burmese, Lao, Filipino, Tamil, and Khmer. Our work leverages large-scale multilingual continued pre-training with a comprehensive post-training regime involving multiple stages of instruction fine-tuning, alignment, and model merging. Evaluation results on multilingual benchmarks show that our models achieve state-of-the-art performance across LLMs supporting SEA languages. We open-source the models to benefit the wider SEA community."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 68,
+ 587,
+ 154,
+ 599
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 68,
+ 587,
+ 154,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 68,
+ 587,
+ 154,
+ 599
+ ],
+ "type": "text",
+ "content": "1 Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 67,
+ 608,
+ 291,
+ 757
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 608,
+ 291,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 608,
+ 291,
+ 757
+ ],
+ "type": "text",
+ "content": "Large language models (LLMs) have significantly transformed the field of natural language processing, achieving remarkable performance in text generation, summarization and sentiment analysis (Brown et al., 2020; OpenAI, 2023; Dubey et al., 2024; Rivière et al., 2024; Zhang et al., 2024b; Yeo et al., 2024). Despite their impressive capabilities, most LLMs remain heavily English-centric (Wendler et al., 2024; Zhong et al., 2024). Unfortunately, this situation has led LLMs in regions with many under-represented languages such"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 302,
+ 260,
+ 526,
+ 354
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 260,
+ 526,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 260,
+ 526,
+ 354
+ ],
+ "type": "text",
+ "content": "as Southeast Asia (SEA) to suffer. Languages with lower resources, such as Filipino, Lao, Burmese and Khmer in the SEA region, are not supported by many open-source English-centric LLMs. This underscores the need to bridge the resource and representation gap between English and SEA languages."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 302,
+ 358,
+ 526,
+ 641
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 358,
+ 526,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 358,
+ 526,
+ 641
+ ],
+ "type": "text",
+ "content": "Recently, there have been many attempts to create multilingual LLMs in an open-source manner, e.g., BLOOM (Scao et al., 2022), a project aimed at increasing multilingual presence in opensource LLMs by supporting 46 languages. Popular LLM families such as Llama (Dubey et al., 2024), Gemma (Rivière et al., 2024) and Qwen (Yang et al., 2024a) have also introduced multilingual LLMs for their latest iteration. During our evaluations, we found that the performance of these models is acceptable in the general case, i.e., when considering evaluation benchmarks formulated from English datasets. However, we observe that the performance degrades on SEA-specific benchmarks. Moreover, researchers have also introduced LLMs such as SeaLLMs (Nguyen et al., 2024; Zhang et al., 2024a) and Sailor (Dou et al., 2024) to specifically address the LLM gap in SEA languages. However, the performance of these models is less than ideal for languages such as Thai or Tamil"
+ },
+ {
+ "bbox": [
+ 302,
+ 358,
+ 526,
+ 641
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 302,
+ 358,
+ 526,
+ 641
+ ],
+ "type": "text",
+ "content": " (10X et al., 2024; AI Products Team, 2024)."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 302,
+ 645,
+ 526,
+ 739
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 645,
+ 526,
+ 739
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 645,
+ 526,
+ 739
+ ],
+ "type": "text",
+ "content": "In this paper, we address the issues by proposing a robust open-source Southeast Asian model with data transparency for reproducibility, namely SEA-LION - a family of LLMs continued pretrained (CPT) and fine-tuned on Llama-3.1-8B-Instruct for Llama-SEA-LION-8B-IT and Gemma2-9B for Gemma-SEA-LION-9B-IT with a focus"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 81,
+ 762,
+ 197,
+ 773
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 81,
+ 762,
+ 197,
+ 773
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 81,
+ 762,
+ 197,
+ 773
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 81,
+ 762,
+ 197,
+ 773
+ ],
+ "type": "text",
+ "content": "SEA-LION Models Collection"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 302,
+ 752,
+ 525,
+ 775
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 752,
+ 525,
+ 775
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 752,
+ 525,
+ 775
+ ],
+ "type": "text",
+ "content": "2Tamil is one of the official languages in Singapore. It is also spoken in other areas in the SEA region, such as Malaysia."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 13,
+ 244,
+ 36,
+ 593
+ ],
+ "type": "aside_text",
+ "angle": 270,
+ "lines": [
+ {
+ "bbox": [
+ 13,
+ 244,
+ 36,
+ 593
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 13,
+ 244,
+ 36,
+ 593
+ ],
+ "type": "text",
+ "content": "arXiv:2504.05747v4 [cs.CL] 30 Oct 2025"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 273
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 67,
+ 71,
+ 291,
+ 273
+ ],
+ "type": "text",
+ "content": "on SEA languages. To tackle the performance problem, we utilize 200 billion English, code, and SEA languages tokens as well as 16.8 million English and SEA languages instruction and answer pairs for CPT and post-training steps, respectively, to achieve a significant improvement in SEA languages. In order to allow our models to be used by everyone without restrictions, we release our models under the fully open MIT license. We benchmark our models against the SEA-HELM(Susanto et al., 2025) and Open LLM Leaderboard| Source (HuggingFace Datasets) | Languages | Size |
| BAAI/Infinity-Instruct^* | EN | 1,456,927 |
| HuggingFaceTB/smoltalk | EN | 409,537 |
| allenai/tulu-3-sft-personas-math | EN | 149,960 |
| parinzee/seed-free-synthetic-instruct-thai-v1 | TH | 118,898 |
| HuggingFaceTB/smoltalk | EN | 96,356 |
| HuggingFaceTB/smoltalk | EN | 83,144 |
| arcee-ai/EvolKit-75K | EN | 74,174 |
| AI-MO/NuminaMath-TIR | EN | 72,441 |
| Post-training-Data-Flywheel/AutoIF-instruct-61k | EN | 61,492 |
| argilla/ifeval-like-data | EN | 56,339 |
| HuggingFaceTB/smoltalk | EN | 53,342 |
| ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k | EN | 50,000 |
| ai2-adapt-dev/tulu_v3.9_synthetic_finalresp_wildguardmixtrain_decontaminated_50k | EN | 50,000 |
| allenai/tulu-3-sft-personas-math-grade | EN | 49,980 |
| allenai/tulu-3-sft-personas-code | EN | 34,999 |
| HuggingFaceTB/smoltalk | EN | 34,424 |
| allenai/tulu-3-sft-personas-instruction-following | EN | 29,980 |
| airesearch/WangchanThaiInstruct | TH | 25,014 |
| allenai/tulu-3-sft-personas-algebra | EN | 20,000 |
| arcee-ai/EvolKit-20k-vi | VI | 15,378 |
| allenai/coconot | EN | 10,983 |
| ai2-adapt-dev/tulu_v3.9_scirff_10k | EN | 10,000 |
| Source (Generated) | Languages | Size |
| qwen_gemma_synthetic_tamil | TA | 480,000 |
| qwen_gemma_synthetic_thai | TH | 480,000 |
| qwen_gemma_synthetic_indonesian | ID | 465,019 |
| qwen_gemma_synthetic_vietnamese | VI | 465,019 |
| gemma_synthetic_indonesian | ID | 458,149 |
| gemma_synthetic_filipino | TL | 455,093 |
| gemma_synthetic_viet | VI | 291,576 |
| gemma_synthetic_tamil | TA | 276,314 |
| gemma_synthetic_thai | TH | 186,339 |
| gemma_synthetic_javanese | JV | 110,000 |
| gemma_synthetic_sudanese | SU | 110,000 |
| llama_gemma_synthetic_thai | TH | 88,920 |
| llama_gemma_synthetic_tamil | TA | 88,920 |
| llama_gemma_synthetic_vietnamese | VI | 88,920 |
| llama_gemma_synthetic_javanese | JV | 88,920 |
| llama_gemma_synthetic_indonesian | ID | 88,920 |
| llama_gemma_synthetic_filipino | TL | 80,000 |
| enrich_27k | SEA | 27,463 |
| seaMultilingual_systemchat | SEA | 1,903 |
| rewritten_oasst | SEA | 841 |
| rewritten_helpsteer | SEA | 838 |
",
+ "image_path": "db3a7dde14ce0b341033213c154d6a9aa45d69bffbec2d1963b8cbfc70cff39a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 214,
+ 718,
+ 379,
+ 729
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 214,
+ 718,
+ 379,
+ 729
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 214,
+ 718,
+ 379,
+ 729
+ ],
+ "type": "text",
+ "content": "Table 9: List of datasets for Stage-2-IFT."
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "text"
+ }
+ ],
+ "discarded_blocks": [],
+ "page_size": [
+ 595,
+ 841
+ ],
+ "page_idx": 14
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/data/2025/2504_05xxx/2504.05979/c3842c4d-63fe-4775-a7ec-8e60138be3cb_content_list.json b/data/2025/2504_05xxx/2504.05979/c3842c4d-63fe-4775-a7ec-8e60138be3cb_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..e4697cea354bb075c0bacf064a6f6086255e40b9
--- /dev/null
+++ b/data/2025/2504_05xxx/2504.05979/c3842c4d-63fe-4775-a7ec-8e60138be3cb_content_list.json
@@ -0,0 +1,22400 @@
+[
+ {
+ "type": "text",
+ "text": "An Empirical Study of GPT-4o Image Generation Capabilities",
+ "text_level": 1,
+ "bbox": [
+ 127,
+ 122,
+ 869,
+ 148
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Sixiang Chen $^{1*}$ , Jinbin Bai $^{2*}$ , Zhuoran Zhao $^{1*}$ , Tian Ye $^{1*}$ , Qingyu Shi $^{3}$ , Donghao Zhou $^{4}$ , Wenhao Chai $^{5}$ , Xin Lin $^{6}$ , Jianzong Wu $^{3}$ , Chao Tang $^{3}$ , Shilin Xu $^{3}$ , Tao Zhang $^{6}$ , Haobo Yuan $^{6}$ , Yikang Zhou $^{6}$ , Wei Chow $^{2}$ , Linfeng Li $^{2}$ , Xiangtai Li $^{3\\dagger}$ , Lei Zhu $^{1,7\\dagger}$ , Lu Qi $^{6\\dagger}$",
+ "bbox": [
+ 137,
+ 198,
+ 870,
+ 246
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "$^{1}$ The Hong Kong University of Science and Technology (GZ) $^{2}$ National University of Singapore $^{3}$ Peking University $^{4}$ The Chinese University of Hong Kong $^{5}$ University of Washington $^{6}$ Wuhan University $^{7}$ The Hong Kong University of Science and Technology",
+ "bbox": [
+ 153,
+ 246,
+ 854,
+ 290
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 459,
+ 325,
+ 537,
+ 340
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "The landscape of image generation has rapidly evolved, from early GAN-based approaches to diffusion models and, most recently, to unified generative architectures that seek to bridge understanding and generation tasks. Recent advances, especially the GPT-4o, have demonstrated the feasibility of high-fidelity multimodal generation, their architectural design remains mysterious and unpublished. This prompts the question of whether image and text generation have already been successfully integrated into a unified framework for those methods. In this work, we conduct an empirical study of GPT-4o's image generation capabilities, benchmarking it against leading open-source and commercial models. Our evaluation covers four main categories, including text-to-image, image-to-image, image-to-3D, and image-to-X generation, with more than 20 tasks. Our analysis highlights the strengths and limitations of GPT-4o under various settings, and situates it within the broader evolution of generative modeling. Through this investigation, we identify promising directions for future unified generative models, emphasizing the role of architectural design and data scaling. For a high-definition version of the PDF, please refer to the link on GitHub: https://github.com/Ephemeral182/Empirical-Study-of-GPT-4o-Image-Gen.",
+ "bbox": [
+ 184,
+ 357,
+ 815,
+ 551
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 Introduction",
+ "text_level": 1,
+ "bbox": [
+ 127,
+ 578,
+ 267,
+ 594
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Over the past decade, image generation has undergone a remarkable evolution—from the early successes of GANs [35] to the dominance of diffusion models [89, 82, 26], which have significantly advanced image fidelity and diversity [37, 7]. In parallel, Large Language Models (LLMs) have achieved exceptional performance across diverse natural language tasks by scaling autoregressive next-token prediction, demonstrating the power of unified modeling principles. These advances naturally raise a compelling question: can such principles be extended to image generation?",
+ "bbox": [
+ 124,
+ 609,
+ 870,
+ 696
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "However, fundamental differences between autoregressive and diffusion-based paradigms present non-trivial challenges. Autoregressive models excel in sequential text generation, while diffusion models have become the de facto standard for high-quality image synthesis. Bridging these modalities within a unified framework remains an open challenge. Several works [96, 101, 100, 34, 24, 13] attempt to bridge this gap via multimodal connectors or instruction tuning, with LLMs serving as planning modules that produce intermediate representations for image generation. While effective to some extent, these paradigms often exhibit limited interaction between text and image modalities, and struggle with content consistency—particularly in image-to-image generation and complex instruction-based synthesis.",
+ "bbox": [
+ 124,
+ 699,
+ 869,
+ 811
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "To address these limitations, recent research explores unified generation models that integrate understanding and generation within a single architecture, following three main technical paradigms. The first line of work represents both language and vision as discrete token sequences [67, 98, 110, 104, 19, 65, 109], leveraging VQGAN [28] or similar compressors to tokenize images for compatibility with autoregressive models. A second direction integrates",
+ "bbox": [
+ 125,
+ 816,
+ 870,
+ 875
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "aside_text",
+ "text": "arXiv:2504.05979v2 [cs.CV] 10 Apr 2025",
+ "bbox": [
+ 22,
+ 262,
+ 60,
+ 705
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_footnote",
+ "text": "*Equal contributions. ☑: schen691@connect.hkust-gz.edu.cn † Corresponding authors.",
+ "bbox": [
+ 148,
+ 883,
+ 705,
+ 898
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "footer",
+ "text": "Preprint. Work in progress.",
+ "bbox": [
+ 127,
+ 922,
+ 290,
+ 936
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "large language models directly into the diffusion process [128, 126, 112, 72], employing them as denoising backbones for image generation and as unified sequence models for text. While promising, these approaches typically rely on intermediate compression modules such as VAEs or VQVAEs, which may limit visual fidelity or increase architectural complexity. A third and increasingly prominent paradigm investigates discrete diffusion frameworks that natively support both image and text generation within a unified modeling space [71, 73, 93]. Building on this insight, recent works [58, 97] propose fully end-to-end diffusion architectures based on shared Transformer backbones, demonstrating competitive performance and seamless modality integration comparable to similarly sized LLMs.",
+ "bbox": [
+ 125,
+ 90,
+ 870,
+ 203
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Despite these promising directions, such systems still lag behind the sophistication and generalization capabilities of proprietary models like Flux [51] and Midjourney [75], which may lack reasoning capabilities.",
+ "bbox": [
+ 125,
+ 208,
+ 870,
+ 238
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The recent release of GPT-4o [78] marks a significant milestone in multimodal generative modeling. As a native multimodal architecture, GPT-4o demonstrates strong capabilities in generating high-fidelity, photorealistic images while seamlessly unifying vision and language generation—reportedly in an autoregressive fashion. However, its closed-source nature—particularly the lack of disclosure about its architecture, training regimen, and inference mechanisms—poses substantial challenges for scientific scrutiny. This motivates a careful empirical assessment of its capabilities relative to open-source state-of-the-art models.",
+ "bbox": [
+ 125,
+ 242,
+ 870,
+ 325
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Although the visual performance of GPT-4o and Gemini is widely recognized, much of their success likely stems from unprecedented scale in training data, model parameters, and compute resources. Prior studies, including diffusion models and connected-based models, suggest that scaling is a key enabler of generative quality—potentially more so than architectural novelty alone. These trends point to a promising trajectory for unified generative models: with sufficient scale, they may rival or even surpass today's best proprietary systems.",
+ "bbox": [
+ 125,
+ 332,
+ 870,
+ 402
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In this study, we conduct a comprehensive evaluation of GPT-4o's image generation performance, benchmarking its outputs against leading systems including Gemini 2.0 Flash Experimental [99] and other state-of-the-art models. Building upon our comparative evaluation across text-to-image, image-to-image, image-to-3D, and image-to-X generation tasks, GPT-4o demonstrates several distinctive strengths:",
+ "bbox": [
+ 125,
+ 407,
+ 870,
+ 465
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- Exceptional Text Rendering Capability. GPT-4o demonstrates exceptional capability in rendering textual elements within images, maintaining correct spelling, alignment, and formatting even in document-style generation tasks. This level of text fluency is rarely seen in prior models and is crucial for practical applications such as chart generation, document layout synthesis, and instruction-rich visual storytelling.",
+ "- Compositional Generalization and Prompt Following. GPT-4o displays impressive compositional abilities, accurately assembling complex scene elements, styles, or attributes described in prompts. This high prompt following enables it to handle fine-grained multi-attribute conditions in generation tasks with minimal loss of semantic detail.",
+ "- Spatial Reasoning and Multi-View Consistency. In generation tasks involving spatial manipulation, such as 3D view synthesis, camera control, and depth-conditioned rendering, GPT-4o maintains geometric consistency and viewpoint realism. This indicates an inherent capacity for spatial reasoning and structural awareness, even without explicit 3D modeling modules.",
+ "- Comprehensive Image Transformation Capability. GPT-4o shows strong generalization across a wide spectrum of image-to-image tasks, ranging from low-level image restoration to high-level perceptual understanding. Without task-specific tuning, it almost handles diverse transformations such as denoising, deblurring, relighting, segmentation, and depth estimation. This suggests the model has learned robust visual priors and spatial semantics, enabling it to perform correction and abstract structural prediction under a unified framework."
+ ],
+ "bbox": [
+ 171,
+ 476,
+ 869,
+ 744
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "However, limitations remain in inconsistent generation, hallucination, and data bias in underrepresented cultural elements and non-Latin scripts, highlighting current trade-offs in model design and training data coverage.",
+ "bbox": [
+ 125,
+ 756,
+ 867,
+ 786
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "While we do not analyze the internal architecture or implementation details of GPT-4o in this paper*, we believe it plays an important role toward unified multimodal generation. We also emphasize that model architecture is only one part of this progress—training data, model scale, and optimization strategies are equally important. We hope future work will provide more empirical evidence to better understand such proprietary systems and their position within this evolving research landscape.",
+ "bbox": [
+ 125,
+ 790,
+ 870,
+ 861
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_footnote",
+ "text": "*There is currently no definitive evidence regarding the specific implementation details or architectural design of GPT-4o's image generation capabilities. To ensure the credibility and accuracy of our analysis, we will refrain from making speculative claims in current version.",
+ "bbox": [
+ 125,
+ 872,
+ 870,
+ 912
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "2",
+ "bbox": [
+ 493,
+ 935,
+ 504,
+ 946
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2 Evaluation",
+ "text_level": 1,
+ "bbox": [
+ 127,
+ 89,
+ 253,
+ 104
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "As GPT-4o's image generation capability has only recently been released and no API is available, we conduct only qualitative comparisons between GPT-4o, Gemini 2.0 Flash [99], and other state-of-the-art models in their respective domains.",
+ "bbox": [
+ 125,
+ 119,
+ 870,
+ 162
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "To systematically compare these models' performance across diverse image generation tasks including text-to-image generation, image-to-image generation, text/image to 3D generation, and various image-to-X generation, we conduct a detailed case study focused on analyzing the performance of these models. This qualitative analysis provides insight into gpt 4o's strengths and limitations in various tasks, as shown in Table 1.",
+ "bbox": [
+ 125,
+ 169,
+ 870,
+ 226
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Low Visual Quality : The image synthesis model fails to generate fine-grained object details or produces blurry outputs. Typical cases include distorted human bodies or unrealistic hand shapes.",
+ "bbox": [
+ 125,
+ 233,
+ 869,
+ 265
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Inconsistent Generation : The image synthesis model produces inconsistent output or image details with input image.",
+ "bbox": [
+ 125,
+ 272,
+ 867,
+ 301
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Lack of Knowledge : The image synthesis model lacks domain-specific knowledge, such as particular artistic styles, and thus generates visually plausible but incorrect results.",
+ "bbox": [
+ 125,
+ 311,
+ 867,
+ 342
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Failure to Follow Instructions : The image synthesis model misinterprets the input prompt and produces inconsistent results. For example, it may fail to capture specified numbers, colors, or object arrangements.",
+ "bbox": [
+ 125,
+ 349,
+ 867,
+ 381
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "3",
+ "bbox": [
+ 493,
+ 935,
+ 503,
+ 946
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "table",
+ "img_path": "images/69fc7b667b5296c731fc241b01cedfd07a9c8e3fb9b10cdf2db89fc2a34aef2f.jpg",
+ "table_caption": [
+ "Table 1: GPT-4o vs. Baselines: Qualitative error analysis across image generation tasks."
+ ],
+ "table_footnote": [],
+ "table_body": "| Case Figure | Meta-task | Sub-task | GPT-4o | Gemini-2.0-flash | Domain-SOTA |
| Figure 1 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 2 | | Complex Text Following | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 3 | | | Success | Success | Success |
| Figure 4 | | | Success | Success | Success |
| Figure 5 | | | Success | Success | Success |
| Figure 6 | Text-to-Image | Text Rendering | Success | Low Visual Quality | Low Visual Quality |
| Figure 7 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 8 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 9 | | Document Generation | Success | Low Visual Quality | Low Visual Quality |
| Figure 10 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 11 | | Panorama | Lack of Knowledge | Success | Success |
| Figure 12 | | Style Transfer | Success | Lack of Knowledge | Lack of Knowledge |
| Figure 13 | | | Success | Lack of Knowledge | Lack of Knowledge |
| Figure 14 | | | Low Visual Quality | Success | Failure to Follow Instructions |
| Figure 15 | | Image Editing | Failure to Follow Instructions | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 16 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 17 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 18 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 19 | | | Success | Inconsistent Generation | Failure to Follow Instructions |
| Figure 20 | | Single-Concept Customization | Success | Failure to Follow Instructions | Success |
| Figure 21 | | Multi-Concept Customization | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 22 | | Story Image Generation | Success | Failure to Follow Instructions | Success |
| Figure 23 | | | Success | Inconsistent Generation | Success |
| Figure 24 | | Low-Level Vision-Denoising | Low Visual Quality | Low Visual Quality | Success |
| Figure 25 | | Low-Level Vision-Deraining | Success | Inconsistent Generation | Success |
| Figure 26 | | Low-Level Vision-Dehazing | Success | Low Visual Quality | Success |
| Figure 27 | | Low-Level Vision-Low Light Enhancement | Low Visual Quality | Low Visual Quality | Success |
| Figure 28 | | Low-Level Vision-Deblurring | Success | Low Visual Quality | Success |
| Figure 29 | | Low-Level Vision-Super Resolution | Success | Low Visual Quality | Success |
| Figure 30 | | Low-Level Vision-Impainting | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 31 | | Low-Level Vision-Outpainting | Inconsistent Generation | Success | Success |
| Figure 32 | | Low-Level Vision-Colorization | Success | Success | Success |
| Figure 33 | | Low-Level Vision-Shadow Removal | Success | Failure to Follow Instructions | Success |
| Figure 34 | | Low-Level Vision-Reflection Removal | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 35 | | Low-Level Vision-Relighting | Success | Failure to Follow Instructions | Success |
| Figure 36 | | Spatial Control-Canny | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 37 | | Spatial Control-Depth | Success | Failure to Follow Instructions | Success |
| Figure 38 | | Spatial Control-Sketch | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 39 | | Spatial Control-Pose | Success | Inconsistent Generation | Success |
| Figure 40 | | Spatial Control-Mask | Inconsistent Generation | Failure to Follow Instructions | Inconsistent Generation |
| Figure 41 | | Camera Control | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 42 | | | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 43 | | In-Context Visual Prompting | Failure to Follow Instructions | Failure to Follow Instructions | N/A |
| Figure 44 | | Image to 3D Modeling | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 45 | | UV Map to 3D Rendering | Success | Inconsistent Generation | Failure to Follow Instructions |
| Figure 46 | | Novel View Synthesis | Success | Success | Failure to Follow Instructions |
| Figure 47 | | Image Segmentation | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 48 | | | Success | Failure to Follow Instructions | Success |
| Figure 49 | | | Success | Failure to Follow Instructions | Success |
| Figure 50 | | Edge Detection | Success | Success | Success |
| Figure 51 | | | Success | Failure to Follow Instructions | Success |
| Figure 52 | | | Success | Failure to Follow Instructions | Success |
| Figure 53 | | Salient Object | Success | Failure to Follow Instructions | Success |
| Figure 54 | | | Success | Success | Success |
| Figure 55 | | | Success | Success | Success |
| Figure 56 | | Depth Estimation | Success | Failure to Follow Instructions | Success |
| Figure 57 | | Normal Estimation | Success | Failure to Follow Instructions | Success |
| Figure 58 | | Layout Detection | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 59 | | Text Detection | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 60 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 61 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 62 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 63 | | | Inconsistent Generation | Inconsistent Generation | Success |
",
+ "bbox": [
+ 129,
+ 138,
+ 870,
+ 883
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "4",
+ "bbox": [
+ 493,
+ 935,
+ 504,
+ 946
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.1 Text-to-Image Tasks",
+ "text_level": 1,
+ "bbox": [
+ 127,
+ 90,
+ 310,
+ 104
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "2.1.1 Complex Text Following Capability",
+ "text_level": 1,
+ "bbox": [
+ 127,
+ 116,
+ 428,
+ 132
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Recent progress in text-to-image generation has shown impressive abilities in generating diverse and realistic images based on text prompts. However, composing multiple objects with various attributes and relationships accurately into one scene remains a significant challenge for current text-to-image generative models [92, 85, 8, 81, 6]. In this section, we assess models' ability for compositional text-to-image generation from four perspectives following [41], which include attribute binding, numeracy, object relationship, and complex compositions. Attribute binding evaluates whether the model correctly assigns attributes, such as color, shape, and texture to the appropriate objects. Numeracy evaluates whether the number of generated objects matches the quantities specified in the prompt. Object relationships refer to both spatial (2D/3D) and non-spatial interactions among objects. Complex compositions evaluate the model's ability to handle multiple types of constraints simultaneously, especially given long or detailed prompts.",
+ "bbox": [
+ 124,
+ 140,
+ 870,
+ 281
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "As shown in Figure 1 row 1, GPT-4o outperforms both Gemini 2.0 Flash and Midjourney in numeracy tasks. While GPT-4o accurately represents a single plate, Gemini 2.0 and Midjourney represent two plates instead. In terms of understanding object relationships, GPT-4o is the only model that correctly infers the action \"walk towards\" from the ragdoll to the labrador. However, GPT-4o struggles with more complex terms like \"pentagonal pyramid\", failing to interpret it correctly (see Figure 1 row 4). This suggests that GPT-4o may have difficulty accurately interpreting objects with unusual geometries. When it comes to abstract prompts, GPT-4o also appears to lack imagination (see Figure 2 row 2), whereas Midjourney v6.1 demonstrates better creativity in this case, outperforming both GPT-4o and Gemini 2.0 Flash.",
+ "bbox": [
+ 124,
+ 284,
+ 870,
+ 397
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "For complex text-to-image generation, we evaluate GPT-4o's performance with Gemini 2.0 Flash [99] and FLUX.1-Pro [51], using the text prompts collected from [124, 106, 115]. As shown in Figure 3, both GPT-4o and FLUX excel at generating realistic and harmonious scenes align with the text prompts. However, we observe that GPT-4o shows limitations in generating culturally related elements. For example, the generated crown for the Chinese general is western-style rather than chinese-style (see Figure 4 row 2). Additionally, in large scene generation, GPT-4o struggles to maintain boundary continuity, whereas FLUX produces a more natural composition (see Figure 4 row 3).",
+ "bbox": [
+ 124,
+ 401,
+ 870,
+ 501
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Overall, we conclude that GPT-4o excels at text-to-image generation in terms of attribute binding, generative numeracy, object relationship, and complex compositions. However, it exhibits limitations in generating uncommon objects, culturally specific elements and in maintaining continuity when composing large scenes.",
+ "bbox": [
+ 124,
+ 505,
+ 870,
+ 550
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "5",
+ "bbox": [
+ 493,
+ 935,
+ 504,
+ 946
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Text-to-Image Generation",
+ "text_level": 1,
+ "bbox": [
+ 191,
+ 99,
+ 392,
+ 116
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/5e416a9762e2def779608eb1bf6eee5a9cee49745a628b45ccc17073ed461705.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 251,
+ 126,
+ 269,
+ 143
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Visual content precisely following the text instruction.",
+ "text_level": 1,
+ "bbox": [
+ 271,
+ 128,
+ 772,
+ 145
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/b56cffa92dcc8965f109102c1a0975dedfc665539ea8cf6c71d5a2879dfe8dcc.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 171,
+ 146,
+ 362,
+ 292
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/e6d59fa9484b3877034764c9fa8752e51c78a7b4eba2a3dd224cdb9236f23368.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 393,
+ 146,
+ 586,
+ 295
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/114bf170640575674f20a5f89dd3e1c35830102990e7d45e9fe0e42bf870b7d0.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 614,
+ 146,
+ 823,
+ 295
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/e21293e1dde8f2f8a8736c0a4d673afa08874092081ea4b1ec07788bbe177b05.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 171,
+ 313,
+ 366,
+ 460
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/cde20841ace4621a197135b25fda935e422854fc542a786ca15f9aea2eb46486.jpg",
+ "image_caption": [
+ "Input Text: \"A yellow bowl, a blue mug and a pink plate on the table.\"",
+ "Input Text: \"A ragdoll walks towards a labrador.\""
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 393,
+ 311,
+ 589,
+ 464
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/d0a18282d6404ef0ab645148e73d17fe445982eb8c40cb3c0d5a9081efd6d52c.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 614,
+ 310,
+ 821,
+ 464
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/3bffa6f0cb5545336d1dece3985a3fcaff2ad60da5c231db47b90005c63f4ceb.jpg",
+ "image_caption": [
+ "Input Text: \"Three differently colored apples (yellow, green, red from left to right) with a Coca-Cola bottle placed behind the middle apple.\""
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 171,
+ 479,
+ 370,
+ 627
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/f333d581ad570f753e99a7e73c99d827c89d746ae6f1f549c53cf30cf7e7ea33.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 395,
+ 479,
+ 589,
+ 627
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/107a161e8c64a21e6ae887d967ee6aa6bf1c940aafacc290e429d8deb9516754.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 614,
+ 478,
+ 821,
+ 627
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/00b435bd1d8026aa4f0d08be8a9579479f625012cae88f49f81a5d958cf5c199.jpg",
+ "image_caption": [
+ "Input Text: \"The oval sphere was nestled between the rectangular prism and the pentagonal pyramid.\"",
+ "Figure 1: Task: Compositional text-to-image generation. Evaluate the image-text alignment on attribute binding, numeracy, and object relationship. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Midjourney v6.1 [75]. Observation: GPT-4o outperforms Gemini 2.0 Flash and Midjourney v6.1 across all aspects. However, GPT-4o struggles with uncommon objects with a special geometry."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 171,
+ 657,
+ 369,
+ 806
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/b1e74b98b6b45c314db43ad44efec736de0526ea53933c41d6f494dfc222ec5c.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 393,
+ 657,
+ 589,
+ 806
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/ec488303c09dbf9dfb728fd1952368c1c3fa85f708e3b290ecd1a36b50ef4803.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 614,
+ 657,
+ 821,
+ 806
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "GPT 40",
+ "bbox": [
+ 235,
+ 829,
+ 287,
+ 842
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Gemini 2.0 Flash",
+ "bbox": [
+ 436,
+ 830,
+ 550,
+ 844
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Midjourney v6.1",
+ "bbox": [
+ 665,
+ 829,
+ 774,
+ 844
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "6",
+ "bbox": [
+ 493,
+ 936,
+ 504,
+ 946
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/41cf37136782a8224523753684cab11c0ddb36c12dfd39ee89ac851285853af2.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 253,
+ 127,
+ 269,
+ 143
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Visual content precisely following the text instruction.",
+ "text_level": 1,
+ "bbox": [
+ 272,
+ 130,
+ 772,
+ 145
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/742bdab0de5d04587ee79a5c9d93f90cfb4494e5a48dcf258a1406c6265bd627.jpg",
+ "image_caption": [
+ "Input Text: \"The round, juicy watermelon sat in the cool, refreshing bowl of ice, waiting to be sliced open and devoured.\""
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 176,
+ 146,
+ 372,
+ 292
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/c73e3aa340f33cab4cf19354249e37fe1bd7f76aef6ec7fbcc154198f1dad05b.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 393,
+ 146,
+ 611,
+ 292
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/382030f3e3f0298bca0a14e5cda55ee5859de8f05e70a6182873cc5b53987e77.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 633,
+ 146,
+ 823,
+ 292
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/9a87f98dd86139de0d38246d54b1d070591e2fdb526d3285409a473583f28ec6.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 176,
+ 319,
+ 372,
+ 469
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/fce6b55e09781aa60141751625be4a55c7d0fd1eef584be53a6585c781eade33.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 397,
+ 319,
+ 609,
+ 468
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/0f7ad2208e211f067756484bb7a6847038cf13549a2f34c8a13b12eeda49aad7.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 632,
+ 319,
+ 826,
+ 469
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/d87f5d38548a498be4acc2e6175a859f019a900cdcbd8394bd2700b4f145ed8f.jpg",
+ "image_caption": [
+ "Input Text: \"The bold, expressive strokes of the artist's brush brought the blank canvas to life, forming a vibrant and dynamic masterpiece.\""
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 176,
+ 496,
+ 372,
+ 643
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/c87b7b6792c5fda9d5340959a376273a7f5d0b80db6770aea6ec9e15c6ec596d.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 397,
+ 496,
+ 609,
+ 643
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/4d7cf29e3244126ed31062342f32c73674817be1d0079f2da561d4e75f3368e2.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 632,
+ 496,
+ 828,
+ 645
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/c36a7bcd11820fd2f7d3c63fbf0e563dbdebdbcf2828507130fec8b86c2c1cc1.jpg",
+ "image_caption": [
+ "Input Text: \"The heavy raindrops fell on the smooth glass and the textured roof.\""
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 176,
+ 662,
+ 372,
+ 811
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/0e3a0dc448dd3f289f1837b24241f0ae490d1d79799115f52e41475efc038437.jpg",
+ "image_caption": [
+ "Input Text: \"The gentle, soothing melody of the piano filled the concert hall, as the pianist's fingers danced over the keys.\"",
+ "Figure 2: Task: Compositional text-to-image generation. Evaluate the image-text alignment on attribute binding and complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Midjourney v6.1 [75]. Observation: GPT-4o outperforms the other two models in generating objects aligned with the text prompts accurately. But for more abstract and creative tasks, Midjourney v6.1 performs the best."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 397,
+ 662,
+ 611,
+ 813
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/adcb856bb8a351cb62cc8d6fd9dcfe0f715a542653534f39e99b1973a7067ec8.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 632,
+ 662,
+ 828,
+ 813
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "GPT 40",
+ "bbox": [
+ 233,
+ 842,
+ 285,
+ 854
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Gemini 2.0 Flash",
+ "bbox": [
+ 434,
+ 842,
+ 547,
+ 854
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Midjourney v6.1",
+ "bbox": [
+ 663,
+ 840,
+ 771,
+ 854
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "header",
+ "text": "Text-to-Image Generation",
+ "bbox": [
+ 194,
+ 102,
+ 392,
+ 118
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "page_number",
+ "text": "7",
+ "bbox": [
+ 493,
+ 935,
+ 504,
+ 946
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Text-to-Image Generation (with complex text prompt)",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 107,
+ 393,
+ 137
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Visual content precisely following the text instruction.",
+ "text_level": 1,
+ "bbox": [
+ 223,
+ 148,
+ 787,
+ 165
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/27fce9400300fb184ed43e1e4b4f8eb042ae16f63646b6eda324b41fff54aede.jpg",
+ "image_caption": [
+ "Input Text: \"An icy landscape. A vast expanse of snow-covered mountain peaks stretches endlessly. Beneath them is a dense forest and a colossal frozen lake. Three people are boating in three boats separately in the lake. Not far from the lake, a volcano threatens eruption, its rumblings felt even from afar. Above, a ferocious red dragon dominates the sky and commands the heavens, fueled by the volcano's relentless energy flow.\" (Prompt from GenArtist)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 173,
+ 167,
+ 359,
+ 311
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/116d28e49cd90395347ff38e9ec8e1c04b768757b89b7a9a1d0a0f0f317b20f4.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 390,
+ 167,
+ 578,
+ 311
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/bfaf1c18b294719962fa4fcaa9024d109045581572e81ebad1ebe756236b696a.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 609,
+ 167,
+ 797,
+ 311
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/64e6d603fb4bb84b5ebce439ab2fdb3c6e0e566e8af8982c1c50d5378b6fe487.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 173,
+ 390,
+ 359,
+ 532
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/7ed09698c0927be9038e1132fc7243c7e628f5e5c041880670c7d1ac190273ad.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 390,
+ 388,
+ 578,
+ 532
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/484e31e49c1e7a2ebd3c367ae0db8fcd10716584cec8c8e06beaca0f6c7a0381.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 609,
+ 388,
+ 795,
+ 532
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/2de6dd439640e1a491cd2669646ca4deab67272994343185f41956a82e6a40a1.jpg",
+ "image_caption": [
+ "Input Text: \"In a magical seascape, a majestic ship sails through crystal blue waters surrounded by vibrant marine life and soaring birds. Towering cliffs frame the scene, while a stunning rainbow arches across the sky, blending with ethereal clouds. This enchanting journey captures the serene beauty of nature's wonders.\" (Prompt from IterComp)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 173,
+ 599,
+ 359,
+ 742
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/0c0470a214490980a8b144c3e88a00976a14f32edb1762c91f3846a0c05ec5b2.jpg",
+ "image_caption": [
+ "Input Text: \"On the rooftop of a skyscraper in a bustling cyberpunk city, a figure in a trench coat and neon-lit visor stands amidst a garden of bio-luminescent plants, overlooking the maze of flying cars and towering holograms. Robotic birds flit among the foliage, digital billboards flash advertisements in the distance.\" (Prompt from IterComp)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 390,
+ 598,
+ 578,
+ 743
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/745ba50ac275ead363177aa6a6389523a01b4c236062fe89cb966d79aeafe69b.jpg",
+ "image_caption": [
+ "Figure 3: Task: Compositional text-to-image generation. Evaluate the image-text alignment on complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and FLUX.1-Pro [51]. Observation: GPT-4o and FLUX can generate more harmonious and natural scene than Gemini 2.0 Flash."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 609,
+ 598,
+ 795,
+ 742
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "GPT 40",
+ "bbox": [
+ 233,
+ 806,
+ 290,
+ 820
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Gemini 2.0 Flash",
+ "bbox": [
+ 418,
+ 806,
+ 544,
+ 820
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "FLUX",
+ "bbox": [
+ 676,
+ 806,
+ 723,
+ 820
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "page_number",
+ "text": "8",
+ "bbox": [
+ 493,
+ 935,
+ 504,
+ 946
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Text-to-Image Generation (with complex text prompt)",
+ "text_level": 1,
+ "bbox": [
+ 168,
+ 119,
+ 387,
+ 150
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/30ed799a330abd36ec3b456792e50674b62123dda7d183a4909f198292186c06.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 197,
+ 157,
+ 217,
+ 175
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Visual content precisely following the text instruction.",
+ "text_level": 1,
+ "bbox": [
+ 222,
+ 162,
+ 787,
+ 178
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/0a19bafce54aa9b28620506b9c6051449ff283c024b7b8bdde3dea6959a5f2ab.jpg",
+ "image_caption": [
+ "Input Text: \"Under the luminous full moon, a serene Japanese garden with traditional pagodas and a tranquil pond creates a magical night scene. The soft glow from the lantern-lit buildings reflects on the water, blending nature and architecture in harmony. The moonlight bathes the landscape, enhancing the peaceful ambiance.\" (Prompt from IterComp)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 171,
+ 181,
+ 357,
+ 325
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/1660f44ab6725c952d5b0bc43505f77c0d7cb4b552be04d6430ee2b25ec63d61.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 393,
+ 183,
+ 578,
+ 325
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/89e41678c8fcc68e299a5cf2931d67bb1f2339142bca0c32c1593b4888650519.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 611,
+ 183,
+ 795,
+ 325
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/f016fb55ce2f2b9f48a4d23f36970310500e1250abf50d5e88f1902507f424c6.jpg",
+ "image_caption": [
+ "Input Text: \"A Chinese general wearing a crown, with whiskers and golden Chinese style armor, standing with a majestic dragon head on his chest, symbolizing his strength, wearing black and gold boots. His appearance exudes a sense of authority, wisdom, and an unyielding spirit, embodying the ideal ancient Chinese hero.\" (Prompt from RPG)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 197,
+ 393,
+ 333,
+ 547
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/9077e20f899c60d5e1c26e1748dfa141980bf09da6c613467d42c1143bab34a1.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 413,
+ 393,
+ 558,
+ 547
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/c341891a5dd74ee440ca81c75da11adba9b78fd29a19903df5c98dbac7513e19.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 627,
+ 393,
+ 777,
+ 547
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/4569221e0b0de61258bfb9ba425fbed4cda37201c120d48edf92be4e3bff1d5c.jpg",
+ "image_caption": [
+ "Input Text: \"A beautiful landscape with a river in the middle, the left of the river is in the evening and in the winter with a big iceberg and a small village while some people are skiing on the river and some people are skating, the right of the river is in the summer with a volcano in the morning and a small village while some people are playing.\" (Prompt from RPG)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 151,
+ 617,
+ 357,
+ 724
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/475d909f8bd94a4e962a1af7f117f1db9f76e1f2cd7047ccbdf158cec0cbc2b2.jpg",
+ "image_caption": [
+ "Figure 4: Task: Compositional text-to-image generation. Evaluate the image-text alignment on complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and FLUX.1-Pro [51]. Observation: GPT-4o struggles to generate culturally related elements and maintain boundary continuity (see rows 2 and 3), similar to Gemini 2.0 Flash and FLUX."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 361,
+ 618,
+ 504,
+ 724
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/ef1d1bd4564e7e63c547c011879fd964bcadd7553b251ad9931c404cc35e496b.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 504,
+ 617,
+ 651,
+ 724
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/1086a817de2fda2fa73bc2221b337c167aa823711f8d87db081972a8c108577d.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 655,
+ 617,
+ 843,
+ 724
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "GPT 40",
+ "bbox": [
+ 220,
+ 795,
+ 277,
+ 809
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "Gemini 2.0 Flash",
+ "bbox": [
+ 431,
+ 795,
+ 557,
+ 809
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "FLUX",
+ "bbox": [
+ 722,
+ 792,
+ 767,
+ 806
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "page_number",
+ "text": "9",
+ "bbox": [
+ 493,
+ 935,
+ 504,
+ 946
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "2.1.2 Text Rendering",
+ "text_level": 1,
+ "bbox": [
+ 127,
+ 90,
+ 292,
+ 107
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "Text rendering is a task that aims at generating texts (characters, sentences, or even paragraphs) on an image. The text content is usually guided by the input prompt. Previous models [27, 2] show good capability in generating short text (within 10 words, such as signs or short phrases), but their ability to generate long texts remains limited.",
+ "bbox": [
+ 125,
+ 114,
+ 870,
+ 157
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "As shown in Figure 5, GPT-4o demonstrates comparable abilities to existing state-of-the-art (SOTA) baselines when generating short texts. All the methods except FLUX [51] perform well at rendering short text following the prompt. In this section, we primarily focus on long text rendering to examine whether GPT-4o can surpass these baselines for extended textual content.",
+ "bbox": [
+ 125,
+ 162,
+ 869,
+ 219
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "We choose POSTA [12], Gemini 2.0 Flash [99], Ideogram 3.0 [2], and Playground-v3 [64] as the baselines because of their established capabilities in rendering longer texts. The results are shown in Figure 6 and Figure 7.",
+ "bbox": [
+ 125,
+ 224,
+ 867,
+ 253
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "From these examples, we make the following key observations:",
+ "bbox": [
+ 125,
+ 260,
+ 545,
+ 273
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- GPT-4o's strength in long text generation: Compared with other baselines, GPT-4o demonstrates a superior ability to generate long, coherent text. In example 1 and example 3, GPT-4o produces detailed textual information with fewer than three characters generated incorrectly across more than 100 characters of text.",
+ "- Baseline limitations: When the input prompt becomes extremely long, models such as Gemini 2.0 Flash, Ideogram 3.0, and Playground-v3 often produce significantly more errors or produce vague text patches that are difficult to recognize.",
+ "- POSTA's performance: As a model specifically designed for poster-style text generation, POSTA performs closely to, or in some instances slightly more precisely than, GPT-4o. We hypothesize this is due to its multi-step pipeline tailored for long text rendering."
+ ],
+ "bbox": [
+ 169,
+ 285,
+ 867,
+ 431
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "Overall, we conclude that GPT-4o excels at long text rendering, offering overwhelming performance compared to most existing commercial models, and delivering results on par with the latest specialized research models.",
+ "bbox": [
+ 125,
+ 445,
+ 869,
+ 474
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "page_number",
+ "text": "10",
+ "bbox": [
+ 490,
+ 935,
+ 509,
+ 946
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "Short Text Rendering",
+ "text_level": 1,
+ "bbox": [
+ 163,
+ 114,
+ 334,
+ 130
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/9d7e40286ddeb28b93be6916d8101d062aa91f512e1c1ac6b0ccc889ae647631.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 339,
+ 141,
+ 359,
+ 157
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Text Rendering Precision.",
+ "text_level": 1,
+ "bbox": [
+ 362,
+ 145,
+ 656,
+ 160
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/08d9d7d0f4498fa74ab30be14fe461e1fd36819539ce56b0a1505126518de149.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 158,
+ 162,
+ 295,
+ 271
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/234feaba9c9e1d0f65d67e6df566f551271d9b6bc16b92cc18990362a5b21b2f.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 315,
+ 164,
+ 452,
+ 271
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/0ce5a44bd3cd7500b4b78db75c0f1ccc81baa684ffa77ebdb728c28b0fe9e785.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 470,
+ 164,
+ 609,
+ 271
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/204fddddcbcd5e6532fa8831a9afe9f12c4bf879dee56a592fddb35d74766418.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 627,
+ 164,
+ 838,
+ 271
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "Input Text: \"A beautiful painting of flowing colors and styles forming the words 'The GPT-4o/Ideogram/FLUX/SD3 research paper is nowhere!'. the background is speckled with drops and splashes of paint.\"",
+ "bbox": [
+ 163,
+ 279,
+ 818,
+ 324
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/5d4b6580226201ab5042c29511d6da5a3409b9f5c7fee1d2d9c074d14f64765b.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 158,
+ 330,
+ 251,
+ 438
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/6ec2422a775f3f017131805ebc131bf30d8d32953e433ab7a7e35ce013d814bd.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 274,
+ 330,
+ 413,
+ 438
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/4f7ea6846a9f2e71ede47f66f8045aa234b77d2155c2f56eca68f95514f68e80.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 437,
+ 330,
+ 575,
+ 438
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/1b62b31f138cbdfe4ceec53b959cee36fc3c5f0c22524ab44623a5cd565ac9c8.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 599,
+ 330,
+ 839,
+ 438
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "Input Text: \"Beautiful pixel art of a Wizard with hovering text 'Achievement unlocked: Diffusion models can spell now'.\"",
+ "bbox": [
+ 166,
+ 446,
+ 781,
+ 477
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/4ba3110052c7723a701e2ccf5e847ad8c21956fcb34a978c1e55cce340d0131c.jpg",
+ "image_caption": [
+ "Figure 5: Task: Short text rendering. Generate prompt-aligned, concise textual content (typically within 10 words) on an image. Setup: Each sample is produced based on a guiding text prompt. Comparisons are made with prior SOTA models [27, 2] and FLUX [51]. Observations: GPT-4o achieves performance on par with existing SOTA baselines in rendering short texts, consistently following the prompt with minimal errors. All evaluated methods—except FLUX [51]—deliver high-fidelity results in this setting."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 158,
+ 486,
+ 251,
+ 592
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/4fed5bdfe6b3d24177ff8f96422315e81b3ae864616ce0f21770566bf7fd8891.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 282,
+ 486,
+ 501,
+ 592
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/83fabbd535e803f344bff99d3d53a6c8a71d95998a6a377d7bae9a19ebb0830a.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 532,
+ 486,
+ 669,
+ 592
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/63a0fda286dbcf6aab1f70694ab0576b25d5fd247827bdfa027aca3ce2984b08.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 702,
+ 486,
+ 839,
+ 592
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "Input Text: \"A monkey holding a sign reading 'Scaling transformer models is awesome'.\"",
+ "bbox": [
+ 166,
+ 603,
+ 784,
+ 618
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/b44926116e89a328041a0572f04bcac45c5be77916af431b03c4a0ab99dd761f.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 158,
+ 631,
+ 295,
+ 738
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/cb9fbff8e79609bd0c0397e9ff8bdd26c4934e0c043bf90309793e247c53acde.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 305,
+ 631,
+ 442,
+ 738
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/ecba2242bac62c8676019c58e03ec1fc8ba9f66e7431827caf98cf3565ae7429.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 452,
+ 631,
+ 589,
+ 738
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "image",
+ "img_path": "images/eb143f4361a33ce174515f230c11a216ac7590d56a2fc328a2adc862245f0ce7.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 601,
+ 632,
+ 836,
+ 738
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "Input Text: \"A surreal and humorous scene in a classroom with the words 'GPUs go brrrrr' written in white chalk on a blackboard. In front of the blackboard.\"",
+ "bbox": [
+ 165,
+ 744,
+ 803,
+ 773
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "GPT 40",
+ "bbox": [
+ 200,
+ 784,
+ 251,
+ 797
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "Ideogram 3.0",
+ "bbox": [
+ 326,
+ 784,
+ 421,
+ 799
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "FLUX",
+ "bbox": [
+ 500,
+ 784,
+ 542,
+ 797
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "SD3",
+ "bbox": [
+ 700,
+ 784,
+ 736,
+ 797
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "page_number",
+ "text": "11",
+ "bbox": [
+ 490,
+ 935,
+ 506,
+ 946
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Text Rendering Precision.",
+ "text_level": 1,
+ "bbox": [
+ 374,
+ 127,
+ 643,
+ 141
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/afce16d9cf439bcdc04fa94f8ca4f2227195b2cf36f175b299fa1b98664ddbfa.jpg",
+ "image_caption": [
+ "GPT 40"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 181,
+ 146,
+ 320,
+ 306
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/c6b354d42faa888d692b35bdfbc5cd4599018152cf0064da6cd13e513960b86e.jpg",
+ "image_caption": [
+ "POSTA"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 321,
+ 146,
+ 462,
+ 306
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/f4cd9c8e039a64fce1592ca2bb40e8f9f9bf8f8c530f65984c76ff6708a91c6f.jpg",
+ "image_caption": [
+ "Gemini 2.0 Flash"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 464,
+ 146,
+ 604,
+ 306
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/c5582fe68220a5a384efa294ef35989ebef4760bd5b2897bd0ee23ceffe344a0.jpg",
+ "image_caption": [
+ "Ideogram 3.0"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 604,
+ 146,
+ 812,
+ 306
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Input Text:",
+ "text_level": 1,
+ "bbox": [
+ 187,
+ 327,
+ 269,
+ 339
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "\"Generate a movie poster with a sci-fi space theme, a solitary figure standing on an alien planet, facing a massive outpost.",
+ "bbox": [
+ 187,
+ 340,
+ 808,
+ 366
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "The poster displays the following text:",
+ "bbox": [
+ 187,
+ 366,
+ 441,
+ 378
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Title: The Last Outpost",
+ "bbox": [
+ 187,
+ 378,
+ 344,
+ 391
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Subtitle: When the stars fall, the truth rises",
+ "bbox": [
+ 187,
+ 391,
+ 486,
+ 404
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Information:",
+ "text_level": 1,
+ "bbox": [
+ 187,
+ 405,
+ 276,
+ 416
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Produced by Jackson Ward",
+ "bbox": [
+ 187,
+ 417,
+ 367,
+ 430
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Music by Aria Calloway",
+ "bbox": [
+ 187,
+ 431,
+ 336,
+ 443
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Screenplay by Elena Sharpe",
+ "bbox": [
+ 187,
+ 443,
+ 370,
+ 455
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Directed By Sylvia Hartman",
+ "bbox": [
+ 187,
+ 455,
+ 370,
+ 469
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "\"A visually stunning and narratively gripping exploration of the unknown. The Last Outpost masterfully blends elements of science fiction, mystery, and psychological thriller, creating a hauntingly atmospheric journey that will leave audiences on the edge of their seats.\" -- Global Film Review\".",
+ "bbox": [
+ 187,
+ 469,
+ 795,
+ 518
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/e1dc82e6233835aef84e7c0d62fd27535b3b4aee7f6bc4beab765dacdbbdc4c1.jpg",
+ "image_caption": [
+ "GPT 40"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 181,
+ 523,
+ 316,
+ 681
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/59f2bdbb4c18ff5b2eaf1af8462de2d557b98afb5403da1a76ec1f639e6017a6.jpg",
+ "image_caption": [
+ "POSTA"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 318,
+ 523,
+ 460,
+ 683
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/3c2edc361f25ec06384618fcf44a118efef67ed2c14c3a957eec1c7c432abf66.jpg",
+ "image_caption": [
+ "Gemini 2.0 Flash"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 460,
+ 523,
+ 616,
+ 683
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/1751faef203209061a6fae572da8f65fff10f9f74ef783748441d19f7f4a2be9.jpg",
+ "image_caption": [
+ "Ideogram 3.0",
+ "Figure 6: Task: Long text rendering. Generate extended, coherent, and prompt-consistent textual content on an image. Setup: Evaluations are conducted against advanced baselines including POSTA [12], Gemini 2.0 Flash [99], Ideogram 3.0 [2], and Playground-v3 [64]. Observations: GPT-4o excels in long text rendering by producing coherent, detailed textual information with very few character errors. In contrast, models like Gemini 2.0 Flash, Ideogram 3.0, and Playground-v3 often exhibit increased errors or generate vague text when faced with lengthy prompts, while POSTA's tailored multi-step pipeline sometimes yields competitive precision. Overall, GPT-4o outperforms most commercial models and rivals specialized research approaches in extended text generation."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 617,
+ 523,
+ 823,
+ 683
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Input Text:",
+ "text_level": 1,
+ "bbox": [
+ 184,
+ 705,
+ 264,
+ 718
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "\"Create a poster with the theme of a Journey of Solitude. The background should depict a lone figure walking toward an unusable form of transportation. The scene should evoke a sense of being lost, helplessness, and desolation, capturing the emotional weight of losing oneself in a barren, unforgiving landscape.",
+ "bbox": [
+ 183,
+ 719,
+ 800,
+ 770
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Title: Solitary Journeys",
+ "bbox": [
+ 184,
+ 770,
+ 341,
+ 784
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Subtitle: Elara Voss",
+ "bbox": [
+ 184,
+ 784,
+ 316,
+ 795
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Information: WANDERING THROUGH THE UNKNOWN\".",
+ "bbox": [
+ 184,
+ 796,
+ 563,
+ 808
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "header",
+ "text": "Long Text Rendering",
+ "bbox": [
+ 207,
+ 99,
+ 359,
+ 116
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "page_number",
+ "text": "12",
+ "bbox": [
+ 490,
+ 935,
+ 509,
+ 946
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Long Text Rendering",
+ "text_level": 1,
+ "bbox": [
+ 200,
+ 291,
+ 356,
+ 308
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "image",
+ "img_path": "images/072dabf79d619daa8eeb4cd2f8867859c2a5525ca4ceff1ee839637ee46f829f.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 346,
+ 316,
+ 367,
+ 333
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Text Rendering Precision.",
+ "text_level": 1,
+ "bbox": [
+ 367,
+ 320,
+ 647,
+ 335
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "image",
+ "img_path": "images/37c0d9d0ba1eed41834771a2d8df3690d2b37294a08a5844702161e53aa817ab.jpg",
+ "image_caption": [
+ "GPT 40",
+ "Figure 7: Task: Long text rendering. The Setup and Observations are the same as Figure 6."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 166,
+ 339,
+ 307,
+ 503
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "image",
+ "img_path": "images/61b9d5a9c147c446387301c4f13acd7ceeccc06f653954410c1dc398d1fd4bc3.jpg",
+ "image_caption": [
+ "POSTA"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 308,
+ 339,
+ 454,
+ 503
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "image",
+ "img_path": "images/96c1a55c7c5f312a71c89d34f708b0e4f8da4556922ffdee41d673b76ea82dad.jpg",
+ "image_caption": [
+ "Gemini 2.0 Flash"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 454,
+ 339,
+ 614,
+ 503
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "image",
+ "img_path": "images/faf4eca5a3720dff227ce6a4ad8f4898de848e4e08fec90adf87952d14446f7e.jpg",
+ "image_caption": [
+ "Playground-v3"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 616,
+ 339,
+ 828,
+ 503
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Input Text:",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 537,
+ 259,
+ 550
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "\"Please generate an artistic and stylized promotional poster. The style is an artistic painting style. The theme is about nature and city. The poster displays the following information: Title: Fragmented Harmony",
+ "bbox": [
+ 174,
+ 551,
+ 792,
+ 590
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Subtitle Between the steel and sky, life finds its way.",
+ "bbox": [
+ 174,
+ 590,
+ 535,
+ 604
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "Information: Amid the towering strucions and the quiet persistence of nature, a delicate balance emerges. The complex and often contradictory relationship between urban development and the natural world reveals itself in fleeting moments of harmony. Though fragmented, life continues, threading its way through the shadows of progress. Here, conflict and coexistence form an intricate dance--sometimes at odds, sometimes in unexpected unity\".",
+ "bbox": [
+ 173,
+ 604,
+ 813,
+ 671
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "page_number",
+ "text": "13",
+ "bbox": [
+ 490,
+ 935,
+ 506,
+ 946
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "2.1.3 Document Generation",
+ "text_level": 1,
+ "bbox": [
+ 127,
+ 90,
+ 338,
+ 104
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "We also explore a novel task: document image generation with GPT-4o, comparing its performance with Gemini 2.0 Flash [99] and Playground-v3 [64]. As shown in Figure 8 - 10, GPT-4o produces document images with cleaner layouts and more consistent content.",
+ "bbox": [
+ 125,
+ 114,
+ 870,
+ 157
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Document Image Generation",
+ "text_level": 1,
+ "bbox": [
+ 169,
+ 176,
+ 392,
+ 191
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "image",
+ "img_path": "images/7ff25ac80d044220739d61284585a070ee9142eede21fe8dbf47dff3cb2ffa9c.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 334,
+ 203,
+ 356,
+ 220
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Text Rendering Precision.",
+ "text_level": 1,
+ "bbox": [
+ 357,
+ 207,
+ 650,
+ 222
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Attention Is All You Need",
+ "text_level": 1,
+ "bbox": [
+ 187,
+ 255,
+ 331,
+ 266
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Ashish Vaswani Noam Shazeer Niki Parmar Jakob Uszkoreit Lilion Jones Aidan N.Gomez Lukasz Kaiser IIIa Polosukhin",
+ "bbox": [
+ 184,
+ 276,
+ 339,
+ 297
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 243,
+ 310,
+ 277,
+ 318
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on tw machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature.",
+ "bbox": [
+ 158,
+ 320,
+ 367,
+ 414
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Attention Is All You Need",
+ "text_level": 1,
+ "bbox": [
+ 403,
+ 323,
+ 501,
+ 330
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Ashish Vourakis, NM Nazarov, NJI Parmar, Jbab Udekoreli, Lillion Jones, Adiinil S. W. Coomce, Lakota Kiber, Pala Poslashov, and M. A. D. G. R. Smith. 2017. The neural network models are based on recurrent or conventional neural networks in an encoder-decoder loss. The best produced models also connect the encoder and decoder loss through an attention mechanism. As, we propose a new study mechanism for a machine, the Transformer, based attention mechanisms, dispensing with the same weight as the encoder-decoder loss. The results of the training tasks show these models to be superior in quality while being more parsibilized and requiring significantly less time to time to move train. Our model achieves 84.2 BLUE in the WMT-50 Go-to-translation task, which is comparable to the performance of our previous work [3]. In addition, our WMT-30 Pragmatic task, our model established a new single-model state-of-the-art-state-of-the-art BLUE score of 41.8 when $\\alpha = 5$ training for 3.5 days on GPUs after fraction of the training costs of the best models from literature. We propose Transformer generalizes well by applying it successfully to English syntacticity parsing both with large and limited training data.",
+ "bbox": [
+ 397,
+ 330,
+ 566,
+ 415
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Attention Is All You Need",
+ "text_level": 1,
+ "bbox": [
+ 602,
+ 228,
+ 705,
+ 236
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Ashish A. [Al] You. Wea a nono' Ainon 1 \nAshish Yawani, Naqadzaree, laokotri \nAlok Uzzotner, Anokalika Sanik, Jokoslav adar, Gosak III Ploosukhaini \nAlok Uzzotner, Anokalika Sanik",
+ "bbox": [
+ 602,
+ 241,
+ 692,
+ 268
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Antext, exotocic sequecra transoedscnncs on cbr be caes baccared on bracococcyne nort bell netiabon an ecocycclion, an ecocyclon. TeTrane: the ensonnnmnsn neeepnckian. Epipcnie rile kceely on meenctiny As adeterdiencr. \nnpopioors sonr tonarwamchim. I mtnr. vorti. inenpoea a dedusum minnyss onomcrh. cortordone.lora ontata tose or uin hoperiosper. \nThe rorner is s oovl dtt maive de acemnccnodkdu aleu cormunb-dlr bing dvl-ndr 016de and mechance \n11 Dae nucnccnng \nceso nucnse \n12 - eannnr er on attonne amehnes asnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnnccnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnn",
+ "bbox": [
+ 602,
+ 280,
+ 828,
+ 417
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "GPT40",
+ "bbox": [
+ 233,
+ 426,
+ 290,
+ 439
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Gemini 2.0 Flash",
+ "bbox": [
+ 416,
+ 426,
+ 545,
+ 441
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Playground-v3",
+ "bbox": [
+ 663,
+ 425,
+ 772,
+ 441
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Input Text:",
+ "text_level": 1,
+ "bbox": [
+ 161,
+ 448,
+ 259,
+ 463
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "\"Generate A realistic screenshot of the first page of the Paper from the following information:",
+ "bbox": [
+ 158,
+ 465,
+ 795,
+ 500
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Title: Attention Is All You Need",
+ "bbox": [
+ 158,
+ 503,
+ 411,
+ 518
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Author List: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin",
+ "bbox": [
+ 158,
+ 522,
+ 833,
+ 556
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Abstract: The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.\"",
+ "bbox": [
+ 158,
+ 560,
+ 831,
+ 824
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Figure 8: Task: Document image generation. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Playground-v3 [64]. Observation: GPT-4o can generate more consistent and accurate font and format than the other two models.",
+ "bbox": [
+ 125,
+ 842,
+ 870,
+ 883
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "page_number",
+ "text": "14",
+ "bbox": [
+ 490,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Document Image Generation",
+ "text_level": 1,
+ "bbox": [
+ 169,
+ 150,
+ 390,
+ 167
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "image",
+ "img_path": "images/e960d89a9332326f5fac58239c1d5784d9e237d7ccede93ea1c82462b3f589b3.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 334,
+ 176,
+ 354,
+ 194
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Text Rendering Precision.",
+ "text_level": 1,
+ "bbox": [
+ 356,
+ 180,
+ 650,
+ 196
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
+ "text_level": 1,
+ "bbox": [
+ 161,
+ 224,
+ 379,
+ 250
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Jacob Devlin Ming-Wei Chang Kenton Lee Kristina Toutanova",
+ "bbox": [
+ 171,
+ 260,
+ 334,
+ 277
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 251,
+ 290,
+ 287,
+ 299
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result; the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.",
+ "bbox": [
+ 158,
+ 301,
+ 383,
+ 366
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to $80.5\\%$ (7.7% point absolute improvement), MultiNLI accuracy to $\\mathcal{S}6.7\\%$ (4.6% absolute improvement), SQAud v1.1 question answering Test F1 to 93.2 (1.5 point absolute",
+ "bbox": [
+ 158,
+ 366,
+ 387,
+ 402
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "GPT 40",
+ "bbox": [
+ 241,
+ 409,
+ 299,
+ 422
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
+ "text_level": 1,
+ "bbox": [
+ 416,
+ 224,
+ 583,
+ 247
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Ashlor Jacob Doslin Ming Wei Chang, Kenton Lee Abstrut Win, Karinlin Touranosa",
+ "bbox": [
+ 419,
+ 250,
+ 578,
+ 265
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Cryostatnet as Transformer. Unlike recnec mean, BERT is designed by pre-train deep bidirectional representations from nonshaded text by jasily selfconforming on both text 1st&4, xavier coint eorect, tate 1st&2 to win 1st, sacp0-twincent BERT model can be fine mused with just one additional output layer to create of dafve for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.",
+ "bbox": [
+ 416,
+ 265,
+ 588,
+ 320
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "BERT is conceptually simple and empirically powerful. It obtains now state-of-the-art results on eleven evertweaven multimoment language processing tasks, including GLUE score to $80.5\\%$ (7.69 absolute improvement), pushing the GLUE alloselect improvement). MulaNLI accuracy to $\\varepsilon_{\\mathrm{M}} = 0.1$ , $\\nu_{\\mathrm{L}} < \\varepsilon_{\\mathrm{M}} / \\varepsilon_{\\mathrm{M}}$ (1.5 point absolute improvement) and SQoAD v2.0 Test F1 to 83.1 (3.1 point absolute improvement).",
+ "bbox": [
+ 416,
+ 327,
+ 589,
+ 375
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Gemini 2.0 Flash",
+ "bbox": [
+ 437,
+ 410,
+ 566,
+ 424
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "BERT: Pre-training on Lepi Bidellar Tansson Translons for Language Understond mderance litting' cetting from t cowf henvaming",
+ "text_level": 1,
+ "bbox": [
+ 620,
+ 229,
+ 823,
+ 247
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "Author, List:",
+ "J Asbad Devlin, Yiw Changuaguagaa Kionn age rspgectangane cans \n $^{a}$ pressin liKistcn-Toutanfa",
+ "represeons-Uintanlvania"
+ ],
+ "bbox": [
+ 620,
+ 250,
+ 820,
+ 273
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "A",
+ "bbox": [
+ 620,
+ 275,
+ 647,
+ 282
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "We introduce a new languagegretrovetercendentiale monoclin klonstionist monole conBldfecarstaadss reprenters from nemer raje Sfiflnonanecones. desessnissrall ranauagaleafdelfe xyn unming on hnlaesaeare two ploddes also the por-entant canteletory a state-vwraon one-on-one coffice of anisotropy, and the ploso-syntropic colective of states. Of s2s212310 pain or questionbmporansuansluangene Tcnf?f to ingest ingf Sf10 tto 46 w. (I, test),marshersonalizne imnance immmens",
+ "bbox": [
+ 620,
+ 286,
+ 834,
+ 328
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "BERT is conceptually imlilienenarplenpholcft-nu surate-fine-ams",
+ "ronen 1 oonranaalwauanu viipopluoteforocnirandinns inget caught anourage, vovlurvulgina for nain. 2004. The use of the word \"sulfur\" in the text is a question envoing SUIF u697.",
+ "- GBFEscors/aanoreqquasurf and Squad w.10 aninlvalte 83.7% 4.6% (X) \n- TeST fto onop:11.3% (x)",
+ "→ BERT is cponconuynlyrsnaintally pocefine-at-ut: ouvah176. (JET v.c.37% quinting anguen linyuH-aCLS sccorts onoonssthe sonea 4000A/AresoVc LEAU pioiHcB: gnrmaeh an epesrourinans A7c)0v o.o35 aed 1170"
+ ],
+ "bbox": [
+ 620,
+ 338,
+ 836,
+ 402
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Playground-v3",
+ "bbox": [
+ 676,
+ 409,
+ 785,
+ 426
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Input Text:",
+ "text_level": 1,
+ "bbox": [
+ 161,
+ 440,
+ 258,
+ 455
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Generate A realistic screenshot of the first page of the Paper from the following information:",
+ "bbox": [
+ 158,
+ 457,
+ 789,
+ 491
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Title: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
+ "bbox": [
+ 158,
+ 494,
+ 733,
+ 532
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Author List: Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova \nAbstract: We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.",
+ "bbox": [
+ 158,
+ 534,
+ 834,
+ 700
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to $80.5\\%$ (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).",
+ "bbox": [
+ 158,
+ 704,
+ 834,
+ 797
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "Figure 9: Task: Document image generation. The Setup and Observations are the same as Fig. 8.",
+ "bbox": [
+ 176,
+ 837,
+ 816,
+ 853
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "page_number",
+ "text": "15",
+ "bbox": [
+ 490,
+ 935,
+ 506,
+ 946
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "image",
+ "img_path": "images/9c2265499ec9f6a0c6e61a8914311c8b94d20424cd7a53a126ce5f30f3f6bf33.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 334,
+ 152,
+ 354,
+ 172
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Evaluation: Text Rendering Precision.",
+ "text_level": 1,
+ "bbox": [
+ 357,
+ 157,
+ 650,
+ 174
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "You Only Look Once: Unified, Real-Time Object Detection",
+ "text_level": 1,
+ "bbox": [
+ 169,
+ 212,
+ 344,
+ 250
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Joseph Redmon, Santosh Divvala, Ross Girshick Ali Farhadi",
+ "bbox": [
+ 174,
+ 258,
+ 359,
+ 273
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 250,
+ 285,
+ 284,
+ 292
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "We present YOLO, a new approach to object detection. Prior work on object detection repurposes classifiers to perform detection. Instead, we frame object detection as a regression problem to spatially separated bounding boxes and associated class probabilities. A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation. Since the whole detection pipeline is a single network, it can be optimized end-to-end directly on detection performance.",
+ "bbox": [
+ 166,
+ 292,
+ 375,
+ 345
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Our unified architecture is extremely fast. Our base YOLO model processes images in real-time at 45 frames per second. A smaller version of the network. Fast YOLO, processes an astounding 155 frames per second while still achieving double the mAP of other real-time detectors. Compared to state-of-the-art detection systems YOLO makes more localization errors but is far less likely to pred",
+ "bbox": [
+ 166,
+ 345,
+ 375,
+ 387
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "You Only Look Once: Unlimited, Time/Indirect Decciption \nAuthor:Lesser Jowesh Redmonial Siptedthi Adri Farhad",
+ "bbox": [
+ 415,
+ 215,
+ 580,
+ 233
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "We present YJLO, a new approach to object detection. Driver search uses oleejeon cieeuses to represen or correct deferential finwication. Istegedel, when farstial, we confirmed anapase, agate trius to a signafal or sialipal staphylococcal preprobed boing boos summarilyour heartbodn frontal nater and the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the",
+ "bbox": [
+ 413,
+ 234,
+ 622,
+ 258
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Acatat Rechun-Yorim Bogae lo rojctc filocly",
+ "bbox": [
+ 413,
+ 260,
+ 496,
+ 265
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Our undersonged tandoce boe + ant - or cemoe - aepocemis in a times, bodingly narmabogus haarban. Jnci enwecstic ancoontie fluy fast YOLO pue moebes, 45f rans perennetioles unperenentio- dorensis in reel thane reobexes petarges. ectcylcfom oene trsoute of sucrose princeia of docuta. d. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..",
+ "bbox": [
+ 413,
+ 268,
+ 622,
+ 295
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "table",
+ "img_path": "images/5789a11c88dd67fe3ecf18424e18808df9be15ba558dca4d60893646f7ed02f2.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Case Figure | Meta-task | Sub-task | GPT-4o | Gemini-2.0-flash | Domain-SOTA |
| Figure 1 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 2 | | Complex Text Following | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 3 | | | Success | Success | Success |
| Figure 4 | | | Success | Success | Success |
| Figure 5 | | | Success | Success | Success |
| Figure 6 | Text-to-Image | Text Rendering | Success | Low Visual Quality | Low Visual Quality |
| Figure 7 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 8 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 9 | | Document Generation | Success | Low Visual Quality | Low Visual Quality |
| Figure 10 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 11 | | Panorama | Lack of Knowledge | Success | Success |
| Figure 12 | | Style Transfer | Success | Lack of Knowledge | Lack of Knowledge |
| Figure 13 | | | Success | Lack of Knowledge | Lack of Knowledge |
| Figure 14 | | | Low Visual Quality | Success | Failure to Follow Instructions |
| Figure 15 | | Image Editing | Failure to Follow Instructions | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 16 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 17 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 18 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 19 | | | Success | Inconsistent Generation | Failure to Follow Instructions |
| Figure 20 | | Single-Concept Customization | Success | Failure to Follow Instructions | Success |
| Figure 21 | | Multi-Concept Customization | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 22 | | Story Image Generation | Success | Failure to Follow Instructions | Success |
| Figure 23 | | | Success | Inconsistent Generation | Success |
| Figure 24 | | Low-Level Vision-Denoising | Low Visual Quality | Low Visual Quality | Success |
| Figure 25 | | Low-Level Vision-Deraining | Success | Inconsistent Generation | Success |
| Figure 26 | | Low-Level Vision-Dehazing | Success | Low Visual Quality | Success |
| Figure 27 | | Low-Level Vision-Low Light Enhancement | Low Visual Quality | Low Visual Quality | Success |
| Figure 28 | | Low-Level Vision-Deblurring | Success | Low Visual Quality | Success |
| Figure 29 | | Low-Level Vision-Super Resolution | Success | Low Visual Quality | Success |
| Figure 30 | | Low-Level Vision-Impainting | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 31 | | Low-Level Vision-Outpainting | Inconsistent Generation | Success | Success |
| Figure 32 | | Low-Level Vision-Colorization | Success | Success | Success |
| Figure 33 | | Low-Level Vision-Shadow Removal | Success | Failure to Follow Instructions | Success |
| Figure 34 | | Low-Level Vision-Reflection Removal | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 35 | | Low-Level Vision-Relighting | Success | Failure to Follow Instructions | Success |
| Figure 36 | | Spatial Control-Canny | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 37 | | Spatial Control-Depth | Success | Failure to Follow Instructions | Success |
| Figure 38 | | Spatial Control-Sketch | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 39 | | Spatial Control-Pose | Success | Inconsistent Generation | Success |
| Figure 40 | | Spatial Control-Mask | Inconsistent Generation | Failure to Follow Instructions | Inconsistent Generation |
| Figure 41 | | Camera Control | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 42 | | | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 43 | | In-Context Visual Prompting | Failure to Follow Instructions | Failure to Follow Instructions | N/A |
| Figure 44 | | Image to 3D Modeling | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 45 | | UV Map to 3D Rendering | Success | Inconsistent Generation | Failure to Follow Instructions |
| Figure 46 | | Novel View Synthesis | Success | Success | Failure to Follow Instructions |
| Figure 47 | | Image Segmentation | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 48 | | | Success | Failure to Follow Instructions | Success |
| Figure 49 | | | Success | Failure to Follow Instructions | Success |
| Figure 50 | | Edge Detection | Success | Success | Success |
| Figure 51 | | | Success | Failure to Follow Instructions | Success |
| Figure 52 | | | Success | Failure to Follow Instructions | Success |
| Figure 53 | | Salient Object | Success | Failure to Follow Instructions | Success |
| Figure 54 | | | Success | Success | Success |
| Figure 55 | | | Success | Success | Success |
| Figure 56 | | Depth Estimation | Success | Failure to Follow Instructions | Success |
| Figure 57 | | Normal Estimation | Success | Failure to Follow Instructions | Success |
| Figure 58 | | Layout Detection | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 59 | | Text Detection | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 60 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 61 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 62 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 63 | | | Inconsistent Generation | Inconsistent Generation | Success |
"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.936,
+ 0.506,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "4"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.128,
+ 0.092,
+ 0.312,
+ 0.106
+ ],
+ "angle": 0,
+ "content": "2.1 Text-to-Image Tasks"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.128,
+ 0.117,
+ 0.429,
+ 0.133
+ ],
+ "angle": 0,
+ "content": "2.1.1 Complex Text Following Capability"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.125,
+ 0.141,
+ 0.872,
+ 0.282
+ ],
+ "angle": 0,
+ "content": "Recent progress in text-to-image generation has shown impressive abilities in generating diverse and realistic images based on text prompts. However, composing multiple objects with various attributes and relationships accurately into one scene remains a significant challenge for current text-to-image generative models [92, 85, 8, 81, 6]. In this section, we assess models' ability for compositional text-to-image generation from four perspectives following [41], which include attribute binding, numeracy, object relationship, and complex compositions. Attribute binding evaluates whether the model correctly assigns attributes, such as color, shape, and texture to the appropriate objects. Numeracy evaluates whether the number of generated objects matches the quantities specified in the prompt. Object relationships refer to both spatial (2D/3D) and non-spatial interactions among objects. Complex compositions evaluate the model's ability to handle multiple types of constraints simultaneously, especially given long or detailed prompts."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.125,
+ 0.285,
+ 0.872,
+ 0.398
+ ],
+ "angle": 0,
+ "content": "As shown in Figure 1 row 1, GPT-4o outperforms both Gemini 2.0 Flash and Midjourney in numeracy tasks. While GPT-4o accurately represents a single plate, Gemini 2.0 and Midjourney represent two plates instead. In terms of understanding object relationships, GPT-4o is the only model that correctly infers the action \"walk towards\" from the ragdoll to the labrador. However, GPT-4o struggles with more complex terms like \"pentagonal pyramid\", failing to interpret it correctly (see Figure 1 row 4). This suggests that GPT-4o may have difficulty accurately interpreting objects with unusual geometries. When it comes to abstract prompts, GPT-4o also appears to lack imagination (see Figure 2 row 2), whereas Midjourney v6.1 demonstrates better creativity in this case, outperforming both GPT-4o and Gemini 2.0 Flash."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.125,
+ 0.402,
+ 0.872,
+ 0.502
+ ],
+ "angle": 0,
+ "content": "For complex text-to-image generation, we evaluate GPT-4o's performance with Gemini 2.0 Flash [99] and FLUX.1-Pro [51], using the text prompts collected from [124, 106, 115]. As shown in Figure 3, both GPT-4o and FLUX excel at generating realistic and harmonious scenes align with the text prompts. However, we observe that GPT-4o shows limitations in generating culturally related elements. For example, the generated crown for the Chinese general is western-style rather than chinese-style (see Figure 4 row 2). Additionally, in large scene generation, GPT-4o struggles to maintain boundary continuity, whereas FLUX produces a more natural composition (see Figure 4 row 3)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.125,
+ 0.506,
+ 0.871,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "Overall, we conclude that GPT-4o excels at text-to-image generation in terms of attribute binding, generative numeracy, object relationship, and complex compositions. However, it exhibits limitations in generating uncommon objects, culturally specific elements and in maintaining continuity when composing large scenes."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.936,
+ 0.506,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "5"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.192,
+ 0.101,
+ 0.393,
+ 0.117
+ ],
+ "angle": 0,
+ "content": "Text-to-Image Generation"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.252,
+ 0.127,
+ 0.271,
+ 0.144
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.272,
+ 0.13,
+ 0.774,
+ 0.146
+ ],
+ "angle": 0,
+ "content": "Evaluation: Visual content precisely following the text instruction."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.172,
+ 0.147,
+ 0.364,
+ 0.294
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.395,
+ 0.147,
+ 0.588,
+ 0.296
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.615,
+ 0.147,
+ 0.824,
+ 0.296
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.176,
+ 0.298,
+ 0.61,
+ 0.312
+ ],
+ "angle": 0,
+ "content": "Input Text: \"A yellow bowl, a blue mug and a pink plate on the table.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.172,
+ 0.314,
+ 0.367,
+ 0.462
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.395,
+ 0.313,
+ 0.59,
+ 0.465
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.615,
+ 0.311,
+ 0.822,
+ 0.465
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.179,
+ 0.465,
+ 0.49,
+ 0.478
+ ],
+ "angle": 0,
+ "content": "Input Text: \"A ragdoll walks towards a labrador.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.172,
+ 0.481,
+ 0.371,
+ 0.628
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.396,
+ 0.481,
+ 0.59,
+ 0.628
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.615,
+ 0.479,
+ 0.822,
+ 0.628
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.174,
+ 0.63,
+ 0.812,
+ 0.656
+ ],
+ "angle": 0,
+ "content": "Input Text: \"Three differently colored apples (yellow, green, red from left to right) with a Coca-Cola bottle placed behind the middle apple.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.172,
+ 0.658,
+ 0.37,
+ 0.807
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.395,
+ 0.658,
+ 0.59,
+ 0.807
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.615,
+ 0.658,
+ 0.822,
+ 0.807
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.174,
+ 0.809,
+ 0.819,
+ 0.824
+ ],
+ "angle": 0,
+ "content": "Input Text: \"The oval sphere was nestled between the rectangular prism and the pentagonal pyramid.\""
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.236,
+ 0.83,
+ 0.288,
+ 0.843
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.437,
+ 0.832,
+ 0.551,
+ 0.845
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.666,
+ 0.83,
+ 0.776,
+ 0.845
+ ],
+ "angle": 0,
+ "content": "Midjourney v6.1"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.127,
+ 0.868,
+ 0.871,
+ 0.926
+ ],
+ "angle": 0,
+ "content": "Figure 1: Task: Compositional text-to-image generation. Evaluate the image-text alignment on attribute binding, numeracy, and object relationship. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Midjourney v6.1 [75]. Observation: GPT-4o outperforms Gemini 2.0 Flash and Midjourney v6.1 across all aspects. However, GPT-4o struggles with uncommon objects with a special geometry."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.937,
+ 0.505,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "6"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.195,
+ 0.103,
+ 0.393,
+ 0.119
+ ],
+ "angle": 0,
+ "content": "Text-to-Image Generation"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.254,
+ 0.128,
+ 0.271,
+ 0.145
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.274,
+ 0.131,
+ 0.774,
+ 0.146
+ ],
+ "angle": 0,
+ "content": "Evaluation: Visual content precisely following the text instruction."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.177,
+ 0.147,
+ 0.373,
+ 0.294
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.395,
+ 0.147,
+ 0.612,
+ 0.294
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.634,
+ 0.147,
+ 0.825,
+ 0.293
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.176,
+ 0.296,
+ 0.832,
+ 0.319
+ ],
+ "angle": 0,
+ "content": "Input Text: \"The round, juicy watermelon sat in the cool, refreshing bowl of ice, waiting to be sliced open and devoured.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.177,
+ 0.32,
+ 0.373,
+ 0.47
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.398,
+ 0.32,
+ 0.611,
+ 0.469
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.633,
+ 0.32,
+ 0.828,
+ 0.47
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.179,
+ 0.472,
+ 0.829,
+ 0.497
+ ],
+ "angle": 0,
+ "content": "Input Text: \"The bold, expressive strokes of the artist's brush brought the blank canvas to life, forming a vibrant and dynamic masterpiece.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.177,
+ 0.497,
+ 0.373,
+ 0.645
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.398,
+ 0.497,
+ 0.611,
+ 0.645
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.633,
+ 0.497,
+ 0.829,
+ 0.646
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.182,
+ 0.648,
+ 0.701,
+ 0.661
+ ],
+ "angle": 0,
+ "content": "Input Text: \"The heavy raindrops fell on the smooth glass and the textured roof.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.177,
+ 0.663,
+ 0.373,
+ 0.813
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.398,
+ 0.663,
+ 0.612,
+ 0.814
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.633,
+ 0.663,
+ 0.829,
+ 0.814
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.178,
+ 0.815,
+ 0.808,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "Input Text: \"The gentle, soothing melody of the piano filled the concert hall, as the pianist's fingers danced over the keys.\""
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.235,
+ 0.843,
+ 0.286,
+ 0.855
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.436,
+ 0.843,
+ 0.548,
+ 0.855
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.664,
+ 0.842,
+ 0.772,
+ 0.856
+ ],
+ "angle": 0,
+ "content": "Midjourney v6.1"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.127,
+ 0.873,
+ 0.87,
+ 0.943
+ ],
+ "angle": 0,
+ "content": "Figure 2: Task: Compositional text-to-image generation. Evaluate the image-text alignment on attribute binding and complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Midjourney v6.1 [75]. Observation: GPT-4o outperforms the other two models in generating objects aligned with the text prompts accurately. But for more abstract and creative tasks, Midjourney v6.1 performs the best."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.936,
+ 0.505,
+ 0.947
+ ],
+ "angle": 0,
+ "content": "7"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.176,
+ 0.108,
+ 0.394,
+ 0.138
+ ],
+ "angle": 0,
+ "content": "Text-to-Image Generation (with complex text prompt)"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.225,
+ 0.149,
+ 0.788,
+ 0.166
+ ],
+ "angle": 0,
+ "content": "Evaluation: Visual content precisely following the text instruction."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.174,
+ 0.168,
+ 0.361,
+ 0.312
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.392,
+ 0.168,
+ 0.579,
+ 0.312
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.611,
+ 0.168,
+ 0.798,
+ 0.312
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.165,
+ 0.318,
+ 0.825,
+ 0.384
+ ],
+ "angle": 0,
+ "content": "Input Text: \"An icy landscape. A vast expanse of snow-covered mountain peaks stretches endlessly. Beneath them is a dense forest and a colossal frozen lake. Three people are boating in three boats separately in the lake. Not far from the lake, a volcano threatens eruption, its rumblings felt even from afar. Above, a ferocious red dragon dominates the sky and commands the heavens, fueled by the volcano's relentless energy flow.\" (Prompt from GenArtist)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.174,
+ 0.391,
+ 0.361,
+ 0.533
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.392,
+ 0.39,
+ 0.579,
+ 0.534
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.61,
+ 0.39,
+ 0.797,
+ 0.534
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.171,
+ 0.541,
+ 0.825,
+ 0.596
+ ],
+ "angle": 0,
+ "content": "Input Text: \"On the rooftop of a skyscraper in a bustling cyberpunk city, a figure in a trench coat and neon-lit visor stands amidst a garden of bio-luminescent plants, overlooking the maze of flying cars and towering holograms. Robotic birds flit among the foliage, digital billboards flash advertisements in the distance.\" (Prompt from IterComp)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.174,
+ 0.6,
+ 0.361,
+ 0.743
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.392,
+ 0.599,
+ 0.58,
+ 0.744
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.611,
+ 0.599,
+ 0.797,
+ 0.743
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.171,
+ 0.747,
+ 0.825,
+ 0.801
+ ],
+ "angle": 0,
+ "content": "Input Text: \"In a magical seascape, a majestic ship sails through crystal blue waters surrounded by vibrant marine life and soaring birds. Towering cliffs frame the scene, while a stunning rainbow arches across the sky, blending with ethereal clouds. This enchanting journey captures the serene beauty of nature's wonders.\" (Prompt from IterComp)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.235,
+ 0.807,
+ 0.292,
+ 0.821
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.419,
+ 0.807,
+ 0.545,
+ 0.821
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.678,
+ 0.807,
+ 0.724,
+ 0.821
+ ],
+ "angle": 0,
+ "content": "FLUX"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.127,
+ 0.84,
+ 0.872,
+ 0.896
+ ],
+ "angle": 0,
+ "content": "Figure 3: Task: Compositional text-to-image generation. Evaluate the image-text alignment on complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and FLUX.1-Pro [51]. Observation: GPT-4o and FLUX can generate more harmonious and natural scene than Gemini 2.0 Flash."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.936,
+ 0.505,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "8"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.169,
+ 0.12,
+ 0.388,
+ 0.151
+ ],
+ "angle": 0,
+ "content": "Text-to-Image Generation (with complex text prompt)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.199,
+ 0.159,
+ 0.218,
+ 0.176
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.223,
+ 0.163,
+ 0.789,
+ 0.179
+ ],
+ "angle": 0,
+ "content": "Evaluation: Visual content precisely following the text instruction."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.172,
+ 0.183,
+ 0.359,
+ 0.326
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.395,
+ 0.184,
+ 0.58,
+ 0.326
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.612,
+ 0.184,
+ 0.796,
+ 0.326
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.156,
+ 0.332,
+ 0.813,
+ 0.386
+ ],
+ "angle": 0,
+ "content": "Input Text: \"Under the luminous full moon, a serene Japanese garden with traditional pagodas and a tranquil pond creates a magical night scene. The soft glow from the lantern-lit buildings reflects on the water, blending nature and architecture in harmony. The moonlight bathes the landscape, enhancing the peaceful ambiance.\" (Prompt from IterComp)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.199,
+ 0.395,
+ 0.334,
+ 0.548
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.415,
+ 0.395,
+ 0.559,
+ 0.548
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.628,
+ 0.395,
+ 0.779,
+ 0.549
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.163,
+ 0.556,
+ 0.82,
+ 0.61
+ ],
+ "angle": 0,
+ "content": "Input Text: \"A Chinese general wearing a crown, with whiskers and golden Chinese style armor, standing with a majestic dragon head on his chest, symbolizing his strength, wearing black and gold boots. His appearance exudes a sense of authority, wisdom, and an unyielding spirit, embodying the ideal ancient Chinese hero.\" (Prompt from RPG)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.152,
+ 0.618,
+ 0.359,
+ 0.726
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.362,
+ 0.619,
+ 0.506,
+ 0.726
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.506,
+ 0.618,
+ 0.653,
+ 0.726
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.656,
+ 0.618,
+ 0.844,
+ 0.726
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.163,
+ 0.734,
+ 0.82,
+ 0.788
+ ],
+ "angle": 0,
+ "content": "Input Text: \"A beautiful landscape with a river in the middle, the left of the river is in the evening and in the winter with a big iceberg and a small village while some people are skiing on the river and some people are skating, the right of the river is in the summer with a volcano in the morning and a small village while some people are playing.\" (Prompt from RPG)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.222,
+ 0.796,
+ 0.279,
+ 0.81
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.432,
+ 0.796,
+ 0.558,
+ 0.81
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.723,
+ 0.793,
+ 0.769,
+ 0.807
+ ],
+ "angle": 0,
+ "content": "FLUX"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.127,
+ 0.826,
+ 0.871,
+ 0.882
+ ],
+ "angle": 0,
+ "content": "Figure 4: Task: Compositional text-to-image generation. Evaluate the image-text alignment on complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and FLUX.1-Pro [51]. Observation: GPT-4o struggles to generate culturally related elements and maintain boundary continuity (see rows 2 and 3), similar to Gemini 2.0 Flash and FLUX."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.936,
+ 0.505,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "9"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.128,
+ 0.092,
+ 0.293,
+ 0.108
+ ],
+ "angle": 0,
+ "content": "2.1.2 Text Rendering"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.126,
+ 0.115,
+ 0.871,
+ 0.159
+ ],
+ "angle": 0,
+ "content": "Text rendering is a task that aims at generating texts (characters, sentences, or even paragraphs) on an image. The text content is usually guided by the input prompt. Previous models [27, 2] show good capability in generating short text (within 10 words, such as signs or short phrases), but their ability to generate long texts remains limited."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.126,
+ 0.164,
+ 0.87,
+ 0.22
+ ],
+ "angle": 0,
+ "content": "As shown in Figure 5, GPT-4o demonstrates comparable abilities to existing state-of-the-art (SOTA) baselines when generating short texts. All the methods except FLUX [51] perform well at rendering short text following the prompt. In this section, we primarily focus on long text rendering to examine whether GPT-4o can surpass these baselines for extended textual content."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.126,
+ 0.226,
+ 0.869,
+ 0.255
+ ],
+ "angle": 0,
+ "content": "We choose POSTA [12], Gemini 2.0 Flash [99], Ideogram 3.0 [2], and Playground-v3 [64] as the baselines because of their established capabilities in rendering longer texts. The results are shown in Figure 6 and Figure 7."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.127,
+ 0.261,
+ 0.547,
+ 0.275
+ ],
+ "angle": 0,
+ "content": "From these examples, we make the following key observations:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.286,
+ 0.865,
+ 0.34
+ ],
+ "angle": 0,
+ "content": "- GPT-4o's strength in long text generation: Compared with other baselines, GPT-4o demonstrates a superior ability to generate long, coherent text. In example 1 and example 3, GPT-4o produces detailed textual information with fewer than three characters generated incorrectly across more than 100 characters of text."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.172,
+ 0.346,
+ 0.868,
+ 0.386
+ ],
+ "angle": 0,
+ "content": "- Baseline limitations: When the input prompt becomes extremely long, models such as Gemini 2.0 Flash, Ideogram 3.0, and Playground-v3 often produce significantly more errors or produce vague text patches that are difficult to recognize."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.172,
+ 0.392,
+ 0.867,
+ 0.433
+ ],
+ "angle": 0,
+ "content": "- POSTA's performance: As a model specifically designed for poster-style text generation, POSTA performs closely to, or in some instances slightly more precisely than, GPT-4o. We hypothesize this is due to its multi-step pipeline tailored for long text rendering."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.171,
+ 0.286,
+ 0.868,
+ 0.433
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.127,
+ 0.446,
+ 0.87,
+ 0.475
+ ],
+ "angle": 0,
+ "content": "Overall, we conclude that GPT-4o excels at long text rendering, offering overwhelming performance compared to most existing commercial models, and delivering results on par with the latest specialized research models."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.936,
+ 0.51,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "10"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.164,
+ 0.115,
+ 0.336,
+ 0.131
+ ],
+ "angle": 0,
+ "content": "Short Text Rendering"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.34,
+ 0.142,
+ 0.361,
+ 0.159
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.363,
+ 0.146,
+ 0.658,
+ 0.161
+ ],
+ "angle": 0,
+ "content": "Evaluation: Text Rendering Precision."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.159,
+ 0.164,
+ 0.297,
+ 0.272
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.316,
+ 0.165,
+ 0.453,
+ 0.272
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.471,
+ 0.165,
+ 0.611,
+ 0.272
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.629,
+ 0.165,
+ 0.839,
+ 0.272
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.165,
+ 0.28,
+ 0.819,
+ 0.325
+ ],
+ "angle": 0,
+ "content": "Input Text: \"A beautiful painting of flowing colors and styles forming the words 'The GPT-4o/Ideogram/FLUX/SD3 research paper is nowhere!'. the background is speckled with drops and splashes of paint.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.159,
+ 0.332,
+ 0.252,
+ 0.439
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.276,
+ 0.332,
+ 0.415,
+ 0.439
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.439,
+ 0.332,
+ 0.576,
+ 0.439
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.601,
+ 0.332,
+ 0.841,
+ 0.439
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.167,
+ 0.448,
+ 0.782,
+ 0.478
+ ],
+ "angle": 0,
+ "content": "Input Text: \"Beautiful pixel art of a Wizard with hovering text 'Achievement unlocked: Diffusion models can spell now'.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.159,
+ 0.487,
+ 0.253,
+ 0.593
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.283,
+ 0.487,
+ 0.503,
+ 0.593
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.534,
+ 0.487,
+ 0.671,
+ 0.593
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.703,
+ 0.487,
+ 0.84,
+ 0.593
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.167,
+ 0.604,
+ 0.785,
+ 0.619
+ ],
+ "angle": 0,
+ "content": "Input Text: \"A monkey holding a sign reading 'Scaling transformer models is awesome'.\""
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.159,
+ 0.632,
+ 0.296,
+ 0.739
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.306,
+ 0.632,
+ 0.444,
+ 0.739
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.453,
+ 0.632,
+ 0.591,
+ 0.739
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.602,
+ 0.633,
+ 0.838,
+ 0.739
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.166,
+ 0.745,
+ 0.805,
+ 0.774
+ ],
+ "angle": 0,
+ "content": "Input Text: \"A surreal and humorous scene in a classroom with the words 'GPUs go brrrrr' written in white chalk on a blackboard. In front of the blackboard.\""
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.201,
+ 0.785,
+ 0.253,
+ 0.798
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.327,
+ 0.785,
+ 0.423,
+ 0.8
+ ],
+ "angle": 0,
+ "content": "Ideogram 3.0"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.501,
+ 0.785,
+ 0.543,
+ 0.798
+ ],
+ "angle": 0,
+ "content": "FLUX"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.702,
+ 0.785,
+ 0.738,
+ 0.798
+ ],
+ "angle": 0,
+ "content": "SD3"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.127,
+ 0.823,
+ 0.87,
+ 0.893
+ ],
+ "angle": 0,
+ "content": "Figure 5: Task: Short text rendering. Generate prompt-aligned, concise textual content (typically within 10 words) on an image. Setup: Each sample is produced based on a guiding text prompt. Comparisons are made with prior SOTA models [27, 2] and FLUX [51]. Observations: GPT-4o achieves performance on par with existing SOTA baselines in rendering short texts, consistently following the prompt with minimal errors. All evaluated methods—except FLUX [51]—deliver high-fidelity results in this setting."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.936,
+ 0.507,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "11"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.208,
+ 0.101,
+ 0.361,
+ 0.117
+ ],
+ "angle": 0,
+ "content": "Long Text Rendering"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.375,
+ 0.128,
+ 0.644,
+ 0.142
+ ],
+ "angle": 0,
+ "content": "Evaluation: Text Rendering Precision."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.183,
+ 0.147,
+ 0.321,
+ 0.307
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.228,
+ 0.312,
+ 0.277,
+ 0.324
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.322,
+ 0.147,
+ 0.464,
+ 0.307
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.37,
+ 0.312,
+ 0.42,
+ 0.324
+ ],
+ "angle": 0,
+ "content": "POSTA"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.465,
+ 0.147,
+ 0.605,
+ 0.307
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.48,
+ 0.312,
+ 0.588,
+ 0.324
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.606,
+ 0.147,
+ 0.813,
+ 0.307
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.663,
+ 0.312,
+ 0.753,
+ 0.326
+ ],
+ "angle": 0,
+ "content": "Ideogram 3.0"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.189,
+ 0.328,
+ 0.271,
+ 0.34
+ ],
+ "angle": 0,
+ "content": "Input Text:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.188,
+ 0.341,
+ 0.81,
+ 0.367
+ ],
+ "angle": 0,
+ "content": "\"Generate a movie poster with a sci-fi space theme, a solitary figure standing on an alien planet, facing a massive outpost."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.189,
+ 0.367,
+ 0.442,
+ 0.38
+ ],
+ "angle": 0,
+ "content": "The poster displays the following text:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.189,
+ 0.38,
+ 0.346,
+ 0.392
+ ],
+ "angle": 0,
+ "content": "Title: The Last Outpost"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.189,
+ 0.392,
+ 0.487,
+ 0.405
+ ],
+ "angle": 0,
+ "content": "Subtitle: When the stars fall, the truth rises"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.189,
+ 0.406,
+ 0.277,
+ 0.417
+ ],
+ "angle": 0,
+ "content": "Information:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.189,
+ 0.418,
+ 0.368,
+ 0.431
+ ],
+ "angle": 0,
+ "content": "Produced by Jackson Ward"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.189,
+ 0.432,
+ 0.338,
+ 0.444
+ ],
+ "angle": 0,
+ "content": "Music by Aria Calloway"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.189,
+ 0.444,
+ 0.371,
+ 0.457
+ ],
+ "angle": 0,
+ "content": "Screenplay by Elena Sharpe"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.189,
+ 0.457,
+ 0.371,
+ 0.47
+ ],
+ "angle": 0,
+ "content": "Directed By Sylvia Hartman"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.188,
+ 0.47,
+ 0.797,
+ 0.52
+ ],
+ "angle": 0,
+ "content": "\"A visually stunning and narratively gripping exploration of the unknown. The Last Outpost masterfully blends elements of science fiction, mystery, and psychological thriller, creating a hauntingly atmospheric journey that will leave audiences on the edge of their seats.\" -- Global Film Review\"."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.182,
+ 0.525,
+ 0.318,
+ 0.683
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.228,
+ 0.691,
+ 0.276,
+ 0.703
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.32,
+ 0.525,
+ 0.462,
+ 0.684
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.37,
+ 0.691,
+ 0.42,
+ 0.703
+ ],
+ "angle": 0,
+ "content": "POSTA"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.462,
+ 0.524,
+ 0.617,
+ 0.684
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.48,
+ 0.691,
+ 0.588,
+ 0.703
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.618,
+ 0.525,
+ 0.824,
+ 0.684
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.663,
+ 0.691,
+ 0.753,
+ 0.704
+ ],
+ "angle": 0,
+ "content": "Ideogram 3.0"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.185,
+ 0.707,
+ 0.266,
+ 0.719
+ ],
+ "angle": 0,
+ "content": "Input Text:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.184,
+ 0.72,
+ 0.802,
+ 0.771
+ ],
+ "angle": 0,
+ "content": "\"Create a poster with the theme of a Journey of Solitude. The background should depict a lone figure walking toward an unusable form of transportation. The scene should evoke a sense of being lost, helplessness, and desolation, capturing the emotional weight of losing oneself in a barren, unforgiving landscape."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.185,
+ 0.771,
+ 0.343,
+ 0.785
+ ],
+ "angle": 0,
+ "content": "Title: Solitary Journeys"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.185,
+ 0.785,
+ 0.318,
+ 0.796
+ ],
+ "angle": 0,
+ "content": "Subtitle: Elara Voss"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.185,
+ 0.797,
+ 0.565,
+ 0.809
+ ],
+ "angle": 0,
+ "content": "Information: WANDERING THROUGH THE UNKNOWN\"."
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.127,
+ 0.83,
+ 0.872,
+ 0.928
+ ],
+ "angle": 0,
+ "content": "Figure 6: Task: Long text rendering. Generate extended, coherent, and prompt-consistent textual content on an image. Setup: Evaluations are conducted against advanced baselines including POSTA [12], Gemini 2.0 Flash [99], Ideogram 3.0 [2], and Playground-v3 [64]. Observations: GPT-4o excels in long text rendering by producing coherent, detailed textual information with very few character errors. In contrast, models like Gemini 2.0 Flash, Ideogram 3.0, and Playground-v3 often exhibit increased errors or generate vague text when faced with lengthy prompts, while POSTA's tailored multi-step pipeline sometimes yields competitive precision. Overall, GPT-4o outperforms most commercial models and rivals specialized research approaches in extended text generation."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.936,
+ 0.51,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "12"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.201,
+ 0.292,
+ 0.357,
+ 0.309
+ ],
+ "angle": 0,
+ "content": "Long Text Rendering"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.347,
+ 0.317,
+ 0.368,
+ 0.334
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.368,
+ 0.321,
+ 0.648,
+ 0.336
+ ],
+ "angle": 0,
+ "content": "Evaluation: Text Rendering Precision."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.167,
+ 0.34,
+ 0.308,
+ 0.505
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.212,
+ 0.511,
+ 0.263,
+ 0.523
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.31,
+ 0.34,
+ 0.455,
+ 0.504
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.356,
+ 0.511,
+ 0.408,
+ 0.523
+ ],
+ "angle": 0,
+ "content": "POSTA"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.455,
+ 0.34,
+ 0.616,
+ 0.504
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.48,
+ 0.511,
+ 0.591,
+ 0.524
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.617,
+ 0.34,
+ 0.829,
+ 0.505
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.684,
+ 0.511,
+ 0.78,
+ 0.525
+ ],
+ "angle": 0,
+ "content": "Playground-v3"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.176,
+ 0.538,
+ 0.261,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "Input Text:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.552,
+ 0.793,
+ 0.592
+ ],
+ "angle": 0,
+ "content": "\"Please generate an artistic and stylized promotional poster. The style is an artistic painting style. The theme is about nature and city. The poster displays the following information: Title: Fragmented Harmony"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.592,
+ 0.536,
+ 0.605
+ ],
+ "angle": 0,
+ "content": "Subtitle Between the steel and sky, life finds its way."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.174,
+ 0.605,
+ 0.815,
+ 0.672
+ ],
+ "angle": 0,
+ "content": "Information: Amid the towering strucions and the quiet persistence of nature, a delicate balance emerges. The complex and often contradictory relationship between urban development and the natural world reveals itself in fleeting moments of harmony. Though fragmented, life continues, threading its way through the shadows of progress. Here, conflict and coexistence form an intricate dance--sometimes at odds, sometimes in unexpected unity\"."
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.196,
+ 0.7,
+ 0.799,
+ 0.715
+ ],
+ "angle": 0,
+ "content": "Figure 7: Task: Long text rendering. The Setup and Observations are the same as Figure 6."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.936,
+ 0.508,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "13"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.128,
+ 0.092,
+ 0.339,
+ 0.106
+ ],
+ "angle": 0,
+ "content": "2.1.3 Document Generation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.126,
+ 0.115,
+ 0.872,
+ 0.158
+ ],
+ "angle": 0,
+ "content": "We also explore a novel task: document image generation with GPT-4o, comparing its performance with Gemini 2.0 Flash [99] and Playground-v3 [64]. As shown in Figure 8 - 10, GPT-4o produces document images with cleaner layouts and more consistent content."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.171,
+ 0.177,
+ 0.393,
+ 0.193
+ ],
+ "angle": 0,
+ "content": "Document Image Generation"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.336,
+ 0.204,
+ 0.357,
+ 0.221
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.358,
+ 0.208,
+ 0.651,
+ 0.223
+ ],
+ "angle": 0,
+ "content": "Evaluation: Text Rendering Precision."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.189,
+ 0.256,
+ 0.333,
+ 0.267
+ ],
+ "angle": 0,
+ "content": "Attention Is All You Need"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.186,
+ 0.277,
+ 0.341,
+ 0.298
+ ],
+ "angle": 0,
+ "content": "Ashish Vaswani Noam Shazeer Niki Parmar Jakob Uszkoreit Lilion Jones Aidan N.Gomez Lukasz Kaiser IIIa Polosukhin"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.245,
+ 0.311,
+ 0.279,
+ 0.319
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.321,
+ 0.369,
+ 0.415
+ ],
+ "angle": 0,
+ "content": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on tw machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.405,
+ 0.324,
+ 0.503,
+ 0.331
+ ],
+ "angle": 0,
+ "content": "Attention Is All You Need"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.398,
+ 0.331,
+ 0.568,
+ 0.416
+ ],
+ "angle": 0,
+ "content": "Ashish Vourakis, NM Nazarov, NJI Parmar, Jbab Udekoreli, Lillion Jones, Adiinil S. W. Coomce, Lakota Kiber, Pala Poslashov, and M. A. D. G. R. Smith. 2017. The neural network models are based on recurrent or conventional neural networks in an encoder-decoder loss. The best produced models also connect the encoder and decoder loss through an attention mechanism. As, we propose a new study mechanism for a machine, the Transformer, based attention mechanisms, dispensing with the same weight as the encoder-decoder loss. The results of the training tasks show these models to be superior in quality while being more parsibilized and requiring significantly less time to time to move train. Our model achieves 84.2 BLUE in the WMT-50 Go-to-translation task, which is comparable to the performance of our previous work [3]. In addition, our WMT-30 Pragmatic task, our model established a new single-model state-of-the-art-state-of-the-art BLUE score of 41.8 when \\(\\alpha = 5\\) training for 3.5 days on GPUs after fraction of the training costs of the best models from literature. We propose Transformer generalizes well by applying it successfully to English syntacticity parsing both with large and limited training data."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.603,
+ 0.229,
+ 0.706,
+ 0.237
+ ],
+ "angle": 0,
+ "content": "Attention Is All You Need"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.603,
+ 0.242,
+ 0.693,
+ 0.27
+ ],
+ "angle": 0,
+ "content": "Ashish A. [Al] You. Wea a nono' Ainon 1 \nAshish Yawani, Naqadzaree, laokotri \nAlok Uzzotner, Anokalika Sanik, Jokoslav adar, Gosak III Ploosukhaini \nAlok Uzzotner, Anokalika Sanik"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.603,
+ 0.281,
+ 0.83,
+ 0.418
+ ],
+ "angle": 0,
+ "content": "Antext, exotocic sequecra transoedscnncs on cbr be caes baccared on bracococcyne nort bell netiabon an ecocycclion, an ecocyclon. TeTrane: the ensonnnmnsn neeepnckian. Epipcnie rile kceely on meenctiny As adeterdiencr. \nnpopioors sonr tonarwamchim. I mtnr. vorti. inenpoea a dedusum minnyss onomcrh. cortordone.lora ontata tose or uin hoperiosper. \nThe rorner is s oovl dtt maive de acemnccnodkdu aleu cormunb-dlr bing dvl-ndr 016de and mechance \n11 Dae nucnccnng \nceso nucnse \n12 - eannnr er on attonne amehnes asnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnnccnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnn"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.234,
+ 0.427,
+ 0.292,
+ 0.44
+ ],
+ "angle": 0,
+ "content": "GPT40"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.418,
+ 0.428,
+ 0.546,
+ 0.442
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.664,
+ 0.426,
+ 0.773,
+ 0.443
+ ],
+ "angle": 0,
+ "content": "Playground-v3"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.163,
+ 0.449,
+ 0.26,
+ 0.464
+ ],
+ "angle": 0,
+ "content": "Input Text:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.466,
+ 0.797,
+ 0.5
+ ],
+ "angle": 0,
+ "content": "\"Generate A realistic screenshot of the first page of the Paper from the following information:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.504,
+ 0.412,
+ 0.519
+ ],
+ "angle": 0,
+ "content": "Title: Attention Is All You Need"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.523,
+ 0.834,
+ 0.557
+ ],
+ "angle": 0,
+ "content": "Author List: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.561,
+ 0.832,
+ 0.825
+ ],
+ "angle": 0,
+ "content": "Abstract: The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.\""
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.126,
+ 0.843,
+ 0.871,
+ 0.884
+ ],
+ "angle": 0,
+ "content": "Figure 8: Task: Document image generation. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Playground-v3 [64]. Observation: GPT-4o can generate more consistent and accurate font and format than the other two models."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.936,
+ 0.509,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "14"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.171,
+ 0.151,
+ 0.392,
+ 0.168
+ ],
+ "angle": 0,
+ "content": "Document Image Generation"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.336,
+ 0.177,
+ 0.355,
+ 0.195
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.357,
+ 0.181,
+ 0.651,
+ 0.198
+ ],
+ "angle": 0,
+ "content": "Evaluation: Text Rendering Precision."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.162,
+ 0.225,
+ 0.38,
+ 0.251
+ ],
+ "angle": 0,
+ "content": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.173,
+ 0.261,
+ 0.335,
+ 0.278
+ ],
+ "angle": 0,
+ "content": "Jacob Devlin Ming-Wei Chang Kenton Lee Kristina Toutanova"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.252,
+ 0.291,
+ 0.288,
+ 0.3
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.302,
+ 0.384,
+ 0.367
+ ],
+ "angle": 0,
+ "content": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result; the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.367,
+ 0.388,
+ 0.404
+ ],
+ "angle": 0,
+ "content": "BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to \\(80.5\\%\\) (7.7% point absolute improvement), MultiNLI accuracy to \\(\\mathcal{S}6.7\\%\\) (4.6% absolute improvement), SQAud v1.1 question answering Test F1 to 93.2 (1.5 point absolute"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.242,
+ 0.41,
+ 0.3,
+ 0.424
+ ],
+ "angle": 0,
+ "content": "GPT 40"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.418,
+ 0.226,
+ 0.584,
+ 0.248
+ ],
+ "angle": 0,
+ "content": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.421,
+ 0.251,
+ 0.58,
+ 0.266
+ ],
+ "angle": 0,
+ "content": "Ashlor Jacob Doslin Ming Wei Chang, Kenton Lee Abstrut Win, Karinlin Touranosa"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.418,
+ 0.266,
+ 0.589,
+ 0.321
+ ],
+ "angle": 0,
+ "content": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Cryostatnet as Transformer. Unlike recnec mean, BERT is designed by pre-train deep bidirectional representations from nonshaded text by jasily selfconforming on both text 1st&4, xavier coint eorect, tate 1st&2 to win 1st, sacp0-twincent BERT model can be fine mused with just one additional output layer to create of dafve for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.418,
+ 0.328,
+ 0.59,
+ 0.376
+ ],
+ "angle": 0,
+ "content": "BERT is conceptually simple and empirically powerful. It obtains now state-of-the-art results on eleven evertweaven multimoment language processing tasks, including GLUE score to \\(80.5\\%\\) (7.69 absolute improvement), pushing the GLUE alloselect improvement). MulaNLI accuracy to \\(\\varepsilon_{\\mathrm{M}} = 0.1\\), \\(\\nu_{\\mathrm{L}} < \\varepsilon_{\\mathrm{M}} / \\varepsilon_{\\mathrm{M}}\\) (1.5 point absolute improvement) and SQoAD v2.0 Test F1 to 83.1 (3.1 point absolute improvement)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.439,
+ 0.411,
+ 0.567,
+ 0.425
+ ],
+ "angle": 0,
+ "content": "Gemini 2.0 Flash"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.621,
+ 0.23,
+ 0.824,
+ 0.248
+ ],
+ "angle": 0,
+ "content": "BERT: Pre-training on Lepi Bidellar Tansson Translons for Language Understond mderance litting' cetting from t cowf henvaming"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.251,
+ 0.653,
+ 0.256
+ ],
+ "angle": 0,
+ "content": "Author, List:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.257,
+ 0.821,
+ 0.269
+ ],
+ "angle": 0,
+ "content": "J Asbad Devlin, Yiw Changuaguagaa Kionn age rspgectangane cans \n\\(^{a}\\)pressin liKistcn-Toutanfa"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.269,
+ 0.689,
+ 0.275
+ ],
+ "angle": 0,
+ "content": "represeons-Uintanlvania"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.621,
+ 0.251,
+ 0.821,
+ 0.275
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.276,
+ 0.648,
+ 0.283
+ ],
+ "angle": 0,
+ "content": "A"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.287,
+ 0.836,
+ 0.329
+ ],
+ "angle": 0,
+ "content": "We introduce a new languagegretrovetercendentiale monoclin klonstionist monole conBldfecarstaadss reprenters from nemer raje Sfiflnonanecones. desessnissrall ranauagaleafdelfe xyn unming on hnlaesaeare two ploddes also the por-entant canteletory a state-vwraon one-on-one coffice of anisotropy, and the ploso-syntropic colective of states. Of s2s212310 pain or questionbmporansuansluangene Tcnf?f to ingest ingf Sf10 tto 46 w. (I, test),marshersonalizne imnance immmens"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.339,
+ 0.801,
+ 0.346
+ ],
+ "angle": 0,
+ "content": "BERT is conceptually imlilienenarplenpholcft-nu surate-fine-ams"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.347,
+ 0.824,
+ 0.365
+ ],
+ "angle": 0,
+ "content": "ronen 1 oonranaalwauanu viipopluoteforocnirandinns inget caught anourage, vovlurvulgina for nain. 2004. The use of the word \"sulfur\" in the text is a question envoing SUIF u697."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.365,
+ 0.837,
+ 0.379
+ ],
+ "angle": 0,
+ "content": "- GBFEscors/aanoreqquasurf and Squad w.10 aninlvalte 83.7% 4.6% (X) \n- TeST fto onop:11.3% (x)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.621,
+ 0.383,
+ 0.836,
+ 0.403
+ ],
+ "angle": 0,
+ "content": "→ BERT is cponconuynlyrsnaintally pocefine-at-ut: ouvah176. (JET v.c.37% quinting anguen linyuH-aCLS sccorts onoonssthe sonea 4000A/AresoVc LEAU pioiHcB: gnrmaeh an epesrourinans A7c)0v o.o35 aed 1170"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.621,
+ 0.339,
+ 0.837,
+ 0.403
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.677,
+ 0.41,
+ 0.786,
+ 0.427
+ ],
+ "angle": 0,
+ "content": "Playground-v3"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.162,
+ 0.441,
+ 0.259,
+ 0.456
+ ],
+ "angle": 0,
+ "content": "Input Text:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.458,
+ 0.79,
+ 0.492
+ ],
+ "angle": 0,
+ "content": "Generate A realistic screenshot of the first page of the Paper from the following information:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.496,
+ 0.735,
+ 0.533
+ ],
+ "angle": 0,
+ "content": "Title: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.535,
+ 0.836,
+ 0.702
+ ],
+ "angle": 0,
+ "content": "Author List: Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova \nAbstract: We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.16,
+ 0.705,
+ 0.836,
+ 0.798
+ ],
+ "angle": 0,
+ "content": "BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to \\(80.5\\%\\) (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.178,
+ 0.838,
+ 0.818,
+ 0.854
+ ],
+ "angle": 0,
+ "content": "Figure 9: Task: Document image generation. The Setup and Observations are the same as Fig. 8."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.936,
+ 0.508,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "15"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.171,
+ 0.128,
+ 0.391,
+ 0.144
+ ],
+ "angle": 0,
+ "content": "Document Image Generation"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.336,
+ 0.154,
+ 0.355,
+ 0.173
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.358,
+ 0.158,
+ 0.651,
+ 0.175
+ ],
+ "angle": 0,
+ "content": "Evaluation: Text Rendering Precision."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.171,
+ 0.213,
+ 0.346,
+ 0.25
+ ],
+ "angle": 0,
+ "content": "You Only Look Once: Unified, Real-Time Object Detection"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.259,
+ 0.361,
+ 0.275
+ ],
+ "angle": 0,
+ "content": "Joseph Redmon, Santosh Divvala, Ross Girshick Ali Farhadi"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.251,
+ 0.286,
+ 0.285,
+ 0.293
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.168,
+ 0.294,
+ 0.376,
+ 0.347
+ ],
+ "angle": 0,
+ "content": "We present YOLO, a new approach to object detection. Prior work on object detection repurposes classifiers to perform detection. Instead, we frame object detection as a regression problem to spatially separated bounding boxes and associated class probabilities. A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation. Since the whole detection pipeline is a single network, it can be optimized end-to-end directly on detection performance."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.168,
+ 0.347,
+ 0.377,
+ 0.388
+ ],
+ "angle": 0,
+ "content": "Our unified architecture is extremely fast. Our base YOLO model processes images in real-time at 45 frames per second. A smaller version of the network. Fast YOLO, processes an astounding 155 frames per second while still achieving double the mAP of other real-time detectors. Compared to state-of-the-art detection systems YOLO makes more localization errors but is far less likely to pred"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.416,
+ 0.216,
+ 0.581,
+ 0.234
+ ],
+ "angle": 0,
+ "content": "You Only Look Once: Unlimited, Time/Indirect Decciption \nAuthor:Lesser Jowesh Redmonial Siptedthi Adri Farhad"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.415,
+ 0.236,
+ 0.624,
+ 0.26
+ ],
+ "angle": 0,
+ "content": "We present YJLO, a new approach to object detection. Driver search uses oleejeon cieeuses to represen or correct deferential finwication. Istegedel, when farstial, we confirmed anapase, agate trius to a signafal or sialipal staphylococcal preprobed boing boos summarilyour heartbodn frontal nater and the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.415,
+ 0.261,
+ 0.498,
+ 0.266
+ ],
+ "angle": 0,
+ "content": "Acatat Rechun-Yorim Bogae lo rojctc filocly"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.415,
+ 0.269,
+ 0.623,
+ 0.296
+ ],
+ "angle": 0,
+ "content": "Our undersonged tandoce boe + ant - or cemoe - aepocemis in a times, bodingly narmabogus haarban. Jnci enwecstic ancoontie fluy fast YOLO pue moebes, 45f rans perennetioles unperenentio- dorensis in reel thane reobexes petarges. ectcylcfom oene trsoute of sucrose princeia of docuta. d. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.418,
+ 0.299,
+ 0.629,
+ 0.345
+ ],
+ "angle": 0,
+ "content": "| Case Figure | Meta-task | Sub-task | GPT-4o | Gemini-2.0-flash | Domain-SOTA |
| Figure 1 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 2 | | Complex Text Following | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 3 | | | Success | Success | Success |
| Figure 4 | | | Success | Success | Success |
| Figure 5 | | | Success | Success | Success |
| Figure 6 | Text-to-Image | Text Rendering | Success | Low Visual Quality | Low Visual Quality |
| Figure 7 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 8 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 9 | | Document Generation | Success | Low Visual Quality | Low Visual Quality |
| Figure 10 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 11 | | Panorama | Lack of Knowledge | Success | Success |
| Figure 12 | | Style Transfer | Success | Lack of Knowledge | Lack of Knowledge |
| Figure 13 | | | Success | Lack of Knowledge | Lack of Knowledge |
| Figure 14 | | | Low Visual Quality | Success | Failure to Follow Instructions |
| Figure 15 | | Image Editing | Failure to Follow Instructions | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 16 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 17 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 18 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 19 | | | Success | Inconsistent Generation | Failure to Follow Instructions |
| Figure 20 | | Single-Concept Customization | Success | Failure to Follow Instructions | Success |
| Figure 21 | | Multi-Concept Customization | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 22 | | Story Image Generation | Success | Failure to Follow Instructions | Success |
| Figure 23 | | | Success | Inconsistent Generation | Success |
| Figure 24 | | Low-Level Vision-Denoising | Low Visual Quality | Low Visual Quality | Success |
| Figure 25 | | Low-Level Vision-Deraining | Success | Inconsistent Generation | Success |
| Figure 26 | | Low-Level Vision-Dehazing | Success | Low Visual Quality | Success |
| Figure 27 | | Low-Level Vision-Low Light Enhancement | Low Visual Quality | Low Visual Quality | Success |
| Figure 28 | | Low-Level Vision-Deblurring | Success | Low Visual Quality | Success |
| Figure 29 | | Low-Level Vision-Super Resolution | Success | Low Visual Quality | Success |
| Figure 30 | | Low-Level Vision-Impainting | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 31 | | Low-Level Vision-Outpainting | Inconsistent Generation | Success | Success |
| Figure 32 | | Low-Level Vision-Colorization | Success | Success | Success |
| Figure 33 | | Low-Level Vision-Shadow Removal | Success | Failure to Follow Instructions | Success |
| Figure 34 | | Low-Level Vision-Reflection Removal | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 35 | | Low-Level Vision-Relighting | Success | Failure to Follow Instructions | Success |
| Figure 36 | | Spatial Control-Canny | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 37 | | Spatial Control-Depth | Success | Failure to Follow Instructions | Success |
| Figure 38 | | Spatial Control-Sketch | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 39 | | Spatial Control-Pose | Success | Inconsistent Generation | Success |
| Figure 40 | | Spatial Control-Mask | Inconsistent Generation | Failure to Follow Instructions | Inconsistent Generation |
| Figure 41 | | Camera Control | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 42 | | | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 43 | | In-Context Visual Prompting | Failure to Follow Instructions | Failure to Follow Instructions | N/A |
| Figure 44 | | Image to 3D Modeling | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 45 | | UV Map to 3D Rendering | Success | Inconsistent Generation | Failure to Follow Instructions |
| Figure 46 | | Novel View Synthesis | Success | Success | Failure to Follow Instructions |
| Figure 47 | | Image Segmentation | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 48 | | | Success | Failure to Follow Instructions | Success |
| Figure 49 | | | Success | Failure to Follow Instructions | Success |
| Figure 50 | | Edge Detection | Success | Success | Success |
| Figure 51 | | | Success | Failure to Follow Instructions | Success |
| Figure 52 | | | Success | Failure to Follow Instructions | Success |
| Figure 53 | | Salient Object | Success | Failure to Follow Instructions | Success |
| Figure 54 | | | Success | Success | Success |
| Figure 55 | | | Success | Success | Success |
| Figure 56 | | Depth Estimation | Success | Failure to Follow Instructions | Success |
| Figure 57 | | Normal Estimation | Success | Failure to Follow Instructions | Success |
| Figure 58 | | Layout Detection | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 59 | | Text Detection | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 60 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 61 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 62 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 63 | | | Inconsistent Generation | Inconsistent Generation | Success |
+
+# 2.1 Text-to-Image Tasks
+
+# 2.1.1 Complex Text Following Capability
+
+Recent progress in text-to-image generation has shown impressive abilities in generating diverse and realistic images based on text prompts. However, composing multiple objects with various attributes and relationships accurately into one scene remains a significant challenge for current text-to-image generative models [92, 85, 8, 81, 6]. In this section, we assess models' ability for compositional text-to-image generation from four perspectives following [41], which include attribute binding, numeracy, object relationship, and complex compositions. Attribute binding evaluates whether the model correctly assigns attributes, such as color, shape, and texture to the appropriate objects. Numeracy evaluates whether the number of generated objects matches the quantities specified in the prompt. Object relationships refer to both spatial (2D/3D) and non-spatial interactions among objects. Complex compositions evaluate the model's ability to handle multiple types of constraints simultaneously, especially given long or detailed prompts.
+
+As shown in Figure 1 row 1, GPT-4o outperforms both Gemini 2.0 Flash and Midjourney in numeracy tasks. While GPT-4o accurately represents a single plate, Gemini 2.0 and Midjourney represent two plates instead. In terms of understanding object relationships, GPT-4o is the only model that correctly infers the action "walk towards" from the ragdoll to the labrador. However, GPT-4o struggles with more complex terms like "pentagonal pyramid", failing to interpret it correctly (see Figure 1 row 4). This suggests that GPT-4o may have difficulty accurately interpreting objects with unusual geometries. When it comes to abstract prompts, GPT-4o also appears to lack imagination (see Figure 2 row 2), whereas Midjourney v6.1 demonstrates better creativity in this case, outperforming both GPT-4o and Gemini 2.0 Flash.
+
+For complex text-to-image generation, we evaluate GPT-4o's performance with Gemini 2.0 Flash [99] and FLUX.1-Pro [51], using the text prompts collected from [124, 106, 115]. As shown in Figure 3, both GPT-4o and FLUX excel at generating realistic and harmonious scenes align with the text prompts. However, we observe that GPT-4o shows limitations in generating culturally related elements. For example, the generated crown for the Chinese general is western-style rather than chinese-style (see Figure 4 row 2). Additionally, in large scene generation, GPT-4o struggles to maintain boundary continuity, whereas FLUX produces a more natural composition (see Figure 4 row 3).
+
+Overall, we conclude that GPT-4o excels at text-to-image generation in terms of attribute binding, generative numeracy, object relationship, and complex compositions. However, it exhibits limitations in generating uncommon objects, culturally specific elements and in maintaining continuity when composing large scenes.
+
+# Text-to-Image Generation
+
+
+
+# Evaluation: Visual content precisely following the text instruction.
+
+
+
+
+
+
+
+
+
+
+Input Text: "A yellow bowl, a blue mug and a pink plate on the table."
+Input Text: "A ragdoll walks towards a labrador."
+
+
+
+
+Input Text: "Three differently colored apples (yellow, green, red from left to right) with a Coca-Cola bottle placed behind the middle apple."
+
+
+
+
+
+
+Input Text: "The oval sphere was nestled between the rectangular prism and the pentagonal pyramid."
+Figure 1: Task: Compositional text-to-image generation. Evaluate the image-text alignment on attribute binding, numeracy, and object relationship. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Midjourney v6.1 [75]. Observation: GPT-4o outperforms Gemini 2.0 Flash and Midjourney v6.1 across all aspects. However, GPT-4o struggles with uncommon objects with a special geometry.
+
+
+
+
+
+GPT 40
+
+Gemini 2.0 Flash
+
+Midjourney v6.1
+
+
+
+# Evaluation: Visual content precisely following the text instruction.
+
+
+Input Text: "The round, juicy watermelon sat in the cool, refreshing bowl of ice, waiting to be sliced open and devoured."
+
+
+
+
+
+
+
+
+
+
+
+
+Input Text: "The bold, expressive strokes of the artist's brush brought the blank canvas to life, forming a vibrant and dynamic masterpiece."
+
+
+
+
+
+
+Input Text: "The heavy raindrops fell on the smooth glass and the textured roof."
+
+
+Input Text: "The gentle, soothing melody of the piano filled the concert hall, as the pianist's fingers danced over the keys."
+Figure 2: Task: Compositional text-to-image generation. Evaluate the image-text alignment on attribute binding and complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Midjourney v6.1 [75]. Observation: GPT-4o outperforms the other two models in generating objects aligned with the text prompts accurately. But for more abstract and creative tasks, Midjourney v6.1 performs the best.
+
+
+
+GPT 40
+
+Gemini 2.0 Flash
+
+Midjourney v6.1
+
+# Text-to-Image Generation (with complex text prompt)
+
+# Evaluation: Visual content precisely following the text instruction.
+
+
+Input Text: "An icy landscape. A vast expanse of snow-covered mountain peaks stretches endlessly. Beneath them is a dense forest and a colossal frozen lake. Three people are boating in three boats separately in the lake. Not far from the lake, a volcano threatens eruption, its rumblings felt even from afar. Above, a ferocious red dragon dominates the sky and commands the heavens, fueled by the volcano's relentless energy flow." (Prompt from GenArtist)
+
+
+
+
+
+
+
+
+
+
+
+
+Input Text: "In a magical seascape, a majestic ship sails through crystal blue waters surrounded by vibrant marine life and soaring birds. Towering cliffs frame the scene, while a stunning rainbow arches across the sky, blending with ethereal clouds. This enchanting journey captures the serene beauty of nature's wonders." (Prompt from IterComp)
+
+
+Input Text: "On the rooftop of a skyscraper in a bustling cyberpunk city, a figure in a trench coat and neon-lit visor stands amidst a garden of bio-luminescent plants, overlooking the maze of flying cars and towering holograms. Robotic birds flit among the foliage, digital billboards flash advertisements in the distance." (Prompt from IterComp)
+
+
+Figure 3: Task: Compositional text-to-image generation. Evaluate the image-text alignment on complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and FLUX.1-Pro [51]. Observation: GPT-4o and FLUX can generate more harmonious and natural scene than Gemini 2.0 Flash.
+
+GPT 40
+
+Gemini 2.0 Flash
+
+FLUX
+
+# Text-to-Image Generation (with complex text prompt)
+
+
+
+# Evaluation: Visual content precisely following the text instruction.
+
+
+Input Text: "Under the luminous full moon, a serene Japanese garden with traditional pagodas and a tranquil pond creates a magical night scene. The soft glow from the lantern-lit buildings reflects on the water, blending nature and architecture in harmony. The moonlight bathes the landscape, enhancing the peaceful ambiance." (Prompt from IterComp)
+
+
+
+
+
+
+Input Text: "A Chinese general wearing a crown, with whiskers and golden Chinese style armor, standing with a majestic dragon head on his chest, symbolizing his strength, wearing black and gold boots. His appearance exudes a sense of authority, wisdom, and an unyielding spirit, embodying the ideal ancient Chinese hero." (Prompt from RPG)
+
+
+
+
+
+
+Input Text: "A beautiful landscape with a river in the middle, the left of the river is in the evening and in the winter with a big iceberg and a small village while some people are skiing on the river and some people are skating, the right of the river is in the summer with a volcano in the morning and a small village while some people are playing." (Prompt from RPG)
+
+
+Figure 4: Task: Compositional text-to-image generation. Evaluate the image-text alignment on complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and FLUX.1-Pro [51]. Observation: GPT-4o struggles to generate culturally related elements and maintain boundary continuity (see rows 2 and 3), similar to Gemini 2.0 Flash and FLUX.
+
+
+
+
+
+GPT 40
+
+Gemini 2.0 Flash
+
+FLUX
+
+# 2.1.2 Text Rendering
+
+Text rendering is a task that aims at generating texts (characters, sentences, or even paragraphs) on an image. The text content is usually guided by the input prompt. Previous models [27, 2] show good capability in generating short text (within 10 words, such as signs or short phrases), but their ability to generate long texts remains limited.
+
+As shown in Figure 5, GPT-4o demonstrates comparable abilities to existing state-of-the-art (SOTA) baselines when generating short texts. All the methods except FLUX [51] perform well at rendering short text following the prompt. In this section, we primarily focus on long text rendering to examine whether GPT-4o can surpass these baselines for extended textual content.
+
+We choose POSTA [12], Gemini 2.0 Flash [99], Ideogram 3.0 [2], and Playground-v3 [64] as the baselines because of their established capabilities in rendering longer texts. The results are shown in Figure 6 and Figure 7.
+
+From these examples, we make the following key observations:
+
+- GPT-4o's strength in long text generation: Compared with other baselines, GPT-4o demonstrates a superior ability to generate long, coherent text. In example 1 and example 3, GPT-4o produces detailed textual information with fewer than three characters generated incorrectly across more than 100 characters of text.
+- Baseline limitations: When the input prompt becomes extremely long, models such as Gemini 2.0 Flash, Ideogram 3.0, and Playground-v3 often produce significantly more errors or produce vague text patches that are difficult to recognize.
+- POSTA's performance: As a model specifically designed for poster-style text generation, POSTA performs closely to, or in some instances slightly more precisely than, GPT-4o. We hypothesize this is due to its multi-step pipeline tailored for long text rendering.
+
+Overall, we conclude that GPT-4o excels at long text rendering, offering overwhelming performance compared to most existing commercial models, and delivering results on par with the latest specialized research models.
+
+# Short Text Rendering
+
+
+
+# Evaluation: Text Rendering Precision.
+
+
+
+
+
+
+
+
+
+Input Text: "A beautiful painting of flowing colors and styles forming the words 'The GPT-4o/Ideogram/FLUX/SD3 research paper is nowhere!'. the background is speckled with drops and splashes of paint."
+
+
+
+
+
+
+
+
+
+Input Text: "Beautiful pixel art of a Wizard with hovering text 'Achievement unlocked: Diffusion models can spell now'."
+
+
+Figure 5: Task: Short text rendering. Generate prompt-aligned, concise textual content (typically within 10 words) on an image. Setup: Each sample is produced based on a guiding text prompt. Comparisons are made with prior SOTA models [27, 2] and FLUX [51]. Observations: GPT-4o achieves performance on par with existing SOTA baselines in rendering short texts, consistently following the prompt with minimal errors. All evaluated methods—except FLUX [51]—deliver high-fidelity results in this setting.
+
+
+
+
+
+
+
+Input Text: "A monkey holding a sign reading 'Scaling transformer models is awesome'."
+
+
+
+
+
+
+
+
+
+Input Text: "A surreal and humorous scene in a classroom with the words 'GPUs go brrrrr' written in white chalk on a blackboard. In front of the blackboard."
+
+GPT 40
+
+Ideogram 3.0
+
+FLUX
+
+SD3
+
+# Evaluation: Text Rendering Precision.
+
+
+GPT 40
+
+
+POSTA
+
+
+Gemini 2.0 Flash
+
+
+Ideogram 3.0
+
+# Input Text:
+
+"Generate a movie poster with a sci-fi space theme, a solitary figure standing on an alien planet, facing a massive outpost.
+
+The poster displays the following text:
+
+Title: The Last Outpost
+
+Subtitle: When the stars fall, the truth rises
+
+# Information:
+
+Produced by Jackson Ward
+
+Music by Aria Calloway
+
+Screenplay by Elena Sharpe
+
+Directed By Sylvia Hartman
+
+"A visually stunning and narratively gripping exploration of the unknown. The Last Outpost masterfully blends elements of science fiction, mystery, and psychological thriller, creating a hauntingly atmospheric journey that will leave audiences on the edge of their seats." -- Global Film Review".
+
+
+GPT 40
+
+
+POSTA
+
+
+Gemini 2.0 Flash
+
+
+Ideogram 3.0
+Figure 6: Task: Long text rendering. Generate extended, coherent, and prompt-consistent textual content on an image. Setup: Evaluations are conducted against advanced baselines including POSTA [12], Gemini 2.0 Flash [99], Ideogram 3.0 [2], and Playground-v3 [64]. Observations: GPT-4o excels in long text rendering by producing coherent, detailed textual information with very few character errors. In contrast, models like Gemini 2.0 Flash, Ideogram 3.0, and Playground-v3 often exhibit increased errors or generate vague text when faced with lengthy prompts, while POSTA's tailored multi-step pipeline sometimes yields competitive precision. Overall, GPT-4o outperforms most commercial models and rivals specialized research approaches in extended text generation.
+
+# Input Text:
+
+"Create a poster with the theme of a Journey of Solitude. The background should depict a lone figure walking toward an unusable form of transportation. The scene should evoke a sense of being lost, helplessness, and desolation, capturing the emotional weight of losing oneself in a barren, unforgiving landscape.
+
+Title: Solitary Journeys
+
+Subtitle: Elara Voss
+
+Information: WANDERING THROUGH THE UNKNOWN".
+
+# Long Text Rendering
+
+
+
+# Evaluation: Text Rendering Precision.
+
+
+GPT 40
+Figure 7: Task: Long text rendering. The Setup and Observations are the same as Figure 6.
+
+
+POSTA
+
+
+Gemini 2.0 Flash
+
+
+Playground-v3
+
+# Input Text:
+
+"Please generate an artistic and stylized promotional poster. The style is an artistic painting style. The theme is about nature and city. The poster displays the following information: Title: Fragmented Harmony
+
+Subtitle Between the steel and sky, life finds its way.
+
+Information: Amid the towering strucions and the quiet persistence of nature, a delicate balance emerges. The complex and often contradictory relationship between urban development and the natural world reveals itself in fleeting moments of harmony. Though fragmented, life continues, threading its way through the shadows of progress. Here, conflict and coexistence form an intricate dance--sometimes at odds, sometimes in unexpected unity".
+
+# 2.1.3 Document Generation
+
+We also explore a novel task: document image generation with GPT-4o, comparing its performance with Gemini 2.0 Flash [99] and Playground-v3 [64]. As shown in Figure 8 - 10, GPT-4o produces document images with cleaner layouts and more consistent content.
+
+# Document Image Generation
+
+
+
+# Evaluation: Text Rendering Precision.
+
+# Attention Is All You Need
+
+Ashish Vaswani Noam Shazeer Niki Parmar Jakob Uszkoreit Lilion Jones Aidan N.Gomez Lukasz Kaiser IIIa Polosukhin
+
+# Abstract
+
+The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on tw machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature.
+
+# Attention Is All You Need
+
+Ashish Vourakis, NM Nazarov, NJI Parmar, Jbab Udekoreli, Lillion Jones, Adiinil S. W. Coomce, Lakota Kiber, Pala Poslashov, and M. A. D. G. R. Smith. 2017. The neural network models are based on recurrent or conventional neural networks in an encoder-decoder loss. The best produced models also connect the encoder and decoder loss through an attention mechanism. As, we propose a new study mechanism for a machine, the Transformer, based attention mechanisms, dispensing with the same weight as the encoder-decoder loss. The results of the training tasks show these models to be superior in quality while being more parsibilized and requiring significantly less time to time to move train. Our model achieves 84.2 BLUE in the WMT-50 Go-to-translation task, which is comparable to the performance of our previous work [3]. In addition, our WMT-30 Pragmatic task, our model established a new single-model state-of-the-art-state-of-the-art BLUE score of 41.8 when $\alpha = 5$ training for 3.5 days on GPUs after fraction of the training costs of the best models from literature. We propose Transformer generalizes well by applying it successfully to English syntacticity parsing both with large and limited training data.
+
+# Attention Is All You Need
+
+Ashish A. [Al] You. Wea a nono' Ainon 1
+Ashish Yawani, Naqadzaree, laokotri
+Alok Uzzotner, Anokalika Sanik, Jokoslav adar, Gosak III Ploosukhaini
+Alok Uzzotner, Anokalika Sanik
+
+Antext, exotocic sequecra transoedscnncs on cbr be caes baccared on bracococcyne nort bell netiabon an ecocycclion, an ecocyclon. TeTrane: the ensonnnmnsn neeepnckian. Epipcnie rile kceely on meenctiny As adeterdiencr.
+npopioors sonr tonarwamchim. I mtnr. vorti. inenpoea a dedusum minnyss onomcrh. cortordone.lora ontata tose or uin hoperiosper.
+The rorner is s oovl dtt maive de acemnccnodkdu aleu cormunb-dlr bing dvl-ndr 016de and mechance
+11 Dae nucnccnng
+ceso nucnse
+12 - eannnr er on attonne amehnes asnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnnccnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnn
+
+GPT40
+
+Gemini 2.0 Flash
+
+Playground-v3
+
+# Input Text:
+
+"Generate A realistic screenshot of the first page of the Paper from the following information:
+
+Title: Attention Is All You Need
+
+Author List: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin
+
+Abstract: The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data."
+
+Figure 8: Task: Document image generation. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Playground-v3 [64]. Observation: GPT-4o can generate more consistent and accurate font and format than the other two models.
+
+# Document Image Generation
+
+
+
+# Evaluation: Text Rendering Precision.
+
+# BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
+
+Jacob Devlin Ming-Wei Chang Kenton Lee Kristina Toutanova
+
+# Abstract
+
+We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result; the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.
+
+BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to $80.5\%$ (7.7% point absolute improvement), MultiNLI accuracy to $\mathcal{S}6.7\%$ (4.6% absolute improvement), SQAud v1.1 question answering Test F1 to 93.2 (1.5 point absolute
+
+GPT 40
+
+# BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
+
+Ashlor Jacob Doslin Ming Wei Chang, Kenton Lee Abstrut Win, Karinlin Touranosa
+
+We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Cryostatnet as Transformer. Unlike recnec mean, BERT is designed by pre-train deep bidirectional representations from nonshaded text by jasily selfconforming on both text 1st&4, xavier coint eorect, tate 1st&2 to win 1st, sacp0-twincent BERT model can be fine mused with just one additional output layer to create of dafve for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.
+
+BERT is conceptually simple and empirically powerful. It obtains now state-of-the-art results on eleven evertweaven multimoment language processing tasks, including GLUE score to $80.5\%$ (7.69 absolute improvement), pushing the GLUE alloselect improvement). MulaNLI accuracy to $\varepsilon_{\mathrm{M}} = 0.1$ , $\nu_{\mathrm{L}} < \varepsilon_{\mathrm{M}} / \varepsilon_{\mathrm{M}}$ (1.5 point absolute improvement) and SQoAD v2.0 Test F1 to 83.1 (3.1 point absolute improvement).
+
+Gemini 2.0 Flash
+
+# BERT: Pre-training on Lepi Bidellar Tansson Translons for Language Understond mderance litting' cetting from t cowf henvaming
+
+Author, List:
+J Asbad Devlin, Yiw Changuaguagaa Kionn age rspgectangane cans
+ $^{a}$ pressin liKistcn-Toutanfa
+represeons-Uintanlvania
+
+A
+
+We introduce a new languagegretrovetercendentiale monoclin klonstionist monole conBldfecarstaadss reprenters from nemer raje Sfiflnonanecones. desessnissrall ranauagaleafdelfe xyn unming on hnlaesaeare two ploddes also the por-entant canteletory a state-vwraon one-on-one coffice of anisotropy, and the ploso-syntropic colective of states. Of s2s212310 pain or questionbmporansuansluangene Tcnf?f to ingest ingf Sf10 tto 46 w. (I, test),marshersonalizne imnance immmens
+
+BERT is conceptually imlilienenarplenpholcft-nu surate-fine-ams
+ronen 1 oonranaalwauanu viipopluoteforocnirandinns inget caught anourage, vovlurvulgina for nain. 2004. The use of the word "sulfur" in the text is a question envoing SUIF u697.
+- GBFEscors/aanoreqquasurf and Squad w.10 aninlvalte 83.7% 4.6% (X)
+- TeST fto onop:11.3% (x)
+→ BERT is cponconuynlyrsnaintally pocefine-at-ut: ouvah176. (JET v.c.37% quinting anguen linyuH-aCLS sccorts onoonssthe sonea 4000A/AresoVc LEAU pioiHcB: gnrmaeh an epesrourinans A7c)0v o.o35 aed 1170
+
+Playground-v3
+
+# Input Text:
+
+Generate A realistic screenshot of the first page of the Paper from the following information:
+
+Title: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
+
+Author List: Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova
+Abstract: We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.
+
+BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to $80.5\%$ (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).
+
+Figure 9: Task: Document image generation. The Setup and Observations are the same as Fig. 8.
+
+
+
+# Evaluation: Text Rendering Precision.
+
+# You Only Look Once: Unified, Real-Time Object Detection
+
+Joseph Redmon, Santosh Divvala, Ross Girshick Ali Farhadi
+
+# Abstract
+
+We present YOLO, a new approach to object detection. Prior work on object detection repurposes classifiers to perform detection. Instead, we frame object detection as a regression problem to spatially separated bounding boxes and associated class probabilities. A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation. Since the whole detection pipeline is a single network, it can be optimized end-to-end directly on detection performance.
+
+Our unified architecture is extremely fast. Our base YOLO model processes images in real-time at 45 frames per second. A smaller version of the network. Fast YOLO, processes an astounding 155 frames per second while still achieving double the mAP of other real-time detectors. Compared to state-of-the-art detection systems YOLO makes more localization errors but is far less likely to pred
+
+You Only Look Once: Unlimited, Time/Indirect Decciption
+Author:Lesser Jowesh Redmonial Siptedthi Adri Farhad
+
+We present YJLO, a new approach to object detection. Driver search uses oleejeon cieeuses to represen or correct deferential finwication. Istegedel, when farstial, we confirmed anapase, agate trius to a signafal or sialipal staphylococcal preprobed boing boos summarilyour heartbodn frontal nater and the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the
+
+Acatat Rechun-Yorim Bogae lo rojctc filocly
+
+Our undersonged tandoce boe + ant - or cemoe - aepocemis in a times, bodingly narmabogus haarban. Jnci enwecstic ancoontie fluy fast YOLO pue moebes, 45f rans perennetioles unperenentio- dorensis in reel thane reobexes petarges. ectcylcfom oene trsoute of sucrose princeia of docuta. d. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
+
+| Case Figure | Meta-task | Sub-task | GPT-4o | Gemini-2.0-flash | Domain-SOTA |
| Figure 1 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 2 | | Complex Text Following | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 3 | | | Success | Success | Success |
| Figure 4 | | | Success | Success | Success |
| Figure 5 | | | Success | Success | Success |
| Figure 6 | Text-to-Image | Text Rendering | Success | Low Visual Quality | Low Visual Quality |
| Figure 7 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 8 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 9 | | Document Generation | Success | Low Visual Quality | Low Visual Quality |
| Figure 10 | | | Success | Low Visual Quality | Low Visual Quality |
| Figure 11 | | Panorama | Lack of Knowledge | Success | Success |
| Figure 12 | | Style Transfer | Success | Lack of Knowledge | Lack of Knowledge |
| Figure 13 | | | Success | Lack of Knowledge | Lack of Knowledge |
| Figure 14 | | | Low Visual Quality | Success | Failure to Follow Instructions |
| Figure 15 | | Image Editing | Failure to Follow Instructions | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 16 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 17 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 18 | | | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 19 | | | Success | Inconsistent Generation | Failure to Follow Instructions |
| Figure 20 | | Single-Concept Customization | Success | Failure to Follow Instructions | Success |
| Figure 21 | | Multi-Concept Customization | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 22 | | Story Image Generation | Success | Failure to Follow Instructions | Success |
| Figure 23 | | | Success | Inconsistent Generation | Success |
| Figure 24 | | Low-Level Vision-Denoising | Low Visual Quality | Low Visual Quality | Success |
| Figure 25 | | Low-Level Vision-Deraining | Success | Inconsistent Generation | Success |
| Figure 26 | | Low-Level Vision-Dehazing | Success | Low Visual Quality | Success |
| Figure 27 | | Low-Level Vision-Low Light Enhancement | Low Visual Quality | Low Visual Quality | Success |
| Figure 28 | | Low-Level Vision-Deblurring | Success | Low Visual Quality | Success |
| Figure 29 | | Low-Level Vision-Super Resolution | Success | Low Visual Quality | Success |
| Figure 30 | | Low-Level Vision-Impainting | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 31 | | Low-Level Vision-Outpainting | Inconsistent Generation | Success | Success |
| Figure 32 | | Low-Level Vision-Colorization | Success | Success | Success |
| Figure 33 | | Low-Level Vision-Shadow Removal | Success | Failure to Follow Instructions | Success |
| Figure 34 | | Low-Level Vision-Reflection Removal | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 35 | | Low-Level Vision-Relighting | Success | Failure to Follow Instructions | Success |
| Figure 36 | | Spatial Control-Canny | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 37 | | Spatial Control-Depth | Success | Failure to Follow Instructions | Success |
| Figure 38 | | Spatial Control-Sketch | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 39 | | Spatial Control-Pose | Success | Inconsistent Generation | Success |
| Figure 40 | | Spatial Control-Mask | Inconsistent Generation | Failure to Follow Instructions | Inconsistent Generation |
| Figure 41 | | Camera Control | Inconsistent Generation | Failure to Follow Instructions | Success |
| Figure 42 | | | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 43 | | In-Context Visual Prompting | Failure to Follow Instructions | Failure to Follow Instructions | N/A |
| Figure 44 | | Image to 3D Modeling | Success | Failure to Follow Instructions | Failure to Follow Instructions |
| Figure 45 | | UV Map to 3D Rendering | Success | Inconsistent Generation | Failure to Follow Instructions |
| Figure 46 | | Novel View Synthesis | Success | Success | Failure to Follow Instructions |
| Figure 47 | | Image Segmentation | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 48 | | | Success | Failure to Follow Instructions | Success |
| Figure 49 | | | Success | Failure to Follow Instructions | Success |
| Figure 50 | | Edge Detection | Success | Success | Success |
| Figure 51 | | | Success | Failure to Follow Instructions | Success |
| Figure 52 | | | Success | Failure to Follow Instructions | Success |
| Figure 53 | | Salient Object | Success | Failure to Follow Instructions | Success |
| Figure 54 | | | Success | Success | Success |
| Figure 55 | | | Success | Success | Success |
| Figure 56 | | Depth Estimation | Success | Failure to Follow Instructions | Success |
| Figure 57 | | Normal Estimation | Success | Failure to Follow Instructions | Success |
| Figure 58 | | Layout Detection | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 59 | | Text Detection | Failure to Follow Instructions | Failure to Follow Instructions | Success |
| Figure 60 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 61 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 62 | | | Inconsistent Generation | Inconsistent Generation | Success |
| Figure 63 | | | Inconsistent Generation | Inconsistent Generation | Success |
",
+ "image_path": "69fc7b667b5296c731fc241b01cedfd07a9c8e3fb9b10cdf2db89fc2a34aef2f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "text",
+ "content": "4"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 190,
+ 83
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 190,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 190,
+ 83
+ ],
+ "type": "text",
+ "content": "2.1 Text-to-Image Tasks"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 78,
+ 92,
+ 262,
+ 105
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 92,
+ 262,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 92,
+ 262,
+ 105
+ ],
+ "type": "text",
+ "content": "2.1.1 Complex Text Following Capability"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 76,
+ 111,
+ 533,
+ 223
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 76,
+ 111,
+ 533,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 76,
+ 111,
+ 533,
+ 223
+ ],
+ "type": "text",
+ "content": "Recent progress in text-to-image generation has shown impressive abilities in generating diverse and realistic images based on text prompts. However, composing multiple objects with various attributes and relationships accurately into one scene remains a significant challenge for current text-to-image generative models [92, 85, 8, 81, 6]. In this section, we assess models' ability for compositional text-to-image generation from four perspectives following [41], which include attribute binding, numeracy, object relationship, and complex compositions. Attribute binding evaluates whether the model correctly assigns attributes, such as color, shape, and texture to the appropriate objects. Numeracy evaluates whether the number of generated objects matches the quantities specified in the prompt. Object relationships refer to both spatial (2D/3D) and non-spatial interactions among objects. Complex compositions evaluate the model's ability to handle multiple types of constraints simultaneously, especially given long or detailed prompts."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 76,
+ 225,
+ 533,
+ 315
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 76,
+ 225,
+ 533,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 76,
+ 225,
+ 533,
+ 315
+ ],
+ "type": "text",
+ "content": "As shown in Figure 1 row 1, GPT-4o outperforms both Gemini 2.0 Flash and Midjourney in numeracy tasks. While GPT-4o accurately represents a single plate, Gemini 2.0 and Midjourney represent two plates instead. In terms of understanding object relationships, GPT-4o is the only model that correctly infers the action \"walk towards\" from the ragdoll to the labrador. However, GPT-4o struggles with more complex terms like \"pentagonal pyramid\", failing to interpret it correctly (see Figure 1 row 4). This suggests that GPT-4o may have difficulty accurately interpreting objects with unusual geometries. When it comes to abstract prompts, GPT-4o also appears to lack imagination (see Figure 2 row 2), whereas Midjourney v6.1 demonstrates better creativity in this case, outperforming both GPT-4o and Gemini 2.0 Flash."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 76,
+ 318,
+ 533,
+ 397
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 76,
+ 318,
+ 533,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 76,
+ 318,
+ 533,
+ 397
+ ],
+ "type": "text",
+ "content": "For complex text-to-image generation, we evaluate GPT-4o's performance with Gemini 2.0 Flash [99] and FLUX.1-Pro [51], using the text prompts collected from [124, 106, 115]. As shown in Figure 3, both GPT-4o and FLUX excel at generating realistic and harmonious scenes align with the text prompts. However, we observe that GPT-4o shows limitations in generating culturally related elements. For example, the generated crown for the Chinese general is western-style rather than chinese-style (see Figure 4 row 2). Additionally, in large scene generation, GPT-4o struggles to maintain boundary continuity, whereas FLUX produces a more natural composition (see Figure 4 row 3)."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 76,
+ 400,
+ 533,
+ 436
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 76,
+ 400,
+ 533,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 76,
+ 400,
+ 533,
+ 436
+ ],
+ "type": "text",
+ "content": "Overall, we conclude that GPT-4o excels at text-to-image generation in terms of attribute binding, generative numeracy, object relationship, and complex compositions. However, it exhibits limitations in generating uncommon objects, culturally specific elements and in maintaining continuity when composing large scenes."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "text",
+ "content": "5"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 117,
+ 79,
+ 240,
+ 92
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 79,
+ 240,
+ 92
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 79,
+ 240,
+ 92
+ ],
+ "type": "text",
+ "content": "Text-to-Image Generation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 154,
+ 100,
+ 165,
+ 114
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 154,
+ 100,
+ 165,
+ 114
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 154,
+ 100,
+ 165,
+ 114
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 154,
+ 100,
+ 165,
+ 114
+ ],
+ "type": "image",
+ "image_path": "5e416a9762e2def779608eb1bf6eee5a9cee49745a628b45ccc17073ed461705.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 166,
+ 102,
+ 473,
+ 115
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 102,
+ 473,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 102,
+ 473,
+ 115
+ ],
+ "type": "text",
+ "content": "Evaluation: Visual content precisely following the text instruction."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 116,
+ 222,
+ 232
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 222,
+ 232
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 222,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 116,
+ 222,
+ 232
+ ],
+ "type": "image",
+ "image_path": "b56cffa92dcc8965f109102c1a0975dedfc665539ea8cf6c71d5a2879dfe8dcc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 241,
+ 116,
+ 359,
+ 234
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 241,
+ 116,
+ 359,
+ 234
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 241,
+ 116,
+ 359,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 241,
+ 116,
+ 359,
+ 234
+ ],
+ "type": "image",
+ "image_path": "e6d59fa9484b3877034764c9fa8752e51c78a7b4eba2a3dd224cdb9236f23368.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 376,
+ 116,
+ 504,
+ 234
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 376,
+ 116,
+ 504,
+ 234
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 376,
+ 116,
+ 504,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 376,
+ 116,
+ 504,
+ 234
+ ],
+ "type": "image",
+ "image_path": "114bf170640575674f20a5f89dd3e1c35830102990e7d45e9fe0e42bf870b7d0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 248,
+ 224,
+ 365
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 224,
+ 365
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 224,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 224,
+ 365
+ ],
+ "type": "image",
+ "image_path": "e21293e1dde8f2f8a8736c0a4d673afa08874092081ea4b1ec07788bbe177b05.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 241,
+ 247,
+ 361,
+ 368
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 107,
+ 236,
+ 373,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 236,
+ 373,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 236,
+ 373,
+ 247
+ ],
+ "type": "text",
+ "content": "Input Text: \"A yellow bowl, a blue mug and a pink plate on the table.\""
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 241,
+ 247,
+ 361,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 241,
+ 247,
+ 361,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 241,
+ 247,
+ 361,
+ 368
+ ],
+ "type": "image",
+ "image_path": "cde20841ace4621a197135b25fda935e422854fc542a786ca15f9aea2eb46486.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 109,
+ 368,
+ 299,
+ 378
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 368,
+ 299,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 368,
+ 299,
+ 378
+ ],
+ "type": "text",
+ "content": "Input Text: \"A ragdoll walks towards a labrador.\""
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 376,
+ 246,
+ 503,
+ 368
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 376,
+ 246,
+ 503,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 376,
+ 246,
+ 503,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 376,
+ 246,
+ 503,
+ 368
+ ],
+ "type": "image",
+ "image_path": "d0a18282d6404ef0ab645148e73d17fe445982eb8c40cb3c0d5a9081efd6d52c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 380,
+ 227,
+ 497
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 227,
+ 497
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 227,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 380,
+ 227,
+ 497
+ ],
+ "type": "image",
+ "image_path": "3bffa6f0cb5545336d1dece3985a3fcaff2ad60da5c231db47b90005c63f4ceb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 106,
+ 498,
+ 496,
+ 519
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 496,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 496,
+ 519
+ ],
+ "type": "text",
+ "content": "Input Text: \"Three differently colored apples (yellow, green, red from left to right) with a Coca-Cola bottle placed behind the middle apple.\""
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 242,
+ 380,
+ 361,
+ 497
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 242,
+ 380,
+ 361,
+ 497
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 242,
+ 380,
+ 361,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 242,
+ 380,
+ 361,
+ 497
+ ],
+ "type": "image",
+ "image_path": "f333d581ad570f753e99a7e73c99d827c89d746ae6f1f549c53cf30cf7e7ea33.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 376,
+ 379,
+ 503,
+ 497
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 376,
+ 379,
+ 503,
+ 497
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 376,
+ 379,
+ 503,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 376,
+ 379,
+ 503,
+ 497
+ ],
+ "type": "image",
+ "image_path": "107a161e8c64a21e6ae887d967ee6aa6bf1c940aafacc290e429d8deb9516754.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 521,
+ 226,
+ 639
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 226,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 226,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 226,
+ 639
+ ],
+ "type": "image",
+ "image_path": "00b435bd1d8026aa4f0d08be8a9579479f625012cae88f49f81a5d958cf5c199.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 501,
+ 652
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 501,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 501,
+ 652
+ ],
+ "type": "text",
+ "content": "Input Text: \"The oval sphere was nestled between the rectangular prism and the pentagonal pyramid.\""
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 77,
+ 687,
+ 533,
+ 733
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 687,
+ 533,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 687,
+ 533,
+ 733
+ ],
+ "type": "text",
+ "content": "Figure 1: Task: Compositional text-to-image generation. Evaluate the image-text alignment on attribute binding, numeracy, and object relationship. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Midjourney v6.1 [75]. Observation: GPT-4o outperforms Gemini 2.0 Flash and Midjourney v6.1 across all aspects. However, GPT-4o struggles with uncommon objects with a special geometry."
+ }
+ ]
+ }
+ ],
+ "index": 22,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 241,
+ 521,
+ 361,
+ 639
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 241,
+ 521,
+ 361,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 241,
+ 521,
+ 361,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 241,
+ 521,
+ 361,
+ 639
+ ],
+ "type": "image",
+ "image_path": "b1e74b98b6b45c314db43ad44efec736de0526ea53933c41d6f494dfc222ec5c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 376,
+ 521,
+ 503,
+ 639
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 376,
+ 521,
+ 503,
+ 639
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 376,
+ 521,
+ 503,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 376,
+ 521,
+ 503,
+ 639
+ ],
+ "type": "image",
+ "image_path": "ec488303c09dbf9dfb728fd1952368c1c3fa85f708e3b290ecd1a36b50ef4803.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 17,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 144,
+ 657,
+ 176,
+ 667
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 144,
+ 657,
+ 176,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 657,
+ 176,
+ 667
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 267,
+ 658,
+ 337,
+ 669
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 267,
+ 658,
+ 337,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 267,
+ 658,
+ 337,
+ 669
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 407,
+ 657,
+ 474,
+ 669
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 407,
+ 657,
+ 474,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 407,
+ 657,
+ 474,
+ 669
+ ],
+ "type": "text",
+ "content": "Midjourney v6.1"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 302,
+ 742,
+ 309,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 742,
+ 309,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 742,
+ 309,
+ 750
+ ],
+ "type": "text",
+ "content": "6"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 5
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 155,
+ 101,
+ 165,
+ 114
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 155,
+ 101,
+ 165,
+ 114
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 155,
+ 101,
+ 165,
+ 114
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 155,
+ 101,
+ 165,
+ 114
+ ],
+ "type": "image",
+ "image_path": "41cf37136782a8224523753684cab11c0ddb36c12dfd39ee89ac851285853af2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 167,
+ 103,
+ 473,
+ 115
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 103,
+ 473,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 103,
+ 473,
+ 115
+ ],
+ "type": "text",
+ "content": "Evaluation: Visual content precisely following the text instruction."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 116,
+ 228,
+ 232
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 108,
+ 116,
+ 228,
+ 232
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 116,
+ 228,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 116,
+ 228,
+ 232
+ ],
+ "type": "image",
+ "image_path": "742bdab0de5d04587ee79a5c9d93f90cfb4494e5a48dcf258a1406c6265bd627.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 107,
+ 234,
+ 509,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 234,
+ 509,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 234,
+ 509,
+ 252
+ ],
+ "type": "text",
+ "content": "Input Text: \"The round, juicy watermelon sat in the cool, refreshing bowl of ice, waiting to be sliced open and devoured.\""
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 241,
+ 116,
+ 374,
+ 232
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 241,
+ 116,
+ 374,
+ 232
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 241,
+ 116,
+ 374,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 241,
+ 116,
+ 374,
+ 232
+ ],
+ "type": "image",
+ "image_path": "c73e3aa340f33cab4cf19354249e37fe1bd7f76aef6ec7fbcc154198f1dad05b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 388,
+ 116,
+ 504,
+ 232
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 388,
+ 116,
+ 504,
+ 232
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 388,
+ 116,
+ 504,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 388,
+ 116,
+ 504,
+ 232
+ ],
+ "type": "image",
+ "image_path": "382030f3e3f0298bca0a14e5cda55ee5859de8f05e70a6182873cc5b53987e77.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 253,
+ 228,
+ 372
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 108,
+ 253,
+ 228,
+ 372
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 253,
+ 228,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 253,
+ 228,
+ 372
+ ],
+ "type": "image",
+ "image_path": "9a87f98dd86139de0d38246d54b1d070591e2fdb526d3285409a473583f28ec6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 243,
+ 253,
+ 373,
+ 371
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 243,
+ 253,
+ 373,
+ 371
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 243,
+ 253,
+ 373,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 243,
+ 253,
+ 373,
+ 371
+ ],
+ "type": "image",
+ "image_path": "fce6b55e09781aa60141751625be4a55c7d0fd1eef584be53a6585c781eade33.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 387,
+ 253,
+ 506,
+ 372
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 387,
+ 253,
+ 506,
+ 372
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 387,
+ 253,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 253,
+ 506,
+ 372
+ ],
+ "type": "image",
+ "image_path": "0f7ad2208e211f067756484bb7a6847038cf13549a2f34c8a13b12eeda49aad7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 393,
+ 228,
+ 510
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 109,
+ 373,
+ 507,
+ 393
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 373,
+ 507,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 373,
+ 507,
+ 393
+ ],
+ "type": "text",
+ "content": "Input Text: \"The bold, expressive strokes of the artist's brush brought the blank canvas to life, forming a vibrant and dynamic masterpiece.\""
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 108,
+ 393,
+ 228,
+ 510
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 393,
+ 228,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 393,
+ 228,
+ 510
+ ],
+ "type": "image",
+ "image_path": "d87f5d38548a498be4acc2e6175a859f019a900cdcbd8394bd2700b4f145ed8f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 243,
+ 393,
+ 373,
+ 510
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 243,
+ 393,
+ 373,
+ 510
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 243,
+ 393,
+ 373,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 243,
+ 393,
+ 373,
+ 510
+ ],
+ "type": "image",
+ "image_path": "c87b7b6792c5fda9d5340959a376273a7f5d0b80db6770aea6ec9e15c6ec596d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 387,
+ 393,
+ 507,
+ 511
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 387,
+ 393,
+ 507,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 387,
+ 393,
+ 507,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 393,
+ 507,
+ 511
+ ],
+ "type": "image",
+ "image_path": "4d7cf29e3244126ed31062342f32c73674817be1d0079f2da561d4e75f3368e2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 525,
+ 228,
+ 643
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 111,
+ 513,
+ 429,
+ 523
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 513,
+ 429,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 513,
+ 429,
+ 523
+ ],
+ "type": "text",
+ "content": "Input Text: \"The heavy raindrops fell on the smooth glass and the textured roof.\""
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 108,
+ 525,
+ 228,
+ 643
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 525,
+ 228,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 525,
+ 228,
+ 643
+ ],
+ "type": "image",
+ "image_path": "c36a7bcd11820fd2f7d3c63fbf0e563dbdebdbcf2828507130fec8b86c2c1cc1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 243,
+ 525,
+ 374,
+ 644
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 243,
+ 525,
+ 374,
+ 644
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 243,
+ 525,
+ 374,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 243,
+ 525,
+ 374,
+ 644
+ ],
+ "type": "image",
+ "image_path": "0e3a0dc448dd3f289f1837b24241f0ae490d1d79799115f52e41475efc038437.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 108,
+ 645,
+ 494,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 645,
+ 494,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 645,
+ 494,
+ 666
+ ],
+ "type": "text",
+ "content": "Input Text: \"The gentle, soothing melody of the piano filled the concert hall, as the pianist's fingers danced over the keys.\""
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 77,
+ 691,
+ 532,
+ 746
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 691,
+ 532,
+ 746
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 691,
+ 532,
+ 746
+ ],
+ "type": "text",
+ "content": "Figure 2: Task: Compositional text-to-image generation. Evaluate the image-text alignment on attribute binding and complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Midjourney v6.1 [75]. Observation: GPT-4o outperforms the other two models in generating objects aligned with the text prompts accurately. But for more abstract and creative tasks, Midjourney v6.1 performs the best."
+ }
+ ]
+ }
+ ],
+ "index": 22,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 387,
+ 525,
+ 507,
+ 644
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 387,
+ 525,
+ 507,
+ 644
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 387,
+ 525,
+ 507,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 525,
+ 507,
+ 644
+ ],
+ "type": "image",
+ "image_path": "adcb856bb8a351cb62cc8d6fd9dcfe0f715a542653534f39e99b1973a7067ec8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 17,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 143,
+ 667,
+ 175,
+ 677
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 143,
+ 667,
+ 175,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 667,
+ 175,
+ 677
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 266,
+ 667,
+ 335,
+ 677
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 266,
+ 667,
+ 335,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 266,
+ 667,
+ 335,
+ 677
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 406,
+ 666,
+ 472,
+ 677
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 406,
+ 666,
+ 472,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 406,
+ 666,
+ 472,
+ 677
+ ],
+ "type": "text",
+ "content": "Midjourney v6.1"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 119,
+ 81,
+ 240,
+ 94
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 81,
+ 240,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 81,
+ 240,
+ 94
+ ],
+ "type": "text",
+ "content": "Text-to-Image Generation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "text",
+ "content": "7"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 107,
+ 85,
+ 241,
+ 109
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 85,
+ 241,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 85,
+ 241,
+ 109
+ ],
+ "type": "text",
+ "content": "Text-to-Image Generation (with complex text prompt)"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 137,
+ 118,
+ 482,
+ 131
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 137,
+ 118,
+ 482,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 137,
+ 118,
+ 482,
+ 131
+ ],
+ "type": "text",
+ "content": "Evaluation: Visual content precisely following the text instruction."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 133,
+ 220,
+ 247
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 133,
+ 220,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 133,
+ 220,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 133,
+ 220,
+ 247
+ ],
+ "type": "image",
+ "image_path": "27fce9400300fb184ed43e1e4b4f8eb042ae16f63646b6eda324b41fff54aede.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 100,
+ 251,
+ 504,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 100,
+ 251,
+ 504,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 100,
+ 251,
+ 504,
+ 304
+ ],
+ "type": "text",
+ "content": "Input Text: \"An icy landscape. A vast expanse of snow-covered mountain peaks stretches endlessly. Beneath them is a dense forest and a colossal frozen lake. Three people are boating in three boats separately in the lake. Not far from the lake, a volcano threatens eruption, its rumblings felt even from afar. Above, a ferocious red dragon dominates the sky and commands the heavens, fueled by the volcano's relentless energy flow.\" (Prompt from GenArtist)"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 239,
+ 133,
+ 354,
+ 247
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 239,
+ 133,
+ 354,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 239,
+ 133,
+ 354,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 239,
+ 133,
+ 354,
+ 247
+ ],
+ "type": "image",
+ "image_path": "116d28e49cd90395347ff38e9ec8e1c04b768757b89b7a9a1d0a0f0f317b20f4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 373,
+ 133,
+ 488,
+ 247
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 373,
+ 133,
+ 488,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 373,
+ 133,
+ 488,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 373,
+ 133,
+ 488,
+ 247
+ ],
+ "type": "image",
+ "image_path": "bfaf1c18b294719962fa4fcaa9024d109045581572e81ebad1ebe756236b696a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 309,
+ 220,
+ 422
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 220,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 220,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 220,
+ 422
+ ],
+ "type": "image",
+ "image_path": "64e6d603fb4bb84b5ebce439ab2fdb3c6e0e566e8af8982c1c50d5378b6fe487.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 239,
+ 308,
+ 354,
+ 422
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 239,
+ 308,
+ 354,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 239,
+ 308,
+ 354,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 239,
+ 308,
+ 354,
+ 422
+ ],
+ "type": "image",
+ "image_path": "7ed09698c0927be9038e1132fc7243c7e628f5e5c041880670c7d1ac190273ad.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 373,
+ 308,
+ 487,
+ 422
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 373,
+ 308,
+ 487,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 373,
+ 308,
+ 487,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 373,
+ 308,
+ 487,
+ 422
+ ],
+ "type": "image",
+ "image_path": "484e31e49c1e7a2ebd3c367ae0db8fcd10716584cec8c8e06beaca0f6c7a0381.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 475,
+ 220,
+ 588
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 220,
+ 588
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 220,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 220,
+ 588
+ ],
+ "type": "image",
+ "image_path": "2de6dd439640e1a491cd2669646ca4deab67272994343185f41956a82e6a40a1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 104,
+ 591,
+ 504,
+ 634
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 591,
+ 504,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 591,
+ 504,
+ 634
+ ],
+ "type": "text",
+ "content": "Input Text: \"In a magical seascape, a majestic ship sails through crystal blue waters surrounded by vibrant marine life and soaring birds. Towering cliffs frame the scene, while a stunning rainbow arches across the sky, blending with ethereal clouds. This enchanting journey captures the serene beauty of nature's wonders.\" (Prompt from IterComp)"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 239,
+ 474,
+ 354,
+ 589
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 104,
+ 428,
+ 504,
+ 472
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 428,
+ 504,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 428,
+ 504,
+ 472
+ ],
+ "type": "text",
+ "content": "Input Text: \"On the rooftop of a skyscraper in a bustling cyberpunk city, a figure in a trench coat and neon-lit visor stands amidst a garden of bio-luminescent plants, overlooking the maze of flying cars and towering holograms. Robotic birds flit among the foliage, digital billboards flash advertisements in the distance.\" (Prompt from IterComp)"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 239,
+ 474,
+ 354,
+ 589
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 239,
+ 474,
+ 354,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 239,
+ 474,
+ 354,
+ 589
+ ],
+ "type": "image",
+ "image_path": "0c0470a214490980a8b144c3e88a00976a14f32edb1762c91f3846a0c05ec5b2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 373,
+ 474,
+ 487,
+ 588
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 373,
+ 474,
+ 487,
+ 588
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 373,
+ 474,
+ 487,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 373,
+ 474,
+ 487,
+ 588
+ ],
+ "type": "image",
+ "image_path": "745ba50ac275ead363177aa6a6389523a01b4c236062fe89cb966d79aeafe69b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 77,
+ 665,
+ 533,
+ 709
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 665,
+ 533,
+ 709
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 665,
+ 533,
+ 709
+ ],
+ "type": "text",
+ "content": "Figure 3: Task: Compositional text-to-image generation. Evaluate the image-text alignment on complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and FLUX.1-Pro [51]. Observation: GPT-4o and FLUX can generate more harmonious and natural scene than Gemini 2.0 Flash."
+ }
+ ]
+ }
+ ],
+ "index": 17,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 143,
+ 639,
+ 178,
+ 650
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 143,
+ 639,
+ 178,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 639,
+ 178,
+ 650
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 256,
+ 639,
+ 333,
+ 650
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 256,
+ 639,
+ 333,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 256,
+ 639,
+ 333,
+ 650
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 414,
+ 639,
+ 443,
+ 650
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 414,
+ 639,
+ 443,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 414,
+ 639,
+ 443,
+ 650
+ ],
+ "type": "text",
+ "content": "FLUX"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "text",
+ "content": "8"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 7
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 103,
+ 95,
+ 237,
+ 119
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 103,
+ 95,
+ 237,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 95,
+ 237,
+ 119
+ ],
+ "type": "text",
+ "content": "Text-to-Image Generation (with complex text prompt)"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 121,
+ 125,
+ 133,
+ 139
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 121,
+ 125,
+ 133,
+ 139
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 125,
+ 133,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 125,
+ 133,
+ 139
+ ],
+ "type": "image",
+ "image_path": "30ed799a330abd36ec3b456792e50674b62123dda7d183a4909f198292186c06.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 136,
+ 129,
+ 482,
+ 141
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 136,
+ 129,
+ 482,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 136,
+ 129,
+ 482,
+ 141
+ ],
+ "type": "text",
+ "content": "Evaluation: Visual content precisely following the text instruction."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 144,
+ 219,
+ 258
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 219,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 219,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 219,
+ 258
+ ],
+ "type": "image",
+ "image_path": "0a19bafce54aa9b28620506b9c6051449ff283c024b7b8bdde3dea6959a5f2ab.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 95,
+ 262,
+ 497,
+ 305
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 95,
+ 262,
+ 497,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 95,
+ 262,
+ 497,
+ 305
+ ],
+ "type": "text",
+ "content": "Input Text: \"Under the luminous full moon, a serene Japanese garden with traditional pagodas and a tranquil pond creates a magical night scene. The soft glow from the lantern-lit buildings reflects on the water, blending nature and architecture in harmony. The moonlight bathes the landscape, enhancing the peaceful ambiance.\" (Prompt from IterComp)"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 241,
+ 145,
+ 354,
+ 258
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 241,
+ 145,
+ 354,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 241,
+ 145,
+ 354,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 241,
+ 145,
+ 354,
+ 258
+ ],
+ "type": "image",
+ "image_path": "1660f44ab6725c952d5b0bc43505f77c0d7cb4b552be04d6430ee2b25ec63d61.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 374,
+ 145,
+ 487,
+ 258
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 374,
+ 145,
+ 487,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 374,
+ 145,
+ 487,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 374,
+ 145,
+ 487,
+ 258
+ ],
+ "type": "image",
+ "image_path": "89e41678c8fcc68e299a5cf2931d67bb1f2339142bca0c32c1593b4888650519.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 121,
+ 312,
+ 204,
+ 434
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 121,
+ 312,
+ 204,
+ 434
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 312,
+ 204,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 312,
+ 204,
+ 434
+ ],
+ "type": "image",
+ "image_path": "f016fb55ce2f2b9f48a4d23f36970310500e1250abf50d5e88f1902507f424c6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 99,
+ 440,
+ 501,
+ 483
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 99,
+ 440,
+ 501,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 99,
+ 440,
+ 501,
+ 483
+ ],
+ "type": "text",
+ "content": "Input Text: \"A Chinese general wearing a crown, with whiskers and golden Chinese style armor, standing with a majestic dragon head on his chest, symbolizing his strength, wearing black and gold boots. His appearance exudes a sense of authority, wisdom, and an unyielding spirit, embodying the ideal ancient Chinese hero.\" (Prompt from RPG)"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 253,
+ 312,
+ 342,
+ 434
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 253,
+ 312,
+ 342,
+ 434
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 253,
+ 312,
+ 342,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 253,
+ 312,
+ 342,
+ 434
+ ],
+ "type": "image",
+ "image_path": "9077e20f899c60d5e1c26e1748dfa141980bf09da6c613467d42c1143bab34a1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 384,
+ 312,
+ 476,
+ 434
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 384,
+ 312,
+ 476,
+ 434
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 384,
+ 312,
+ 476,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 384,
+ 312,
+ 476,
+ 434
+ ],
+ "type": "image",
+ "image_path": "c341891a5dd74ee440ca81c75da11adba9b78fd29a19903df5c98dbac7513e19.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 93,
+ 489,
+ 219,
+ 574
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 93,
+ 489,
+ 219,
+ 574
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 93,
+ 489,
+ 219,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 93,
+ 489,
+ 219,
+ 574
+ ],
+ "type": "image",
+ "image_path": "4569221e0b0de61258bfb9ba425fbed4cda37201c120d48edf92be4e3bff1d5c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 99,
+ 581,
+ 501,
+ 624
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 99,
+ 581,
+ 501,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 99,
+ 581,
+ 501,
+ 624
+ ],
+ "type": "text",
+ "content": "Input Text: \"A beautiful landscape with a river in the middle, the left of the river is in the evening and in the winter with a big iceberg and a small village while some people are skiing on the river and some people are skating, the right of the river is in the summer with a volcano in the morning and a small village while some people are playing.\" (Prompt from RPG)"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 221,
+ 490,
+ 309,
+ 574
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 221,
+ 490,
+ 309,
+ 574
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 221,
+ 490,
+ 309,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 221,
+ 490,
+ 309,
+ 574
+ ],
+ "type": "image",
+ "image_path": "475d909f8bd94a4e962a1af7f117f1db9f76e1f2cd7047ccbdf158cec0cbc2b2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 77,
+ 654,
+ 533,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 654,
+ 533,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 654,
+ 533,
+ 698
+ ],
+ "type": "text",
+ "content": "Figure 4: Task: Compositional text-to-image generation. Evaluate the image-text alignment on complex compositions. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and FLUX.1-Pro [51]. Observation: GPT-4o struggles to generate culturally related elements and maintain boundary continuity (see rows 2 and 3), similar to Gemini 2.0 Flash and FLUX."
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 309,
+ 489,
+ 399,
+ 574
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 309,
+ 489,
+ 399,
+ 574
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 489,
+ 399,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 489,
+ 399,
+ 574
+ ],
+ "type": "image",
+ "image_path": "ef1d1bd4564e7e63c547c011879fd964bcadd7553b251ad9931c404cc35e496b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 401,
+ 489,
+ 516,
+ 574
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 401,
+ 489,
+ 516,
+ 574
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 401,
+ 489,
+ 516,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 401,
+ 489,
+ 516,
+ 574
+ ],
+ "type": "image",
+ "image_path": "1086a817de2fda2fa73bc2221b337c167aa823711f8d87db081972a8c108577d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 135,
+ 630,
+ 170,
+ 641
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 630,
+ 170,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 630,
+ 170,
+ 641
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 264,
+ 630,
+ 341,
+ 641
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 264,
+ 630,
+ 341,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 264,
+ 630,
+ 341,
+ 641
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 442,
+ 628,
+ 470,
+ 639
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 442,
+ 628,
+ 470,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 442,
+ 628,
+ 470,
+ 639
+ ],
+ "type": "text",
+ "content": "FLUX"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "type": "text",
+ "content": "9"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 8
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 179,
+ 85
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 179,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 179,
+ 85
+ ],
+ "type": "text",
+ "content": "2.1.2 Text Rendering"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 77,
+ 91,
+ 533,
+ 125
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 91,
+ 533,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 91,
+ 533,
+ 125
+ ],
+ "type": "text",
+ "content": "Text rendering is a task that aims at generating texts (characters, sentences, or even paragraphs) on an image. The text content is usually guided by the input prompt. Previous models [27, 2] show good capability in generating short text (within 10 words, such as signs or short phrases), but their ability to generate long texts remains limited."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 77,
+ 129,
+ 532,
+ 174
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 129,
+ 532,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 129,
+ 532,
+ 174
+ ],
+ "type": "text",
+ "content": "As shown in Figure 5, GPT-4o demonstrates comparable abilities to existing state-of-the-art (SOTA) baselines when generating short texts. All the methods except FLUX [51] perform well at rendering short text following the prompt. In this section, we primarily focus on long text rendering to examine whether GPT-4o can surpass these baselines for extended textual content."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 77,
+ 178,
+ 531,
+ 201
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 178,
+ 531,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 178,
+ 531,
+ 201
+ ],
+ "type": "text",
+ "content": "We choose POSTA [12], Gemini 2.0 Flash [99], Ideogram 3.0 [2], and Playground-v3 [64] as the baselines because of their established capabilities in rendering longer texts. The results are shown in Figure 6 and Figure 7."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 77,
+ 206,
+ 334,
+ 217
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 206,
+ 334,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 206,
+ 334,
+ 217
+ ],
+ "type": "text",
+ "content": "From these examples, we make the following key observations:"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 104,
+ 226,
+ 531,
+ 342
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 8,
+ "blocks": [
+ {
+ "bbox": [
+ 104,
+ 226,
+ 529,
+ 269
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 226,
+ 529,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 226,
+ 529,
+ 269
+ ],
+ "type": "text",
+ "content": "- GPT-4o's strength in long text generation: Compared with other baselines, GPT-4o demonstrates a superior ability to generate long, coherent text. In example 1 and example 3, GPT-4o produces detailed textual information with fewer than three characters generated incorrectly across more than 100 characters of text."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 531,
+ 305
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 531,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 531,
+ 305
+ ],
+ "type": "text",
+ "content": "- Baseline limitations: When the input prompt becomes extremely long, models such as Gemini 2.0 Flash, Ideogram 3.0, and Playground-v3 often produce significantly more errors or produce vague text patches that are difficult to recognize."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 530,
+ 342
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 530,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 530,
+ 342
+ ],
+ "type": "text",
+ "content": "- POSTA's performance: As a model specifically designed for poster-style text generation, POSTA performs closely to, or in some instances slightly more precisely than, GPT-4o. We hypothesize this is due to its multi-step pipeline tailored for long text rendering."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 77,
+ 353,
+ 532,
+ 376
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 353,
+ 532,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 353,
+ 532,
+ 376
+ ],
+ "type": "text",
+ "content": "Overall, we conclude that GPT-4o excels at long text rendering, offering overwhelming performance compared to most existing commercial models, and delivering results on par with the latest specialized research models."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 312,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 312,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 312,
+ 750
+ ],
+ "type": "text",
+ "content": "10"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 9
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 100,
+ 91,
+ 205,
+ 103
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 100,
+ 91,
+ 205,
+ 103
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 100,
+ 91,
+ 205,
+ 103
+ ],
+ "type": "text",
+ "content": "Short Text Rendering"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 208,
+ 112,
+ 220,
+ 125
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 208,
+ 112,
+ 220,
+ 125
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 208,
+ 112,
+ 220,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 208,
+ 112,
+ 220,
+ 125
+ ],
+ "type": "image",
+ "image_path": "9d7e40286ddeb28b93be6916d8101d062aa91f512e1c1ac6b0ccc889ae647631.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 222,
+ 115,
+ 402,
+ 127
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 222,
+ 115,
+ 402,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 222,
+ 115,
+ 402,
+ 127
+ ],
+ "type": "text",
+ "content": "Evaluation: Text Rendering Precision."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 97,
+ 129,
+ 181,
+ 215
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 97,
+ 129,
+ 181,
+ 215
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 129,
+ 181,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 129,
+ 181,
+ 215
+ ],
+ "type": "image",
+ "image_path": "08d9d7d0f4498fa74ab30be14fe461e1fd36819539ce56b0a1505126518de149.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 193,
+ 130,
+ 277,
+ 215
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 193,
+ 130,
+ 277,
+ 215
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 193,
+ 130,
+ 277,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 193,
+ 130,
+ 277,
+ 215
+ ],
+ "type": "image",
+ "image_path": "234feaba9c9e1d0f65d67e6df566f551271d9b6bc16b92cc18990362a5b21b2f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 288,
+ 130,
+ 373,
+ 215
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 288,
+ 130,
+ 373,
+ 215
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 288,
+ 130,
+ 373,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 288,
+ 130,
+ 373,
+ 215
+ ],
+ "type": "image",
+ "image_path": "0ce5a44bd3cd7500b4b78db75c0f1ccc81baa684ffa77ebdb728c28b0fe9e785.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 384,
+ 130,
+ 513,
+ 215
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 384,
+ 130,
+ 513,
+ 215
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 384,
+ 130,
+ 513,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 384,
+ 130,
+ 513,
+ 215
+ ],
+ "type": "image",
+ "image_path": "204fddddcbcd5e6532fa8831a9afe9f12c4bf879dee56a592fddb35d74766418.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 100,
+ 221,
+ 501,
+ 257
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 100,
+ 221,
+ 501,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 100,
+ 221,
+ 501,
+ 257
+ ],
+ "type": "text",
+ "content": "Input Text: \"A beautiful painting of flowing colors and styles forming the words 'The GPT-4o/Ideogram/FLUX/SD3 research paper is nowhere!'. the background is speckled with drops and splashes of paint.\""
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 97,
+ 262,
+ 154,
+ 347
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 97,
+ 262,
+ 154,
+ 347
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 262,
+ 154,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 262,
+ 154,
+ 347
+ ],
+ "type": "image",
+ "image_path": "5d4b6580226201ab5042c29511d6da5a3409b9f5c7fee1d2d9c074d14f64765b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 168,
+ 262,
+ 253,
+ 347
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 168,
+ 262,
+ 253,
+ 347
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 168,
+ 262,
+ 253,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 168,
+ 262,
+ 253,
+ 347
+ ],
+ "type": "image",
+ "image_path": "6ec2422a775f3f017131805ebc131bf30d8d32953e433ab7a7e35ce013d814bd.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 268,
+ 262,
+ 352,
+ 347
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 268,
+ 262,
+ 352,
+ 347
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 268,
+ 262,
+ 352,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 268,
+ 262,
+ 352,
+ 347
+ ],
+ "type": "image",
+ "image_path": "4f7ea6846a9f2e71ede47f66f8045aa234b77d2155c2f56eca68f95514f68e80.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 367,
+ 262,
+ 514,
+ 347
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 367,
+ 262,
+ 514,
+ 347
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 367,
+ 262,
+ 514,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 367,
+ 262,
+ 514,
+ 347
+ ],
+ "type": "image",
+ "image_path": "1b62b31f138cbdfe4ceec53b959cee36fc3c5f0c22524ab44623a5cd565ac9c8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 102,
+ 354,
+ 478,
+ 378
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 102,
+ 354,
+ 478,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 354,
+ 478,
+ 378
+ ],
+ "type": "text",
+ "content": "Input Text: \"Beautiful pixel art of a Wizard with hovering text 'Achievement unlocked: Diffusion models can spell now'.\""
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 97,
+ 385,
+ 154,
+ 469
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 97,
+ 385,
+ 154,
+ 469
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 385,
+ 154,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 385,
+ 154,
+ 469
+ ],
+ "type": "image",
+ "image_path": "4ba3110052c7723a701e2ccf5e847ad8c21956fcb34a978c1e55cce340d0131c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 77,
+ 651,
+ 532,
+ 707
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 651,
+ 532,
+ 707
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 651,
+ 532,
+ 707
+ ],
+ "type": "text",
+ "content": "Figure 5: Task: Short text rendering. Generate prompt-aligned, concise textual content (typically within 10 words) on an image. Setup: Each sample is produced based on a guiding text prompt. Comparisons are made with prior SOTA models [27, 2] and FLUX [51]. Observations: GPT-4o achieves performance on par with existing SOTA baselines in rendering short texts, consistently following the prompt with minimal errors. All evaluated methods—except FLUX [51]—deliver high-fidelity results in this setting."
+ }
+ ]
+ }
+ ],
+ "index": 27,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 173,
+ 385,
+ 307,
+ 469
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 173,
+ 385,
+ 307,
+ 469
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 173,
+ 385,
+ 307,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 173,
+ 385,
+ 307,
+ 469
+ ],
+ "type": "image",
+ "image_path": "4fed5bdfe6b3d24177ff8f96422315e81b3ae864616ce0f21770566bf7fd8891.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 326,
+ 385,
+ 410,
+ 469
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 326,
+ 385,
+ 410,
+ 469
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 326,
+ 385,
+ 410,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 385,
+ 410,
+ 469
+ ],
+ "type": "image",
+ "image_path": "83fabbd535e803f344bff99d3d53a6c8a71d95998a6a377d7bae9a19ebb0830a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 430,
+ 385,
+ 514,
+ 469
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 430,
+ 385,
+ 514,
+ 469
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 430,
+ 385,
+ 514,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 430,
+ 385,
+ 514,
+ 469
+ ],
+ "type": "image",
+ "image_path": "63a0fda286dbcf6aab1f70694ab0576b25d5fd247827bdfa027aca3ce2984b08.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 102,
+ 478,
+ 480,
+ 490
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 102,
+ 478,
+ 480,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 478,
+ 480,
+ 490
+ ],
+ "type": "text",
+ "content": "Input Text: \"A monkey holding a sign reading 'Scaling transformer models is awesome'.\""
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 97,
+ 500,
+ 181,
+ 585
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 97,
+ 500,
+ 181,
+ 585
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 500,
+ 181,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 500,
+ 181,
+ 585
+ ],
+ "type": "image",
+ "image_path": "b44926116e89a328041a0572f04bcac45c5be77916af431b03c4a0ab99dd761f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 187,
+ 500,
+ 271,
+ 585
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 187,
+ 500,
+ 271,
+ 585
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 187,
+ 500,
+ 271,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 187,
+ 500,
+ 271,
+ 585
+ ],
+ "type": "image",
+ "image_path": "cb9fbff8e79609bd0c0397e9ff8bdd26c4934e0c043bf90309793e247c53acde.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 277,
+ 500,
+ 361,
+ 585
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 277,
+ 500,
+ 361,
+ 585
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 277,
+ 500,
+ 361,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 277,
+ 500,
+ 361,
+ 585
+ ],
+ "type": "image",
+ "image_path": "ecba2242bac62c8676019c58e03ec1fc8ba9f66e7431827caf98cf3565ae7429.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 20,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 368,
+ 501,
+ 512,
+ 585
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 368,
+ 501,
+ 512,
+ 585
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 368,
+ 501,
+ 512,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 368,
+ 501,
+ 512,
+ 585
+ ],
+ "type": "image",
+ "image_path": "eb143f4361a33ce174515f230c11a216ac7590d56a2fc328a2adc862245f0ce7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 101,
+ 590,
+ 492,
+ 613
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 101,
+ 590,
+ 492,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 590,
+ 492,
+ 613
+ ],
+ "type": "text",
+ "content": "Input Text: \"A surreal and humorous scene in a classroom with the words 'GPUs go brrrrr' written in white chalk on a blackboard. In front of the blackboard.\""
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 123,
+ 621,
+ 154,
+ 632
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 123,
+ 621,
+ 154,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 621,
+ 154,
+ 632
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 200,
+ 621,
+ 258,
+ 633
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 621,
+ 258,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 621,
+ 258,
+ 633
+ ],
+ "type": "text",
+ "content": "Ideogram 3.0"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 306,
+ 621,
+ 332,
+ 632
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 621,
+ 332,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 621,
+ 332,
+ 632
+ ],
+ "type": "text",
+ "content": "FLUX"
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 429,
+ 621,
+ 451,
+ 632
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 429,
+ 621,
+ 451,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 429,
+ 621,
+ 451,
+ 632
+ ],
+ "type": "text",
+ "content": "SD3"
+ }
+ ]
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "text",
+ "content": "11"
+ }
+ ]
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 10
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 229,
+ 101,
+ 394,
+ 112
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 229,
+ 101,
+ 394,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 229,
+ 101,
+ 394,
+ 112
+ ],
+ "type": "text",
+ "content": "Evaluation: Text Rendering Precision."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 111,
+ 116,
+ 196,
+ 243
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 111,
+ 116,
+ 196,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 116,
+ 196,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 116,
+ 196,
+ 243
+ ],
+ "type": "image",
+ "image_path": "afce16d9cf439bcdc04fa94f8ca4f2227195b2cf36f175b299fa1b98664ddbfa.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 139,
+ 247,
+ 169,
+ 256
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 247,
+ 169,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 247,
+ 169,
+ 256
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 197,
+ 116,
+ 283,
+ 243
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 197,
+ 116,
+ 283,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 197,
+ 116,
+ 283,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 197,
+ 116,
+ 283,
+ 243
+ ],
+ "type": "image",
+ "image_path": "c6b354d42faa888d692b35bdfbc5cd4599018152cf0064da6cd13e513960b86e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 226,
+ 247,
+ 257,
+ 256
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 226,
+ 247,
+ 257,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 226,
+ 247,
+ 257,
+ 256
+ ],
+ "type": "text",
+ "content": "POSTA"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 284,
+ 116,
+ 370,
+ 243
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 284,
+ 116,
+ 370,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 116,
+ 370,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 116,
+ 370,
+ 243
+ ],
+ "type": "image",
+ "image_path": "f4cd9c8e039a64fce1592ca2bb40e8f9f9bf8f8c530f65984c76ff6708a91c6f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 293,
+ 247,
+ 359,
+ 256
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 247,
+ 359,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 247,
+ 359,
+ 256
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 370,
+ 116,
+ 497,
+ 243
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 370,
+ 116,
+ 497,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 370,
+ 116,
+ 497,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 370,
+ 116,
+ 497,
+ 243
+ ],
+ "type": "image",
+ "image_path": "c5582fe68220a5a384efa294ef35989ebef4760bd5b2897bd0ee23ceffe344a0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 405,
+ 247,
+ 460,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 405,
+ 247,
+ 460,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 405,
+ 247,
+ 460,
+ 258
+ ],
+ "type": "text",
+ "content": "Ideogram 3.0"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 259,
+ 165,
+ 269
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 259,
+ 165,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 259,
+ 165,
+ 269
+ ],
+ "type": "text",
+ "content": "Input Text:"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 270,
+ 495,
+ 290
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 270,
+ 495,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 270,
+ 495,
+ 290
+ ],
+ "type": "text",
+ "content": "\"Generate a movie poster with a sci-fi space theme, a solitary figure standing on an alien planet, facing a massive outpost."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 290,
+ 270,
+ 300
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 290,
+ 270,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 290,
+ 270,
+ 300
+ ],
+ "type": "text",
+ "content": "The poster displays the following text:"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 300,
+ 211,
+ 310
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 300,
+ 211,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 300,
+ 211,
+ 310
+ ],
+ "type": "text",
+ "content": "Title: The Last Outpost"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 310,
+ 298,
+ 320
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 310,
+ 298,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 310,
+ 298,
+ 320
+ ],
+ "type": "text",
+ "content": "Subtitle: When the stars fall, the truth rises"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 115,
+ 321,
+ 169,
+ 330
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 321,
+ 169,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 321,
+ 169,
+ 330
+ ],
+ "type": "text",
+ "content": "Information:"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 331,
+ 225,
+ 341
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 331,
+ 225,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 331,
+ 225,
+ 341
+ ],
+ "type": "text",
+ "content": "Produced by Jackson Ward"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 342,
+ 206,
+ 351
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 342,
+ 206,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 342,
+ 206,
+ 351
+ ],
+ "type": "text",
+ "content": "Music by Aria Calloway"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 351,
+ 227,
+ 361
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 351,
+ 227,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 351,
+ 227,
+ 361
+ ],
+ "type": "text",
+ "content": "Screenplay by Elena Sharpe"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 361,
+ 227,
+ 372
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 361,
+ 227,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 361,
+ 227,
+ 372
+ ],
+ "type": "text",
+ "content": "Directed By Sylvia Hartman"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 372,
+ 487,
+ 411
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 372,
+ 487,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 372,
+ 487,
+ 411
+ ],
+ "type": "text",
+ "content": "\"A visually stunning and narratively gripping exploration of the unknown. The Last Outpost masterfully blends elements of science fiction, mystery, and psychological thriller, creating a hauntingly atmospheric journey that will leave audiences on the edge of their seats.\" -- Global Film Review\"."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 111,
+ 415,
+ 194,
+ 540
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 111,
+ 415,
+ 194,
+ 540
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 415,
+ 194,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 415,
+ 194,
+ 540
+ ],
+ "type": "image",
+ "image_path": "e1dc82e6233835aef84e7c0d62fd27535b3b4aee7f6bc4beab765dacdbbdc4c1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 139,
+ 547,
+ 168,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 547,
+ 168,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 547,
+ 168,
+ 556
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 22,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 195,
+ 415,
+ 282,
+ 541
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 195,
+ 415,
+ 282,
+ 541
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 195,
+ 415,
+ 282,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 195,
+ 415,
+ 282,
+ 541
+ ],
+ "type": "image",
+ "image_path": "59f2bdbb4c18ff5b2eaf1af8462de2d557b98afb5403da1a76ec1f639e6017a6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 23,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 226,
+ 547,
+ 257,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 226,
+ 547,
+ 257,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 226,
+ 547,
+ 257,
+ 556
+ ],
+ "type": "text",
+ "content": "POSTA"
+ }
+ ]
+ }
+ ],
+ "index": 24,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 282,
+ 415,
+ 377,
+ 541
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 282,
+ 415,
+ 377,
+ 541
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 282,
+ 415,
+ 377,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 282,
+ 415,
+ 377,
+ 541
+ ],
+ "type": "image",
+ "image_path": "3c2edc361f25ec06384618fcf44a118efef67ed2c14c3a957eec1c7c432abf66.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 293,
+ 547,
+ 359,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 547,
+ 359,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 547,
+ 359,
+ 556
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 26,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 378,
+ 415,
+ 504,
+ 541
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 378,
+ 415,
+ 504,
+ 541
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 378,
+ 415,
+ 504,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 378,
+ 415,
+ 504,
+ 541
+ ],
+ "type": "image",
+ "image_path": "1751faef203209061a6fae572da8f65fff10f9f74ef783748441d19f7f4a2be9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 27,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 405,
+ 547,
+ 460,
+ 557
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 405,
+ 547,
+ 460,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 405,
+ 547,
+ 460,
+ 557
+ ],
+ "type": "text",
+ "content": "Ideogram 3.0"
+ }
+ ]
+ }
+ ],
+ "index": 28,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 77,
+ 657,
+ 533,
+ 734
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 657,
+ 533,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 657,
+ 533,
+ 734
+ ],
+ "type": "text",
+ "content": "Figure 6: Task: Long text rendering. Generate extended, coherent, and prompt-consistent textual content on an image. Setup: Evaluations are conducted against advanced baselines including POSTA [12], Gemini 2.0 Flash [99], Ideogram 3.0 [2], and Playground-v3 [64]. Observations: GPT-4o excels in long text rendering by producing coherent, detailed textual information with very few character errors. In contrast, models like Gemini 2.0 Flash, Ideogram 3.0, and Playground-v3 often exhibit increased errors or generate vague text when faced with lengthy prompts, while POSTA's tailored multi-step pipeline sometimes yields competitive precision. Overall, GPT-4o outperforms most commercial models and rivals specialized research approaches in extended text generation."
+ }
+ ]
+ }
+ ],
+ "index": 34,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 113,
+ 559,
+ 162,
+ 569
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 559,
+ 162,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 559,
+ 162,
+ 569
+ ],
+ "type": "text",
+ "content": "Input Text:"
+ }
+ ]
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 112,
+ 570,
+ 490,
+ 610
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 570,
+ 490,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 570,
+ 490,
+ 610
+ ],
+ "type": "text",
+ "content": "\"Create a poster with the theme of a Journey of Solitude. The background should depict a lone figure walking toward an unusable form of transportation. The scene should evoke a sense of being lost, helplessness, and desolation, capturing the emotional weight of losing oneself in a barren, unforgiving landscape."
+ }
+ ]
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 113,
+ 610,
+ 209,
+ 621
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 610,
+ 209,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 610,
+ 209,
+ 621
+ ],
+ "type": "text",
+ "content": "Title: Solitary Journeys"
+ }
+ ]
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 113,
+ 621,
+ 194,
+ 630
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 621,
+ 194,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 621,
+ 194,
+ 630
+ ],
+ "type": "text",
+ "content": "Subtitle: Elara Voss"
+ }
+ ]
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 113,
+ 631,
+ 345,
+ 640
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 631,
+ 345,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 631,
+ 345,
+ 640
+ ],
+ "type": "text",
+ "content": "Information: WANDERING THROUGH THE UNKNOWN\"."
+ }
+ ]
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 127,
+ 79,
+ 220,
+ 92
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 127,
+ 79,
+ 220,
+ 92
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 79,
+ 220,
+ 92
+ ],
+ "type": "text",
+ "content": "Long Text Rendering"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 300,
+ 741,
+ 312,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 312,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 312,
+ 750
+ ],
+ "type": "text",
+ "content": "12"
+ }
+ ]
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 11
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 123,
+ 231,
+ 218,
+ 244
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 123,
+ 231,
+ 218,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 231,
+ 218,
+ 244
+ ],
+ "type": "text",
+ "content": "Long Text Rendering"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 212,
+ 251,
+ 225,
+ 264
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 212,
+ 251,
+ 225,
+ 264
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 212,
+ 251,
+ 225,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 212,
+ 251,
+ 225,
+ 264
+ ],
+ "type": "image",
+ "image_path": "072dabf79d619daa8eeb4cd2f8867859c2a5525ca4ceff1ee839637ee46f829f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 225,
+ 254,
+ 396,
+ 266
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 225,
+ 254,
+ 396,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 225,
+ 254,
+ 396,
+ 266
+ ],
+ "type": "text",
+ "content": "Evaluation: Text Rendering Precision."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 102,
+ 269,
+ 188,
+ 399
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 102,
+ 269,
+ 188,
+ 399
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 102,
+ 269,
+ 188,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 269,
+ 188,
+ 399
+ ],
+ "type": "image",
+ "image_path": "37c0d9d0ba1eed41834771a2d8df3690d2b37294a08a5844702161e53aa817ab.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 129,
+ 404,
+ 160,
+ 414
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 404,
+ 160,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 404,
+ 160,
+ 414
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 119,
+ 554,
+ 488,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 554,
+ 488,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 554,
+ 488,
+ 566
+ ],
+ "type": "text",
+ "content": "Figure 7: Task: Long text rendering. The Setup and Observations are the same as Figure 6."
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 189,
+ 269,
+ 278,
+ 399
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 189,
+ 269,
+ 278,
+ 399
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 189,
+ 269,
+ 278,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 189,
+ 269,
+ 278,
+ 399
+ ],
+ "type": "image",
+ "image_path": "61b9d5a9c147c446387301c4f13acd7ceeccc06f653954410c1dc398d1fd4bc3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 217,
+ 404,
+ 249,
+ 414
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 217,
+ 404,
+ 249,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 217,
+ 404,
+ 249,
+ 414
+ ],
+ "type": "text",
+ "content": "POSTA"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 278,
+ 269,
+ 376,
+ 399
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 278,
+ 269,
+ 376,
+ 399
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 278,
+ 269,
+ 376,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 278,
+ 269,
+ 376,
+ 399
+ ],
+ "type": "image",
+ "image_path": "96c1a55c7c5f312a71c89d34f708b0e4f8da4556922ffdee41d673b76ea82dad.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 293,
+ 404,
+ 361,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 404,
+ 361,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 404,
+ 361,
+ 415
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 377,
+ 269,
+ 507,
+ 399
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 377,
+ 269,
+ 507,
+ 399
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 377,
+ 269,
+ 507,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 377,
+ 269,
+ 507,
+ 399
+ ],
+ "type": "image",
+ "image_path": "faf4eca5a3720dff227ce6a4ad8f4898de848e4e08fec90adf87952d14446f7e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 418,
+ 404,
+ 477,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 418,
+ 404,
+ 477,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 418,
+ 404,
+ 477,
+ 415
+ ],
+ "type": "text",
+ "content": "Playground-v3"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 107,
+ 426,
+ 159,
+ 436
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 426,
+ 159,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 426,
+ 159,
+ 436
+ ],
+ "type": "text",
+ "content": "Input Text:"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 107,
+ 437,
+ 485,
+ 468
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 437,
+ 485,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 437,
+ 485,
+ 468
+ ],
+ "type": "text",
+ "content": "\"Please generate an artistic and stylized promotional poster. The style is an artistic painting style. The theme is about nature and city. The poster displays the following information: Title: Fragmented Harmony"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 107,
+ 468,
+ 328,
+ 479
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 468,
+ 328,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 468,
+ 328,
+ 479
+ ],
+ "type": "text",
+ "content": "Subtitle Between the steel and sky, life finds its way."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 479,
+ 498,
+ 532
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 479,
+ 498,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 479,
+ 498,
+ 532
+ ],
+ "type": "text",
+ "content": "Information: Amid the towering strucions and the quiet persistence of nature, a delicate balance emerges. The complex and often contradictory relationship between urban development and the natural world reveals itself in fleeting moments of harmony. Though fragmented, life continues, threading its way through the shadows of progress. Here, conflict and coexistence form an intricate dance--sometimes at odds, sometimes in unexpected unity\"."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "text",
+ "content": "13"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 12
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 207,
+ 83
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 207,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 72,
+ 207,
+ 83
+ ],
+ "type": "text",
+ "content": "2.1.3 Document Generation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 77,
+ 91,
+ 533,
+ 125
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 91,
+ 533,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 91,
+ 533,
+ 125
+ ],
+ "type": "text",
+ "content": "We also explore a novel task: document image generation with GPT-4o, comparing its performance with Gemini 2.0 Flash [99] and Playground-v3 [64]. As shown in Figure 8 - 10, GPT-4o produces document images with cleaner layouts and more consistent content."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 140,
+ 240,
+ 152
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 140,
+ 240,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 140,
+ 240,
+ 152
+ ],
+ "type": "text",
+ "content": "Document Image Generation"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 205,
+ 161,
+ 218,
+ 175
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 205,
+ 161,
+ 218,
+ 175
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 205,
+ 161,
+ 218,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 205,
+ 161,
+ 218,
+ 175
+ ],
+ "type": "image",
+ "image_path": "7ff25ac80d044220739d61284585a070ee9142eede21fe8dbf47dff3cb2ffa9c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 219,
+ 164,
+ 398,
+ 176
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 219,
+ 164,
+ 398,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 219,
+ 164,
+ 398,
+ 176
+ ],
+ "type": "text",
+ "content": "Evaluation: Text Rendering Precision."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 202,
+ 203,
+ 211
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 202,
+ 203,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 202,
+ 203,
+ 211
+ ],
+ "type": "text",
+ "content": "Attention Is All You Need"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 113,
+ 219,
+ 208,
+ 236
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 219,
+ 208,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 219,
+ 208,
+ 236
+ ],
+ "type": "text",
+ "content": "Ashish Vaswani Noam Shazeer Niki Parmar Jakob Uszkoreit Lilion Jones Aidan N.Gomez Lukasz Kaiser IIIa Polosukhin"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 149,
+ 246,
+ 170,
+ 252
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 149,
+ 246,
+ 170,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 149,
+ 246,
+ 170,
+ 252
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 97,
+ 254,
+ 225,
+ 328
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 254,
+ 225,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 254,
+ 225,
+ 328
+ ],
+ "type": "text",
+ "content": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on tw machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 247,
+ 256,
+ 307,
+ 262
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 247,
+ 256,
+ 307,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 247,
+ 256,
+ 307,
+ 262
+ ],
+ "type": "text",
+ "content": "Attention Is All You Need"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 243,
+ 262,
+ 347,
+ 329
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 243,
+ 262,
+ 347,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 243,
+ 262,
+ 347,
+ 329
+ ],
+ "type": "text",
+ "content": "Ashish Vourakis, NM Nazarov, NJI Parmar, Jbab Udekoreli, Lillion Jones, Adiinil S. W. Coomce, Lakota Kiber, Pala Poslashov, and M. A. D. G. R. Smith. 2017. The neural network models are based on recurrent or conventional neural networks in an encoder-decoder loss. The best produced models also connect the encoder and decoder loss through an attention mechanism. As, we propose a new study mechanism for a machine, the Transformer, based attention mechanisms, dispensing with the same weight as the encoder-decoder loss. The results of the training tasks show these models to be superior in quality while being more parsibilized and requiring significantly less time to time to move train. Our model achieves 84.2 BLUE in the WMT-50 Go-to-translation task, which is comparable to the performance of our previous work [3]. In addition, our WMT-30 Pragmatic task, our model established a new single-model state-of-the-art-state-of-the-art BLUE score of 41.8 when "
+ },
+ {
+ "bbox": [
+ 243,
+ 262,
+ 347,
+ 329
+ ],
+ "type": "inline_equation",
+ "content": "\\alpha = 5"
+ },
+ {
+ "bbox": [
+ 243,
+ 262,
+ 347,
+ 329
+ ],
+ "type": "text",
+ "content": " training for 3.5 days on GPUs after fraction of the training costs of the best models from literature. We propose Transformer generalizes well by applying it successfully to English syntacticity parsing both with large and limited training data."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 369,
+ 181,
+ 432,
+ 187
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 369,
+ 181,
+ 432,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 369,
+ 181,
+ 432,
+ 187
+ ],
+ "type": "text",
+ "content": "Attention Is All You Need"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 369,
+ 191,
+ 424,
+ 213
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 369,
+ 191,
+ 424,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 369,
+ 191,
+ 424,
+ 213
+ ],
+ "type": "text",
+ "content": "Ashish A. [Al] You. Wea a nono' Ainon 1 \nAshish Yawani, Naqadzaree, laokotri \nAlok Uzzotner, Anokalika Sanik, Jokoslav adar, Gosak III Ploosukhaini \nAlok Uzzotner, Anokalika Sanik"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 369,
+ 222,
+ 507,
+ 331
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 369,
+ 222,
+ 507,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 369,
+ 222,
+ 507,
+ 331
+ ],
+ "type": "text",
+ "content": "Antext, exotocic sequecra transoedscnncs on cbr be caes baccared on bracococcyne nort bell netiabon an ecocycclion, an ecocyclon. TeTrane: the ensonnnmnsn neeepnckian. Epipcnie rile kceely on meenctiny As adeterdiencr. \nnpopioors sonr tonarwamchim. I mtnr. vorti. inenpoea a dedusum minnyss onomcrh. cortordone.lora ontata tose or uin hoperiosper. \nThe rorner is s oovl dtt maive de acemnccnodkdu aleu cormunb-dlr bing dvl-ndr 016de and mechance \n11 Dae nucnccnng \nceso nucnse \n12 - eannnr er on attonne amehnes asnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnncnnccnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnn"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 143,
+ 338,
+ 178,
+ 348
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 143,
+ 338,
+ 178,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 338,
+ 178,
+ 348
+ ],
+ "type": "text",
+ "content": "GPT40"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 255,
+ 338,
+ 334,
+ 350
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 255,
+ 338,
+ 334,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 255,
+ 338,
+ 334,
+ 350
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 406,
+ 337,
+ 473,
+ 350
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 406,
+ 337,
+ 473,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 406,
+ 337,
+ 473,
+ 350
+ ],
+ "type": "text",
+ "content": "Playground-v3"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 99,
+ 355,
+ 159,
+ 367
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 99,
+ 355,
+ 159,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 99,
+ 355,
+ 159,
+ 367
+ ],
+ "type": "text",
+ "content": "Input Text:"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 97,
+ 369,
+ 487,
+ 396
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 369,
+ 487,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 369,
+ 487,
+ 396
+ ],
+ "type": "text",
+ "content": "\"Generate A realistic screenshot of the first page of the Paper from the following information:"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 97,
+ 399,
+ 252,
+ 411
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 399,
+ 252,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 399,
+ 252,
+ 411
+ ],
+ "type": "text",
+ "content": "Title: Attention Is All You Need"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 97,
+ 414,
+ 510,
+ 441
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 414,
+ 510,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 414,
+ 510,
+ 441
+ ],
+ "type": "text",
+ "content": "Author List: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 97,
+ 444,
+ 509,
+ 653
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 444,
+ 509,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 444,
+ 509,
+ 653
+ ],
+ "type": "text",
+ "content": "Abstract: The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.\""
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 77,
+ 667,
+ 533,
+ 700
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 667,
+ 533,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 667,
+ 533,
+ 700
+ ],
+ "type": "text",
+ "content": "Figure 8: Task: Document image generation. Setup: Each row shows a text prompt and the generated outputs from GPT-4o, Gemini 2.0 Flash [99], and Playground-v3 [64]. Observation: GPT-4o can generate more consistent and accurate font and format than the other two models."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "14"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 13
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 104,
+ 119,
+ 239,
+ 133
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 119,
+ 239,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 119,
+ 239,
+ 133
+ ],
+ "type": "text",
+ "content": "Document Image Generation"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 205,
+ 140,
+ 217,
+ 154
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 205,
+ 140,
+ 217,
+ 154
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 205,
+ 140,
+ 217,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 205,
+ 140,
+ 217,
+ 154
+ ],
+ "type": "image",
+ "image_path": "e960d89a9332326f5fac58239c1d5784d9e237d7ccede93ea1c82462b3f589b3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 218,
+ 143,
+ 398,
+ 156
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 218,
+ 143,
+ 398,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 218,
+ 143,
+ 398,
+ 156
+ ],
+ "type": "text",
+ "content": "Evaluation: Text Rendering Precision."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 99,
+ 178,
+ 232,
+ 198
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 99,
+ 178,
+ 232,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 99,
+ 178,
+ 232,
+ 198
+ ],
+ "type": "text",
+ "content": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 206,
+ 205,
+ 220
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 206,
+ 205,
+ 220
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 206,
+ 205,
+ 220
+ ],
+ "type": "text",
+ "content": "Jacob Devlin Ming-Wei Chang Kenton Lee Kristina Toutanova"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 154,
+ 230,
+ 176,
+ 237
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 154,
+ 230,
+ 176,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 154,
+ 230,
+ 176,
+ 237
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 97,
+ 239,
+ 235,
+ 290
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 239,
+ 235,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 239,
+ 235,
+ 290
+ ],
+ "type": "text",
+ "content": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result; the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 97,
+ 290,
+ 237,
+ 319
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 290,
+ 237,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 290,
+ 237,
+ 319
+ ],
+ "type": "text",
+ "content": "BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to "
+ },
+ {
+ "bbox": [
+ 97,
+ 290,
+ 237,
+ 319
+ ],
+ "type": "inline_equation",
+ "content": "80.5\\%"
+ },
+ {
+ "bbox": [
+ 97,
+ 290,
+ 237,
+ 319
+ ],
+ "type": "text",
+ "content": " (7.7% point absolute improvement), MultiNLI accuracy to "
+ },
+ {
+ "bbox": [
+ 97,
+ 290,
+ 237,
+ 319
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{S}6.7\\%"
+ },
+ {
+ "bbox": [
+ 97,
+ 290,
+ 237,
+ 319
+ ],
+ "type": "text",
+ "content": " (4.6% absolute improvement), SQAud v1.1 question answering Test F1 to 93.2 (1.5 point absolute"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 148,
+ 324,
+ 183,
+ 335
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 148,
+ 324,
+ 183,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 148,
+ 324,
+ 183,
+ 335
+ ],
+ "type": "text",
+ "content": "GPT 40"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 255,
+ 178,
+ 357,
+ 196
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 255,
+ 178,
+ 357,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 255,
+ 178,
+ 357,
+ 196
+ ],
+ "type": "text",
+ "content": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 257,
+ 198,
+ 354,
+ 210
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 257,
+ 198,
+ 354,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 257,
+ 198,
+ 354,
+ 210
+ ],
+ "type": "text",
+ "content": "Ashlor Jacob Doslin Ming Wei Chang, Kenton Lee Abstrut Win, Karinlin Touranosa"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 255,
+ 210,
+ 360,
+ 254
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 255,
+ 210,
+ 360,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 255,
+ 210,
+ 360,
+ 254
+ ],
+ "type": "text",
+ "content": "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Cryostatnet as Transformer. Unlike recnec mean, BERT is designed by pre-train deep bidirectional representations from nonshaded text by jasily selfconforming on both text 1st&4, xavier coint eorect, tate 1st&2 to win 1st, sacp0-twincent BERT model can be fine mused with just one additional output layer to create of dafve for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "type": "text",
+ "content": "BERT is conceptually simple and empirically powerful. It obtains now state-of-the-art results on eleven evertweaven multimoment language processing tasks, including GLUE score to "
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "type": "inline_equation",
+ "content": "80.5\\%"
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "type": "text",
+ "content": " (7.69 absolute improvement), pushing the GLUE alloselect improvement). MulaNLI accuracy to "
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "type": "inline_equation",
+ "content": "\\varepsilon_{\\mathrm{M}} = 0.1"
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "type": "inline_equation",
+ "content": "\\nu_{\\mathrm{L}} < \\varepsilon_{\\mathrm{M}} / \\varepsilon_{\\mathrm{M}}"
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 361,
+ 297
+ ],
+ "type": "text",
+ "content": " (1.5 point absolute improvement) and SQoAD v2.0 Test F1 to 83.1 (3.1 point absolute improvement)."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 268,
+ 325,
+ 347,
+ 336
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 268,
+ 325,
+ 347,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 268,
+ 325,
+ 347,
+ 336
+ ],
+ "type": "text",
+ "content": "Gemini 2.0 Flash"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 380,
+ 182,
+ 504,
+ 196
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 182,
+ 504,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 182,
+ 504,
+ 196
+ ],
+ "type": "text",
+ "content": "BERT: Pre-training on Lepi Bidellar Tansson Translons for Language Understond mderance litting' cetting from t cowf henvaming"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 380,
+ 198,
+ 502,
+ 217
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 18,
+ "blocks": [
+ {
+ "bbox": [
+ 380,
+ 198,
+ 399,
+ 202
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 198,
+ 399,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 198,
+ 399,
+ 202
+ ],
+ "type": "text",
+ "content": "Author, List:"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 380,
+ 203,
+ 502,
+ 213
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 203,
+ 502,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 203,
+ 502,
+ 213
+ ],
+ "type": "text",
+ "content": "J Asbad Devlin, Yiw Changuaguagaa Kionn age rspgectangane cans \n"
+ },
+ {
+ "bbox": [
+ 380,
+ 203,
+ 502,
+ 213
+ ],
+ "type": "inline_equation",
+ "content": "^{a}"
+ },
+ {
+ "bbox": [
+ 380,
+ 203,
+ 502,
+ 213
+ ],
+ "type": "text",
+ "content": "pressin liKistcn-Toutanfa"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 380,
+ 213,
+ 421,
+ 217
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 213,
+ 421,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 213,
+ 421,
+ 217
+ ],
+ "type": "text",
+ "content": "represeons-Uintanlvania"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 218,
+ 396,
+ 224
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 218,
+ 396,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 218,
+ 396,
+ 224
+ ],
+ "type": "text",
+ "content": "A"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 380,
+ 227,
+ 511,
+ 260
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 227,
+ 511,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 227,
+ 511,
+ 260
+ ],
+ "type": "text",
+ "content": "We introduce a new languagegretrovetercendentiale monoclin klonstionist monole conBldfecarstaadss reprenters from nemer raje Sfiflnonanecones. desessnissrall ranauagaleafdelfe xyn unming on hnlaesaeare two ploddes also the por-entant canteletory a state-vwraon one-on-one coffice of anisotropy, and the ploso-syntropic colective of states. Of s2s212310 pain or questionbmporansuansluangene Tcnf?f to ingest ingf Sf10 tto 46 w. (I, test),marshersonalizne imnance immmens"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 380,
+ 268,
+ 512,
+ 319
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 25,
+ "blocks": [
+ {
+ "bbox": [
+ 380,
+ 268,
+ 490,
+ 274
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 268,
+ 490,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 268,
+ 490,
+ 274
+ ],
+ "type": "text",
+ "content": "BERT is conceptually imlilienenarplenpholcft-nu surate-fine-ams"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 380,
+ 274,
+ 504,
+ 289
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 274,
+ 504,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 274,
+ 504,
+ 289
+ ],
+ "type": "text",
+ "content": "ronen 1 oonranaalwauanu viipopluoteforocnirandinns inget caught anourage, vovlurvulgina for nain. 2004. The use of the word \"sulfur\" in the text is a question envoing SUIF u697."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 380,
+ 289,
+ 512,
+ 300
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 289,
+ 512,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 289,
+ 512,
+ 300
+ ],
+ "type": "text",
+ "content": "- GBFEscors/aanoreqquasurf and Squad w.10 aninlvalte 83.7% 4.6% (X) \n- TeST fto onop:11.3% (x)"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 380,
+ 303,
+ 511,
+ 319
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 380,
+ 303,
+ 511,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 380,
+ 303,
+ 511,
+ 319
+ ],
+ "type": "text",
+ "content": "→ BERT is cponconuynlyrsnaintally pocefine-at-ut: ouvah176. (JET v.c.37% quinting anguen linyuH-aCLS sccorts onoonssthe sonea 4000A/AresoVc LEAU pioiHcB: gnrmaeh an epesrourinans A7c)0v o.o35 aed 1170"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 324,
+ 481,
+ 338
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 414,
+ 324,
+ 481,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 414,
+ 324,
+ 481,
+ 338
+ ],
+ "type": "text",
+ "content": "Playground-v3"
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 99,
+ 349,
+ 158,
+ 361
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 99,
+ 349,
+ 158,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 99,
+ 349,
+ 158,
+ 361
+ ],
+ "type": "text",
+ "content": "Input Text:"
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 97,
+ 362,
+ 483,
+ 389
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 362,
+ 483,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 362,
+ 483,
+ 389
+ ],
+ "type": "text",
+ "content": "Generate A realistic screenshot of the first page of the Paper from the following information:"
+ }
+ ]
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 97,
+ 392,
+ 449,
+ 422
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 392,
+ 449,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 392,
+ 449,
+ 422
+ ],
+ "type": "text",
+ "content": "Title: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding"
+ }
+ ]
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 97,
+ 423,
+ 511,
+ 555
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 423,
+ 511,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 423,
+ 511,
+ 555
+ ],
+ "type": "text",
+ "content": "Author List: Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova \nAbstract: We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications."
+ }
+ ]
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 97,
+ 558,
+ 511,
+ 632
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 558,
+ 511,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 558,
+ 511,
+ 632
+ ],
+ "type": "text",
+ "content": "BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to "
+ },
+ {
+ "bbox": [
+ 97,
+ 558,
+ 511,
+ 632
+ ],
+ "type": "inline_equation",
+ "content": "80.5\\%"
+ },
+ {
+ "bbox": [
+ 97,
+ 558,
+ 511,
+ 632
+ ],
+ "type": "text",
+ "content": " (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)."
+ }
+ ]
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 108,
+ 663,
+ 500,
+ 676
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 663,
+ 500,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 663,
+ 500,
+ 676
+ ],
+ "type": "text",
+ "content": "Figure 9: Task: Document image generation. The Setup and Observations are the same as Fig. 8."
+ }
+ ]
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "text",
+ "content": "15"
+ }
+ ]
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 14
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 205,
+ 121,
+ 217,
+ 137
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 205,
+ 121,
+ 217,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 205,
+ 121,
+ 217,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 205,
+ 121,
+ 217,
+ 137
+ ],
+ "type": "image",
+ "image_path": "9c2265499ec9f6a0c6e61a8914311c8b94d20424cd7a53a126ce5f30f3f6bf33.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 219,
+ 125,
+ 398,
+ 138
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 219,
+ 125,
+ 398,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 219,
+ 125,
+ 398,
+ 138
+ ],
+ "type": "text",
+ "content": "Evaluation: Text Rendering Precision."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 168,
+ 211,
+ 198
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 168,
+ 211,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 168,
+ 211,
+ 198
+ ],
+ "type": "text",
+ "content": "You Only Look Once: Unified, Real-Time Object Detection"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 205,
+ 220,
+ 217
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 205,
+ 220,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 205,
+ 220,
+ 217
+ ],
+ "type": "text",
+ "content": "Joseph Redmon, Santosh Divvala, Ross Girshick Ali Farhadi"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 153,
+ 226,
+ 174,
+ 232
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 153,
+ 226,
+ 174,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 153,
+ 226,
+ 174,
+ 232
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 102,
+ 232,
+ 230,
+ 274
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 102,
+ 232,
+ 230,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 232,
+ 230,
+ 274
+ ],
+ "type": "text",
+ "content": "We present YOLO, a new approach to object detection. Prior work on object detection repurposes classifiers to perform detection. Instead, we frame object detection as a regression problem to spatially separated bounding boxes and associated class probabilities. A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation. Since the whole detection pipeline is a single network, it can be optimized end-to-end directly on detection performance."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 102,
+ 274,
+ 230,
+ 307
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 102,
+ 274,
+ 230,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 274,
+ 230,
+ 307
+ ],
+ "type": "text",
+ "content": "Our unified architecture is extremely fast. Our base YOLO model processes images in real-time at 45 frames per second. A smaller version of the network. Fast YOLO, processes an astounding 155 frames per second while still achieving double the mAP of other real-time detectors. Compared to state-of-the-art detection systems YOLO makes more localization errors but is far less likely to pred"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 254,
+ 171,
+ 355,
+ 185
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 254,
+ 171,
+ 355,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 254,
+ 171,
+ 355,
+ 185
+ ],
+ "type": "text",
+ "content": "You Only Look Once: Unlimited, Time/Indirect Decciption \nAuthor:Lesser Jowesh Redmonial Siptedthi Adri Farhad"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 253,
+ 186,
+ 381,
+ 205
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 253,
+ 186,
+ 381,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 253,
+ 186,
+ 381,
+ 205
+ ],
+ "type": "text",
+ "content": "We present YJLO, a new approach to object detection. Driver search uses oleejeon cieeuses to represen or correct deferential finwication. Istegedel, when farstial, we confirmed anapase, agate trius to a signafal or sialipal staphylococcal preprobed boing boos summarilyour heartbodn frontal nater and the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the 100% of the"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 253,
+ 206,
+ 304,
+ 210
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 253,
+ 206,
+ 304,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 253,
+ 206,
+ 304,
+ 210
+ ],
+ "type": "text",
+ "content": "Acatat Rechun-Yorim Bogae lo rojctc filocly"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 253,
+ 213,
+ 381,
+ 234
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 253,
+ 213,
+ 381,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 253,
+ 213,
+ 381,
+ 234
+ ],
+ "type": "text",
+ "content": "Our undersonged tandoce boe + ant - or cemoe - aepocemis in a times, bodingly narmabogus haarban. Jnci enwecstic ancoontie fluy fast YOLO pue moebes, 45f rans perennetioles unperenentio- dorensis in reel thane reobexes petarges. ectcylcfom oene trsoute of sucrose princeia of docuta. d. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 255,
+ 236,
+ 384,
+ 273
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 255,
+ 236,
+ 384,
+ 273
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 255,
+ 236,
+ 384,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 255,
+ 236,
+ 384,
+ 273
+ ],
+ "type": "table",
+ "html": "