| Field | Sub-field | Subsub-field | Papers |
| Pre-retrieval | Query Rewriting | Learning to Personalized Query Rewrite | CLE-QR [60], CGF [38], PEARL [80] |
| LLM to Personalized Query Rewrite | Least-to-Most Prompting [173], ERAGent [112], CoPS [174], Agent4Ranking [61], FIG [22], BASES [99] |
| Query Expansion | Tagging-based query expansion | Gossiple [10], Biancalana and Micarelli [12], SoQuES [15], Zhou et al. [172] |
| Else | Lin and Huang [66], Bender et al. [9], Axiomatic PQEC [79], WE-LM [144], PSQE [14], PQEWC [7] |
| Others | Bobo [33], Kannadasan and Aslanyan [52], PSQE [8] |
| Retrieval | Indexing | PEARL [80], KG-Retriever [21], EMG-RAG [137], PGraphRAG [5] |
| Retrieval | Dense Retrieval | MeMemo [138], RECAP [71], LAPDOG [43], Gu et al. [37], PersonalLM [77], UIA [155], XPERT [125], DPSR [157], RTM [11], Pearl [80], MemPrompt [74], EERRA [23], MALP [160], USER-LLM [84], PER-PCS [120] |
| Sparse Retrieval | OPPU [121], PAG [101], Au et al. [5], UniMS-RAG [128], Deng et al. [29], |
| Prompt-based Retrieval | LAPS [50], UniIMP [140], Shen et al. [111] |
| Others | Salem et al. [103], PersonalITM [65], Zhang et al. [165] |
| Post-retrieval | PersonaRAG [156], Pavliukevich et al. [89], UniMS-RAG [128], Salemi and Zamani [106], Zhang et al. [164], AutoCompressors [24], FIT-RAG [76] |
| Generation | Generation from Explicit Preferences | Direct Prompting | P² [49], Character Profiling [154] OpinionQA [107], Kang et al. [51], Liu et al. [67], Cue-CoT [129], TICL [26] |
| Profile-Augmented Prompting | GPG [158], Richardson et al. [101], ONCE [70], LLMTreeRec [163], KAR [145], Matryoshka [58] |
| Personalized-Prompt Prompting | Li et al. [57], RecGPT [166], PEPLER-D [59], GRAPA [94], SGPT [28], PFCL [152] |
| Generation from Implicit Preferences | Fine-tuning-Based Methods | PLoRA [165], LM-P [142], MiLP [165], OPPU [122], PER-PCS [120], Review-LLM [91], UserIdentifier [78], UserAdapter [171], HYDRA [175], PocketLLM [90], CoGenesis [161] |
| Reinforcement Learning-Based Methods | P-RLHF [62], P-SOUPS [47], PAD [20], REST-PG [104], Salemi et al. [103], RewrimerSIRI [57],Kulkarni et al. [54] |
| From RAG to Agent | Personalized Understanding | In user-profile understanding | Xu et al. [148], Abbasian et al. [2], |
| In agent's role understanding | RoleLLM [139], Character-LLM [110], Wang et al. [134], |
| In agent's user-role joint understanding | SocialBench [18], Dai et al. [27], Ran et al. [96], Wang et al. [126], Tu et al. [123], Necko [153] |
| Personalized Planning and Execution | Memory Management | EMG-RAG [137], Park et al. [87], Abbasian et al. [2], RecAgent [133], TravelPlanner+ [114], PersonalWAB [17], VOYAGER [127], MemoeryLLM [136] |
| Tool and API Calling | VOYAGER [127], Zhang et al. [159], PUMA [17], Wang et al. [126], PenetrativeAI [148], Huang et al. [44], [87], MetaGPT [40], OKR-Agent [169] |
| Personalized Generation | Alignment with User Fact | Character-LLM [110], Wang et al. [135], Dai et al. [27] |
| Alignment with User Preferences | Wang et al. [139], Ran et al. [96], Wang et al. [134], Chen et al. [18] |
",
+ "bbox": [
+ 178,
+ 138,
+ 879,
+ 500
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- User Historical Content: Implicit personalization derived from user-generated content, such as chat history, emails, reviews, and social media interactions.",
+ "- Persona-Based User Simulation: The use of LLM-based agents to simulate and generate personalized interactions."
+ ],
+ "bbox": [
+ 174,
+ 530,
+ 879,
+ 578
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Integrating this personalized information at various stages of the RAG and agent workflows enables dynamic alignment with human preferences, thereby making responses more user-centric and adaptive.",
+ "bbox": [
+ 174,
+ 587,
+ 879,
+ 619
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "3 HOW TO ADOPT PERSONALIZATION",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 637,
+ 460,
+ 648
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "We define the process of introducing personalization within the RAG pipeline as follows:",
+ "bbox": [
+ 174,
+ 659,
+ 710,
+ 672
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\ng = \\mathcal {G} (\\mathcal {R} (Q (q, p), C, p), \\text {p r o m p t}, p, \\theta) \\tag {1}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 411,
+ 684,
+ 880,
+ 698
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $p$ denotes personalized information, and the process unfolds in three steps. In the pre-retrieval phase, query processing $(Q)$ refines the query $q$ using personalized information, such as through query rewriting or expansion. During the retrieval phase, the retriever $(\\mathcal{R})$ leverages $p$ to fetch relevant documents from the corpus $(C)$ . Finally, in the generation phase, the retrieved information, combined with $p$ and structured using the given prompt, id fed into the generator $(\\mathcal{G})$ with parameter $\\theta$ to produce the final response $g$ . It is evident that personalized information directly influences multiple stages of the RAG pipeline. In this survey, we consider the agent system as a specialized application of the RAG framework, where personalization is incorporated in a manner similar to the RAG framework.",
+ "bbox": [
+ 174,
+ 708,
+ 880,
+ 825
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 879,
+ 101
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "4",
+ "bbox": [
+ 524,
+ 830,
+ 532,
+ 839
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "image",
+ "img_path": "images/8f9e5a0db9555c1a4ed96186bb98667b2c4fd2abae37751b9dee8a95b60f65c1.jpg",
+ "image_caption": [
+ "Fig. 2. Overview of the personalized pre-retrieval stage."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 114,
+ 125,
+ 816,
+ 383
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4 WHERE TO ADOPT PERSONALIZATION",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 425,
+ 415,
+ 438
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.1 Pre-retrieval",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 446,
+ 238,
+ 459
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.1.1 Definition. Pre-retrieval is a crucial step in information retrieval systems, where the original user query is enhanced or modified before the retrieval process to improve the relevance and quality of the search results, as shown in Figure 2. This process often incorporates additional contextual or personalized information to better align the query with the user's intent. The process can be formalized as follows:",
+ "bbox": [
+ 112,
+ 469,
+ 821,
+ 535
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nq ^ {*} = Q (q, p) \\tag {2}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 426,
+ 545,
+ 821,
+ 560
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $p$ and $q$ denote the personalized information and original query, and $q^{*}$ is the optimized query after query reformulation.",
+ "bbox": [
+ 112,
+ 570,
+ 820,
+ 599
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.1.2 Query Rewriting. Query rewriting in RAG at the pre-retrieval stage refers to the process of reformulating user queries to enhance retrieval effectiveness by improving relevance, disambiguating intent, or incorporating contextual information before retrieving documents from an external knowledge source. The literature on personalized query rewriting can be broadly classified into two primary categories: (1) Direct Personalized Query Rewriting and (2) Auxiliary Personalized Query Rewriting.",
+ "bbox": [
+ 112,
+ 614,
+ 821,
+ 696
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "(1). Direct Personalized Query Rewriting. The first category focuses on personalized query rewriting by using direct models. For example, Cho et al. [25] presents a personalized search-based query rewrite system for conversational AI that addresses user-specific semantic and phonetic errors. Nguyen et al. [82] apply reinforcement learning techniques to improve query rewriting in online e-commerce systems, leveraging distilled LLMs for personalized performance. CLE-QR [60] explores query rewriting in Taobao's search engine to enhance user satisfaction through customized query adaptation. CGF [38] introduces a constrained generation framework that allows for more flexible and personalized query rewriting in conversational AI. Li et al. [57] investigate learning methods to rewrite prompts for personalized",
+ "bbox": [
+ 112,
+ 709,
+ 821,
+ 827
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 114,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "5",
+ "bbox": [
+ 464,
+ 830,
+ 470,
+ 839
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "text generation, improving the relevance and engagement of AI-generated content. Additionally, PEARL [80] discusses personalizing large language model-based writing assistants through the integration of generation-calibrated retrievers, enhancing AI-generated content.",
+ "bbox": [
+ 174,
+ 121,
+ 880,
+ 170
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "(2). Auxiliary Personalized Query Rewriting. The second category emphasizes personalized query rewriting by using auxiliary mechanisms, such as retrieval, reasoning strategies, and external memory. Zhou et al. [173] propose a least-to-most prompting strategy that aids in complex reasoning within LLMs, which can be adapted for personalized text generation. ERAGent [112] enhance retrieval-augmented LLMs to improve personalization, efficiency, and accuracy, indirectly supporting personalized query rewriting for content generation. CoPS [174] integrate LLMs with memory mechanisms to create more personalized search experiences, which also influences content generation through better query understanding. Further, Agent4Ranking [61] employs multi-agent LLMs to perform semantic robust ranking, including personalized query rewriting to improve search rankings. FIG [22] combine graph-based methods with LLMs to query rewrite, improving personalized content generation and conversational interactions. Lastly, BASES [99] employ LLM-based agents to simulate large-scale web search user interactions, contributing to the development of personalized query rewriting strategies for content generation.",
+ "bbox": [
+ 174,
+ 185,
+ 883,
+ 373
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.1.3 Query Expansion. Query expansion enhances retrieval systems by expanding a user's original query with additional terms, synonyms, or refined structure to better capture intent. This improves the relevance and scope of retrieved documents. Recent advancements in LLMs have reinvigorated this field [46, 48, 132], leveraging their comprehension and generation abilities to expand queries using encoded knowledge or external retrieval, with notable success. Personalized query expansion, a subset, incorporates user-specific data to tailor results, boosting performance and customizing the search experience.",
+ "bbox": [
+ 174,
+ 388,
+ 883,
+ 491
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(1). Tagging-based Query Expansion. By 2009, studies began incorporating tagging information to enhance personalized query expansion. For instance, Gossle [10] introduced the TagMap and TagRank algorithms, which dynamically selected tags from personalized networks constructed using the cosine similarity of user-item tag distances, improving recall performance. Similarly, Biancalana and Micarelli [12] recorded user queries and visited URLs, leveraging social bookmarking to extract relevant tags and build a personalized three-dimensional co-occurrence matrix. Based on this, multiple semantically categorized expanded queries were generated to better reflect user interests. Further advancements include SoQuES [15], which integrated tag semantic similarity with social proximity, and a graph-based approach [172] that utilized Tag-Topic models and pseudo-relevance feedback for term weighting, tailoring the expansion process to individual user preferences.",
+ "(2). Else. Apart from tagging-based techniques, early research on Personalized Query Expansion primarily focused on modeling user personalization based on search history [66], social networks, or preferences derived from friendship networks [9]. The Axiomatic PQEC framework [79] formalized expansion rules using both local (user behavior-driven) and social (network-driven) strategies. In 2017, WE-LM [144] advanced this paradigm by modeling multi-relational networks with word embeddings across tag-word relationships, refining associations through affinity graphs. Later, PSQE [14] further improved tagging-based methods using utf-iuf user profiling, integrating a tag similarity graph with user profiles in the online phase to compute expansion terms relevant to user interests in real-time, achieving dynamic personalized expansion. In addition, PQEWC [7] leveraged clustering and contextual word embeddings to optimize query expansions dynamically."
+ ],
+ "bbox": [
+ 174,
+ 505,
+ 883,
+ 827
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 879,
+ 101
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "6",
+ "bbox": [
+ 524,
+ 830,
+ 532,
+ 840
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/431fadc2fb6b837e6448de02affe85f7a85af23cf964b4cebff641487aab910d.jpg",
+ "image_caption": [
+ "Fig. 3. Overview of the personalized retrieval stage."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 117,
+ 117,
+ 821,
+ 281
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.1.4 Others. Besides query rewriting and query expansion, other personalized query-related research focuses on areas like query disambiguation and query auto-completion [116]. Bobo [33] allows users to input contextual terms reflecting their domain knowledge. In 2019, a method [52] applied fastText embeddings from recent queries to rank candidates. In addition, PSQE [8] employed synthetic user profiles from Wikipedia and word2vec embeddings for query disambiguation.",
+ "bbox": [
+ 114,
+ 332,
+ 821,
+ 416
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.1.5 Discussion. While both query rewriting and query expansion aim to align user input with system understanding to enhance retrieval quality, their roles in personalization differ in fundamental ways. Understanding the distinct operational characteristics and application scenarios of each technique is essential for designing effective personalized retrieval systems. The key takeaways are listed as follows:",
+ "bbox": [
+ 114,
+ 430,
+ 821,
+ 496
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- Query rewriting is most beneficial when the original query is ambiguous, underspecified, or misaligned with retrieval intents, particularly in conversational or multi-turn settings.",
+ "- Query expansion is most effective when the original query is relevant but incomplete - i.e., when it needs to be semantically broadened to cover additional relevant concepts."
+ ],
+ "bbox": [
+ 114,
+ 506,
+ 818,
+ 571
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.2 Retrieval",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 590,
+ 215,
+ 603
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.2.1 Definition. The retrieval process involves finding the most relevant documents $D^{*}$ from a corpus $C$ based on a query $q^{*}$ , as shown in Figure 3. To incorporate personalization, additional user-specific information $p$ is integrated into the retrieval function $\\mathcal{R}$ . This allows the retrieval process to tailor the selected documents to align with individual user preferences or contexts, thereby enhancing the relevance and personalization of the generated outputs.",
+ "bbox": [
+ 114,
+ 612,
+ 818,
+ 679
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nD ^ {*} = \\mathcal {R} (q ^ {*}, C, p) \\tag {3}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 413,
+ 688,
+ 821,
+ 705
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "In the retrieval process, personalization can primarily be introduced by focusing on three steps: indexing, retrieval, and post-retrieval. These steps ensure efficient and accurate retrieval of relevant documents or knowledge, while tailoring the process to individual user preferences. Below, we provide a detailed explanation of each step.",
+ "bbox": [
+ 114,
+ 714,
+ 818,
+ 763
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.2.2 Indexing. Indexing organizes knowledge base data into a structured format to facilitate efficient retrieval. Within the RAG pipeline, documents are either chunked or entirely encoded into representations before being integrated into searchable systems [30, 117]. Conventional encoding methods employ either sparse encoding techniques (e.g.,",
+ "bbox": [
+ 114,
+ 777,
+ 821,
+ 825
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 114,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "page_number",
+ "text": "7",
+ "bbox": [
+ 462,
+ 830,
+ 472,
+ 839
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "TF-IDF [95], BM25 [102]) or dense encoding approaches leveraging pre-trained models, such as BERT [1], Siamese Encoders [98], or LLM-based encoders [64, 147].",
+ "bbox": [
+ 174,
+ 121,
+ 880,
+ 152
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "To introduce personalization at the indexing stage, PEARL [80] generates user embeddings by encoding personal history data with models like DeBERTa. These embeddings are subsequently clustered to create personalized shared indices. Other approaches integrate knowledge graphs into indexing to enhance retrieval performance. For example, KG-R retriever [21] employs a Hierarchical Index Graph, consisting of a knowledge graph layer and a collaborative document layer, to improve RAG retrieval. EMG-RAG [137] incorporates personalized memory within an editable knowledge graph, enabling dynamic retrieval. Similarly, PGraphRAG [5] leverages user-centric knowledge graphs to enhance personalization in retrieval tasks.",
+ "bbox": [
+ 174,
+ 156,
+ 883,
+ 273
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "4.2.3 Retrieval. The Retrieval step matches a user query with the indexed knowledge base to fetch relevant candidates. It can be broadly categorized into four different types: (1) Dense Retrieval, (2) Sparse Retrieval, (3) Prompt-based Retrieval, and (4) Others.",
+ "bbox": [
+ 174,
+ 287,
+ 883,
+ 337
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "(1). Dense Retrieval. Dense retrieval methods often use vector embeddings and similarity metrics (e.g., cosine similarity) and achieve personalization by encoding user preferences, context, or interactions into query or document embeddings, enabling tailored results through similarity-based matching. For instance, MeMemo [138] retrieves personalized information by matching user-specific embeddings with document vectors, focusing on private, on-device text generation. Similarly, RECAP [71] and LAPDOG [43] enhance personalized dialogue generation by encoding queries and user profiles as dense vectors and retrieving top-N results, ensuring user-specific context drives the responses. In chatbots, Gu et al. [37] integrates conversational context and user profiles to align retrieved responses with user personas. PersonalM [77] employs group-wise contrastive learning, training its retrieval model to align user queries with domain-specific text fragments, thereby improving personalization. UIA [155] employs dual encoders to retrieve documents tailored to user preferences. XPERT [125] incorporates temporal events and user interactions into embeddings, enabling large-scale retrieval across millions of items.",
+ "bbox": [
+ 174,
+ 351,
+ 883,
+ 537
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Dense retrieval also enhances specific applications like e-commerce, medical assistance, and language models. DPSR [157] and RTM [11] encode user queries and product information to personalize product searches dynamically. Pearl [80] and MemPrompt [74] retrieve personalized content by leveraging historical user data and memory-assisted mechanisms. EERRA [23] uses review embeddings as dense queries for recommendations. In medical assistance, MALP [160] and User-LLM [84] integrate short- and long-term user interactions into embeddings for contextualized, personalized responses. Finally, PER-PCS [120] retrieves relevant information using individual user histories, enhancing the personalization capabilities of large language models.",
+ "bbox": [
+ 174,
+ 541,
+ 885,
+ 659
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "(2). Sparse Retrieval. Sparse retrieval methods often rely on term-based matching (e.g., BM25) and apply personalization by assigning higher weights to terms or keywords that are more relevant to the user. OPPU [121] uses the BM25 algorithm to select the k most relevant records from the user's historical data for the current query. Similarly, PAG [101] incorporates user input and profiles to enhance summarization and retrieval, aligning sparse representations with personalization objectives for large language models. Au et al. [5] uses BM25 search algorithms to find entries related to the target user or neighboring users through the graph structure. UniMS-RAG [128] combines sparse and dense retrieval by leveraging multi-source knowledge, such as dialogue context and user images, to refine personalized responses in dialogue systems. Lastly, Deng et al. [29] apply sparse retrieval to support fact-based queries, considering user queries and preferences to enhance answer generation for e-commerce applications.",
+ "bbox": [
+ 174,
+ 672,
+ 885,
+ 827
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 879,
+ 101
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "page_number",
+ "text": "8",
+ "bbox": [
+ 524,
+ 830,
+ 534,
+ 840
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(3). Prompt-based Retrieval. Prompt-based retrieval leverages prompts to guide retrieval from the model or external sources and introduces personalization by crafting user-specific prompts that guide the retrieval process. These prompts may include explicit user preferences, historical interactions, or detailed instructions that reflect the user's unique requirements. By embedding this personalized context directly into the prompt, the retrieval process can dynamically adjust to capture and return results that are most relevant to the user. LAPS [50] focuses on multi-session conversational search by storing user preferences and dialogue context, then using prompts to retrieve relevant information tailored to the user's biases and categories of interest. UniMP [140] employs user interaction histories as input to prompt-based retrieval, enabling personalized recommendations for multi-modal tasks, such as vision-language applications, by aligning prompts with user behavioral data. In contrast, Shen et al. [111] explores the use of LLMs to extract empathy and narrative styles from user-provided stories, but this work primarily focuses on style extraction and does not explicitly involve a retrieval component.",
+ "(4). Others. Reinforcement learning-based retrieval personalizes the process by optimizing retrieval policies based on user feedback, learning user preferences over time to adjust strategies. Salemi et al. [103] combines models like BM25, RbR, and dense retrieval, refining them with reinforcement learning (RL) and knowledge distillation (KD) to adapt to user profiles for personalized outputs. Parameter-based retrieval leverages pre-trained model parameters to implicitly store and retrieve user-specific information, allowing direct retrieval from the model without traditional indices. PersonalTM [65] generates document identifiers (Document IDs) using a Transformer model, encoding query, history, and document relationships into its parameters for personalization. Similarly, Zhang et al. [165] uses parameterized representations to integrate user queries and histories, tailoring responses to individual preferences."
+ ],
+ "bbox": [
+ 114,
+ 121,
+ 823,
+ 457
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "4.2.4 Post-retrieval. Current Post-Retrieval methods primarily focus on refining retrieved documents or responses to improve relevance and coherence, current methodologies could be categorized into three parts (1) Re-ranking, (2) Summarization, and (3) Compression.",
+ "bbox": [
+ 114,
+ 470,
+ 820,
+ 518
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(1). Re-ranking. Re-ranking enhances personalized content generation by prioritizing more relevant documents at the top. PersonaRAG [156] extends RAG by integrating user-centric agents, such as the Live Session Agent and the Document Ranking Agent, to refine document ranking and improve overall performance. Pavliukevich et al. [89] propose a cross-encoder BERT model for re-ranking external knowledge within a personalized context. UniMS-RAG [128] introduces a scoring mechanism that evaluates retrieved documents and outputs by optimizing the retriever. Besides, it includes an evidence attention mask, enabling re-ranking during inference and applying it to personalized datasets. Salemi and Zamani [106] present an iterative approach to optimizing ranking results based on the expectation-maximization algorithm, with performance validated in personalized scenarios.",
+ "(2). Summarization. Summarization refers to the process of summarizing retrieved information to enhance performance. For instance, Zhang et al. [164] introduced a role-playing agent system to summarize retrieved history in order to improve the final Personalized Opinion Summarization process.",
+ "(3). Compression. Compression involves condensing embeddings or retrieved content to enhance efficiency and effectiveness. Approaches like AutoCompressor [24] compress contextual embeddings into shorter semantic representations, and FIT-RAG [76] introduces a self-knowledge recognizer along with a sub-document-level token reduction mechanism to minimize tokens within RAG pipeline. However, few studies have specifically explored personalized fields, highlighting a promising direction for future research."
+ ],
+ "bbox": [
+ 114,
+ 532,
+ 823,
+ 825
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 114,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "page_number",
+ "text": "9",
+ "bbox": [
+ 464,
+ 830,
+ 472,
+ 839
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "4.2.5 Discussion. Indexing, retrieval, and post-retrieval methods each play a critical role in ensuring efficient and personalized information processing, with specific applications and trade-offs. Indexing focuses on organizing knowledge bases for efficient retrieval, using techniques such as sparse encoding methods like TF-IDF and BM25, which are efficient but limited in understanding semantics, and dense encoding methods like BERT and DeBERTa, which provide better semantic understanding but require significant computational resources. These methods are widely used in tasks like question answering and personalized recommendation systems. Retrieval involves matching user queries with relevant documents and can be categorized into dense retrieval, which provides high semantic understanding and personalization but is computationally expensive; sparse retrieval, which is efficient and interpretable but less capable of handling semantics; prompt-based retrieval, which is highly flexible and adaptable to user needs but requires careful engineering of prompts; and advanced methods like reinforcement learning-based approaches, which dynamically adapt to user feedback but are complex to implement. This step is essential in applications like personalized dialogue systems, search engines, and e-commerce. Post-retrieval methods refine retrieved results to enhance relevance and coherence through re-ranking, which improves personalization and prioritizes relevant content but increases computational overhead; summarization, which simplifies complex information for better user understanding but risks losing critical details; and compression, which reduces computational costs by condensing information but remains underexplored in personalized contexts. Together, these methods provide a comprehensive pipeline for delivering efficient, relevant, and personalized outputs, balancing their strengths in semantic understanding, relevance, and flexibility with challenges related to computational costs and implementation complexity.",
+ "bbox": [
+ 174,
+ 121,
+ 883,
+ 431
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "4.3 Generation",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 446,
+ 289,
+ 458
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "4.3.1 Definition. Personalized generation incorporates user-specific retrieved documents $D^{*}$ , task-specific prompt prompt, and user preference information $p$ via the generator $\\mathcal{G}$ parameterized by $\\theta$ to produce tailored content $g^{*}$ aligned with individual preference, where the flow is shown in Figure 4. The generation process can be formulated as",
+ "bbox": [
+ 174,
+ 467,
+ 883,
+ 517
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\ng ^ {*} = \\mathcal {G} (D ^ {*}, \\text {p r o m p t}, p, \\theta). \\tag {4}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 447,
+ 526,
+ 880,
+ 542
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "Personalized generation can be achieved by incorporating explicit and implicit preferences. Explicit preference-driven methodologies utilize direct input signals (e.g., $D^{*}$ , prompt, and $p$ ), to tailor outputs to specific user preferences. Conversely, implicit preference-encoded approaches embed personalized information within the parameters $\\theta$ of the generator model, during training, thereby facilitating preference alignment without the necessity for explicit runtime inputs.",
+ "bbox": [
+ 174,
+ 551,
+ 883,
+ 635
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "4.3.2 Generation from Explicit Preferences. Integrating explicit preferences into LLMs facilitates personalized content generation. Explicit preference information encompasses user demographic information (e.g., age, occupation, gender, location), user behavior sequences (reflecting historical behavioral patterns), and user historical output texts (capturing writing style and tone preferences). The injection of explicit preferences for personalized generation can be categorized into three types: (1) Direct-integrated Prompting, (2) Summary-augmented Prompting, and (3) Adaptive Prompting.",
+ "bbox": [
+ 174,
+ 647,
+ 883,
+ 750
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "(1). Direct-integrated Prompting. Integrating user explicit preferences into language models through prompting enables the prediction of users' intent and behavioral patterns, facilitating personalized content generation. For instance, $\\mathrm{P}^2$ [49], Character Profiling [154], and OpinionQA [107] integrate personalized data into LLMs through prompting for role-playing task, thereby aligning the model's responses with specified user profiles. Kang et al. [51] and Liu et al. [67]",
+ "bbox": [
+ 174,
+ 761,
+ 883,
+ 828
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 879,
+ 101
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "page_number",
+ "text": "10",
+ "bbox": [
+ 522,
+ 830,
+ 535,
+ 840
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "image",
+ "img_path": "images/684289db43c065be977860260a5ca5599fa5f676c9663e41befff2d9bcc9c089.jpg",
+ "image_caption": [
+ "Fig. 4. Overview of the personalized generation stage."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 135,
+ 119,
+ 802,
+ 292
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "integrate interaction histories into LLMs via prompting to predict user rating for candidate items. Cue-CoT [129] employs chain-of-thought reasoning to infer user needs from contextual cues, enabling personalized responses to in-depth dialogue questions. Additionally, TICL [26] proposes a trial-and-error framework that critiques initial LLM-generated responses, derives explanations and integrates these negative examples into prompts to improve personalization alignment.",
+ "bbox": [
+ 114,
+ 349,
+ 821,
+ 434
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(2). Summary-augmented Prompting. Direct integration of personalized information via prompting struggles with ambiguous intent signals: Lengthy interaction histories introduce noise that obscures critical behavioral patterns [69], while sparse behavioral data lacks sufficient context for LLMs to derive meaningful user preferences. To address these issues, recent approaches focus on summarizing user personalized intents and integrating them into prompts. For instance, GPG [158] extracts key user habits and preferences from personal contexts, enabling fine-grained personalization. Similarly, LLMs are employed to generate task-specific summaries of user preferences, enhancing retrieval-augmented personalized generation capabilities [101]. In recommendation systems, ONCE [70], LLMTreeRec [163], and KAR [145] leverage historical user-item interactions to summarize user preferences. Furthermore, Matryoshka [58] generates user preference summaries by dynamically retrieving and synthesizing historical data.",
+ "(3). Adaptive Prompting. Manually designing personalized prompts demands both expert knowledge and significant labor, motivating the development of automated methods for personalized prompt generation. For example, Li et al. [57] trains a personalized prompt rewriter via supervised and reinforcement learning. RecGPT [166] and PEPLER-D [59] leverage prompt tuning to generate personalized prompts, enhancing sequential and explainable recommendations, respectively. GRAPA [94] integrates semantic and collaborative signals from user-item interaction graphs with graph neural networks to generate context-aware personalized prompts. SGPT [28] employs prompt tuning to jointly model common and group-specific patterns, bridging generalized and personalized federated learning paradigms. Furthermore, PFCL [152] achieves multi-granularity human preference modeling: coarse-grained prompts distill shared knowledge, while fine-grained prompts adapt to individual user characteristics."
+ ],
+ "bbox": [
+ 112,
+ 446,
+ 823,
+ 765
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "4.3.3 Generation from Implicit Preferences. Unlike explicit preference modeling, which captures user preferences through textual input, implicit preference-based methods incorporate personalization through internal parameters. This personalization is achieved either through Parameter-Efficient Fine-tuning (PEFT) techniques, such as LoRA [42],",
+ "bbox": [
+ 114,
+ 777,
+ 823,
+ 827
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 114,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "page_number",
+ "text": "11",
+ "bbox": [
+ 462,
+ 830,
+ 473,
+ 840
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "text",
+ "text": "or reinforcement learning-based approaches for preference alignment [20, 57]. Based on these strategies, we classify existing methods into two categories: (1) Fine-tuning-Based Methods and (2) Reinforcement Learning-Based Methods.",
+ "bbox": [
+ 174,
+ 121,
+ 883,
+ 154
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "(1). Fine-tuning Based Methods. For fine-tuning methods, LoRA is the most widely adopted since it is resource-efficient and enables rapid adaptation without compromising model performance. PLoRA [165] introduces a personalized knowledge integration framework that combines task-specific LoRA with user-specific knowledge. Similarly, LM-P [142] personalizes information via LoRA by incorporating User ID as a personalization factor. MiLP [165] employs Bayesian optimization to determine the optimal personalization injection configuration, including LoRA settings, to effectively capture and utilize user-specific information. OPPU [122] and PER-PCS [120] follow a similar approach, leveraging user history data for fine-tuning LoRA-based personalization. However, PER-PCS differs by incorporating a gating module that selects the appropriate LoRA, enabling fine-grained personalization. Additionally, Review-LLM [91] integrates LoRA for supervised fine-tuning in the task of personalized review generation.",
+ "bbox": [
+ 174,
+ 166,
+ 883,
+ 319
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "Beyond LoRA-based approaches, alternative pipelines have been proposed for personalized generation. UserIdentifier [78] introduces a user-specific identifier, significantly reducing training costs while enhancing personalized demonstration. UserAdapter [171] proposes user-independent prefix embeddings, leveraging prefix tuning for personalization. Meanwhile, HYDRA [175] achieves implicit personalization by training user-specific headers. Recently, researchers have also explored fine-tuning personalized model on edge devices [90] and collaborative learning between small and large language models to enable more personalized generation [161].",
+ "bbox": [
+ 174,
+ 321,
+ 883,
+ 422
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "(2). Reinforcement Learning Based Methods. Apart from fine-tuning based methods, recent research has explored reinforcement learning based techniques to personalize text generation by aligning outputs with user preferences. P-RLHF [62] has been proposed to jointly learn a user-specific and reward model to enable text generation that aligns with a user's styles or criteria. P-SOUPS [47] models multiple user preferences as a Multi-Objective Reinforcement Learning (MORL) problem, decomposing preferences into multiple dimensions, each trained independently. PAD [20] aligns text generation with human preferences during inference by utilizing token-level personalized rewards to guide the decoding process. REST-PG [104] introduces a framework that trains large language models to reason over personal data during response generation. This approach first generates reasoning paths to enhance the LLM's reasoning ability and then employs Expectation-Maximization Reinforced Self-Training to iteratively refine the model based on its high-reward outputs. Additionally, Salemi et al. [103] incorporate reinforcement learning into the RAG pipeline to improve retrieval accuracy, thereby enhancing the personalization of generated content. Other applications include RewriterSIRI [57], which has been introduced to generate text via RL-based personalized prompt rewriting using API-based LLMs, and Kulkarni et al. [54], who explore the use of reinforcement learning to optimize RAG for improving the relevance and coherence of chatbot responses in specialized domains, ultimately enhancing user satisfaction and engagement.",
+ "bbox": [
+ 174,
+ 436,
+ 883,
+ 694
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "text",
+ "text": "4.3.4 Discussion. Personalized generation can be adopted via both explicit and implicit preference injection, yet they exhibit distinct characteristics that make them suitable for different scenarios. In explicit preference-based generation, personalization is clearly defined through user profile descriptions, contextual information, and similar inputs, which are incorporated into generators via prompts. A key advantage of this approach is explainability, as the personalized information is explicitly provided and easily traceable. Despite leveraging provided preferences and internal knowledge, explicit preference injection's personalization is constrained by model capabilities and irrelevant information interference. In contrast, implicit preference-based generation internalizes personalized information into",
+ "bbox": [
+ 174,
+ 708,
+ 883,
+ 825
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 880,
+ 101
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "page_number",
+ "text": "12",
+ "bbox": [
+ 522,
+ 830,
+ 535,
+ 839
+ ],
+ "page_idx": 11
+ },
+ {
+ "type": "image",
+ "img_path": "images/c1f227eab52914058e52cd6428ef57f5faa6d829581489e901866e44230f1223.jpg",
+ "image_caption": [
+ "Fig. 5. Overview of transition from personalized RAG to personalized agent."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 122,
+ 118,
+ 816,
+ 489
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "the generator's parameters through scene-specific personalized data, thereby adapting the model for more fine-grained personalization. However, these methods typically incur substantial training and computational costs, as they require fine-tuning the generator's internal parameters. Therefore, selecting between these approaches should be guided by the specific application scenario and resource constraints.",
+ "bbox": [
+ 114,
+ 555,
+ 821,
+ 621
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "4.4 From RAG to Agent",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 642,
+ 284,
+ 656
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "4.4.1 Definition. A personalized LLM-based agent is a system designed to dynamically incorporate user context, memory, and external tools or APIs to support highly personalized and goal-oriented interactions [19, 45, 146], and solve problems in a goal-oriented manner [63, 113]. From the previously introduced stages of RAG, we observe that the evolution of personalized RAG reveals a structural convergence with agent architectures. We analyze them from three key perspectives:",
+ "bbox": [
+ 112,
+ 662,
+ 821,
+ 748
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "text",
+ "text": "- Personalized Understanding: This phase within the agent parallels the query understanding and rewriting process of RAG as outlined in Section 4.1. However, it extends beyond static semantic parsing by incorporating dynamic user profiling [139] and role modeling [110]. This integration enables the agent to dynamically align interactions with implicit user preferences, facilitating personalized responses and task-specific adaptations [96].",
+ "bbox": [
+ 114,
+ 758,
+ 821,
+ 825
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 114,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "page_number",
+ "text": "13",
+ "bbox": [
+ 462,
+ 830,
+ 473,
+ 840
+ ],
+ "page_idx": 12
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- Personalized Planning and Execution: This phase in agents mirrors RAG's retrieval process in Section 4.2 yet it advances beyond static document retrieval by incorporating real-time memory management [87] and sophisticated tool and API calling [127]. This approach ensures the dynamic alignment of external knowledge with personalized constraints, such as integrating medical history in healthcare agents [2], to deliver context-aware and user-specific outcomes.",
+ "- Personalized Generation: This phase in agents mirrors RAG's generative process in Section 4.3 but transcends static template-based generation by integrating user preference and fact alignment. Agents dynamically enforce user preferences and ensure fact consistency through role-specific mechanisms (e.g., social adaptability in conversational agents [2]), enabling outputs to evolve in harmony with personalized and situational constraints rather than relying solely on predefined generative frameworks."
+ ],
+ "bbox": [
+ 174,
+ 121,
+ 880,
+ 292
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "In general we frame agent architectures as \"personalized RAG++\", where persistent memory [137] replaces static indexes, and tool APIs [17] serve as dynamic knowledge connectors, enabling complicated, human-aligned interactions beyond one-shot retrieval, as shown in Figure 5. This progression highlights that as RAG systems incorporate deeper personalization—requiring user-state tracking, adaptive tool usage, and context-aware generation, they inherently adopt agent-like capabilities.",
+ "bbox": [
+ 174,
+ 300,
+ 880,
+ 383
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "4.4.2 Personalized Understanding. Personalized understanding refers to an agent's ability to accurately interpret user inputs by integrating user intent recognition and contextual analysis. This process ensures interactions that are both meaningful and contextually appropriate. The rationale behind this classification lies in its capacity to address three core aspects of understanding: recognizing user intent, analyzing context, and leveraging user profiles. Each of these aspects plays a distinct role in improving the agent's performance.",
+ "bbox": [
+ 174,
+ 398,
+ 880,
+ 481
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(1). User-profile Understanding. In user-profile understanding, an agent's personalized ability primarily depends on its capacity to accurately model and understand the user's preferences, context, and intentions. Xu et al. [148] proposes a framework in which LLMs are designed to understand the physical world, thereby facilitating a deeper connection between the agent and its environment, which is essential for accurate task execution. Abbasian et al. [2] further expands this understanding by emphasizing the importance of personalization in health agents, where the user's profile directly influences the behavior and decisions of the agent. This user understanding is foundational to ensuring that the AI agent performs tasks in a way that aligns with individual user needs.",
+ "(2). Role Understanding. In agent's role understanding, the role of the agent within these environments is also crucial. Recent studies focus on enhancing role-playing capabilities within LLMs. Wang et al. [139] introduce RoleLLM, a benchmark that aims to elicit and refine the role-playing abilities of LLMs, demonstrating how role understanding influences agent performance in conversational tasks. Similarly, Shao et al. [110] present Character-LLM, a trainable agent framework for role-playing, which tailors its responses based on predefined roles. Wang et al. [134] introduce a method for evaluating personality fidelity in role-playing agents through psychological interviews, aiming to enhance the realism and consistency of AI-driven characters. This role understanding allows for more contextually appropriate interactions, increasing the relevance and utility of AI agents across various applications.",
+ "(3). User-role Joint Understanding. In agent's user-role joint understanding, the intersection of user and role understanding is explored through frameworks that evaluate and enhance the social and personality aspects of LLMs. SocialBench Chen et al. [18] provides a sociality evaluation framework for role-playing agents. Dai et al. [27], and"
+ ],
+ "bbox": [
+ 174,
+ 494,
+ 880,
+ 825
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 880,
+ 101
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "page_number",
+ "text": "14",
+ "bbox": [
+ 522,
+ 830,
+ 535,
+ 840
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "[96] extend this by incorporating multi-modal data and personality-indicative information, respectively, which allows agents to better adapt to both user and role understanding in dynamic environments. Furthermore, Wang et al. [126] offers a perspective on how role and environment understanding can improve user experience. Tu et al. [123] contribute by providing a benchmark specifically for evaluating role-playing agents in the Chinese context, adding a cultural dimension to role understanding. Finally, Neeko [153] further advances role-based interactions.",
+ "bbox": [
+ 114,
+ 121,
+ 821,
+ 204
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "4.4.3 Personalized Planning and Execution. Personalized planning and execution refer to the process of designing and implementing strategies or actions that are specifically tailored to an individual's unique context, and goals [44, 87, 114, 159]. It requires agents to dynamically integrate long-term memory, real-time reasoning, and external tool utilization [40, 41, 169], as demonstrated in healthcare decision support [2] and travel planning scenarios [17]. We analyze two fundamental components that enable this personalization in the following.",
+ "bbox": [
+ 114,
+ 220,
+ 823,
+ 305
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(1). Memory Management. Effective memory systems allow agents to integrate users' historical preferences, behavioral patterns, and contextual habits, enhancing their ability to make planning and tailor interactions to user-specific needs [17, 127, 136]. The EMG-RAG framework [137] combines editable memory graphs with retrieval-augmented generation to maintain dynamic user profiles, while Park et al. [87] implements memory streams and periodic reflection mechanisms to simulate human-like behavior. In healthcare applications, Abbasian et al. [2] integrates multimodal user data through specialized memory modules to optimize treatment recommendations. For recommendation systems, RecAgent [133] employs hierarchical memory structures to model user interaction patterns across multiple domains. Recent advances like TravelPlanner+ [114] demonstrate how memory-augmented LLMs achieve higher relevance in personalized itinerary generation compared to generic planners.",
+ "(2). Tool and API Calling. The integration of external tools expands agents' capabilities beyond pure linguistic reasoning, enabling agents to interact with users and perform personalized tasks [17, 126, 127, 148, 159]. For instance, VOYAGER [127] establishes a paradigm for lifelong skill acquisition through automatic API curriculum learning and skill library construction. In robotics, Zhang et al. [159] develops a bootstrapping framework where LLMs guide robots in tool-mediated skill discovery, enabling a high success rate in novel object manipulation tasks. The PUMA framework [17] demonstrates how personalized web agents can achieve performance gains in e-commerce tasks through adaptive API orchestration. For mobile interaction, Wang et al. [126] implements few-shot tool learning to handle diverse UI operations with minimal training data. These approaches highlight the importance of tool grounding mechanisms [44] that translate linguistic plans into executable API sequences while maintaining personalization constraints."
+ ],
+ "bbox": [
+ 114,
+ 319,
+ 823,
+ 641
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "This synthesis highlights that modern agent systems achieve enhanced personalization through two primary strategies: 1) Memory-augmented architectures, which leverage editable memory graphs [137], reflection mechanisms [87], and hierarchical memory structures [133] to dynamically adapt to user preferences across various domains; and 2) Tool and API integration, which expand agent capabilities by balancing generalization with specialization. Future work may explore improving the contextual relevance and adaptability of memory systems while optimizing real-time tool interaction for seamless task execution.",
+ "bbox": [
+ 114,
+ 643,
+ 823,
+ 743
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "4.4.4 Personalized Generation. Based on the foundation of personalized planning and execution mechanisms, which enable agents to adapt strategies to user-specific contexts [44, 159], the next critical concern lies in personalized generation. This capability ensures that generated outputs not only align with factual correctness but also resonate with users' unique preferences, personality traits, and situational needs. Personalized generation bridges the gap between",
+ "bbox": [
+ 114,
+ 758,
+ 823,
+ 825
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 114,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "page_number",
+ "text": "15",
+ "bbox": [
+ 462,
+ 830,
+ 473,
+ 840
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "adaptive reasoning and human-aligned outcomes, allowing agents to produce contextually relevant and emotionally appropriate responses.",
+ "bbox": [
+ 174,
+ 121,
+ 880,
+ 154
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "(1). Alignment with User Fact. Alignment with User Fact emphasizes the accuracy, consistency, and factual grounding of personalized responses, ensuring they remain trustworthy across diverse user interactions. This is particularly challenging in personalized agents, where maintaining character authenticity while avoiding hallucinations requires balancing creativity with factual adherence. Recent advances address these challenges through improved training frameworks and evaluation metrics. For instance, Character-LLM [110] integrates memory-augmented architectures to reduce hallucinations while preserving character-specific traits. Wang et al. [135] investigate quantization effects on personality consistency in edge-deployed agents and stabilize outputs under computational constraints. Dai et al. [27] ensures multimodal consistency (text-image) in role-playing. These works highlight the importance of architectural innovations and rigorous evaluation in achieving reliability.",
+ "(2). Alignment with User Preferences. Alignment with user preferences ensures that generated outputs reflect individualized personalities, values, and interaction styles. This requires agents to dynamically interpret implicit user cues and adapt responses accordingly. Wang et al. [139] benchmarks role-specific alignment. Ran et al. [96] improves personality fidelity via psychological scale datasets. Wang et al. [134] quantifies alignment via psychological interviews. Chen et al. [18] evaluates social adaptability in conversations."
+ ],
+ "bbox": [
+ 174,
+ 176,
+ 885,
+ 439
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "4.4.5 Discussion. The architectural evolution from RAG to personalized agents introduces significant advancements in human-AI interaction but also surfaces critical challenges that warrant further investigation.",
+ "bbox": [
+ 174,
+ 460,
+ 880,
+ 492
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Personalized Understanding, while enabling interpretation of user intent and context, faces limitations in real-time adaptability and generalization. Current approaches like RoleLLM [139] and Character-LLM [110] demonstrate robust role-specific comprehension but struggle with dynamic user state tracking, particularly when handling evolving preferences or multi-session interactions. Furthermore, cultural specificity in benchmarks like CharacterEval [123] reveals gaps in global applicability, as agents trained on region-specific data often fail to generalize across diverse sociocultural contexts. Future work could explore hybrid architectures that combine continuous learning mechanisms with privacy-preserving federated learning to address these adaptability constraints while maintaining user trust.",
+ "bbox": [
+ 174,
+ 494,
+ 883,
+ 613
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Personalized Planning and Execution, achieves remarkable task specialization through memory management and tool integration, yet suffers from scalability issues in complex environments. While frameworks like EMG-RAG [137] and VOYAGER [127] effectively manage user-specific constraints, their reliance on predefined API taxonomies limits emergent tool discovery in novel scenarios. The \"cold-start\" problem persists in domains requiring rapid skill acquisition, as seen in healthcare applications [2], where delayed API responses can compromise decision-making efficacy. A promising direction involves developing meta-reasoning architectures that dynamically prioritize memory recall versus tool invocation based on situational urgency and confidence thresholds.",
+ "bbox": [
+ 174,
+ 617,
+ 883,
+ 736
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Personalized Generation balances factual accuracy with preference alignment but risks over-fitting, where excessive finetuning to user profiles may reinforce cognitive biases. Techniques address surface-level alignment but lack mechanisms for ethical boundary detection. For instance, agents might inadvertently propagate harmful stereotypes when mirroring user preferences without critical oversight. Future systems could integrate value-aligned reinforcement learning with human-in-the-loop validation to preserve authenticity while preventing detrimental customization.",
+ "bbox": [
+ 174,
+ 738,
+ 883,
+ 823
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 880,
+ 101
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "page_number",
+ "text": "16",
+ "bbox": [
+ 522,
+ 830,
+ 535,
+ 840
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "table",
+ "img_path": "images/d86abcf5e5ec9f4d1a1e2cf178e6e2a2411500ea7111c2031609d3b0586d7569.jpg",
+ "table_caption": [
+ "Table 2. Datasets and metrics for personalized RAG and Agent."
+ ],
+ "table_footnote": [],
+ "table_body": "| Field | Metrics Category | Metrics | Datasets |
| Pre-retrieval | Textual Quality | BLEU, ROUGE, EM | Avocado Research Email Collection [57, 85], Amazon review [57, 83], Reddit comments[57, 118], Amazon ESCI dataset[82, 97], PIP |
| Information Retrieval | MAP, MRR, NDCG, Precision, Recall, RBP | AOL[88, 174], WARRIORS[99], Personalized Results Re-Ranking benchmark [6], delicio.us [9, 15, 144, 172], Flickr [9, 108], CiteULike [10, 14], LRDP [12], Delicious [141], Bibsonomy [79], Wikipedia [8, 33] |
| Classification | Accuracy, Macro-F1 | SCAN [56, 173], AITA WORKSM[53, 80], Robust04 [61] |
| Others | XEntropy, PMS, Image-Align, PQEC, ProfOverlap | Amazon ESCI dataset[82, 97], PIP, Bibsonomy [79] |
| Retrieval | Textual Quality | BLEU, ROUGE, Dis, PPL | TOPDIAL [130], Pchatbot [93], DuLemon [150] |
| Information Retrieval | Recall, MRR, Precision, F1 | LiveChat [34], Pchatbot [93], DuLemon [150] |
| Classification | Accuracy, Succ | TOPDIAL [130], PersonalityEvid [119], DuLemon [150], PersonalityEdit [75] |
| Others | Fluency, Coherence, Plausibility, ES, DD, TPEI, PAE | PersonalityEvid [119], PersonalityEdit [75] |
| Generation | Textual Quality | BLEU, ROUGE, Dis, PPL, METEOR | LaMP [105], Long LaMP [55], Dulemon [150], PGraphRAG [5], AmazonQA/Products [29], Reddit [170], MedicalDialogue [162] |
| Classification | Accuracy, F1, Persona F1 | LaMP [105], Long LaMP [55], Dulemon [150], AmazonQA/Products [29], Reddit [170], MedicalDialogue [162] |
| Regression | MAE, RMSE | LaMP [105], Long LaMP [55], PGraphRAG [5] |
| Others | Fluency, Mean Success Rate, Median Relative Improvements | Personalized-Gen [3] |
| Agent | Textual Quality | BLEU, ROUGE, METEOR, CIDer, EM, Fluency, Coherence, Instruction Adherence, Consistency related metrics | RICO [126], RoleBench [139], Shao et al. [110], Socialbench [18], MMRole-Data [27], ROLEPERSONALITY [96], ChatHarui [134], Character-LLM-Data [153], Knowledge Behind Persona [41], Wang et al. [137], Wang et al. [135], Zheng et al. [169] |
| Information Retrieval | Recall, F1, Precision | Knowledge Behind Persona [41] |
| Classification | Accuracy, Failure Rate, Classification Accuracy, Preference Rate, Correctness | MIT-BIH Arrhythmia Database [148], VirtualHome [44], Socialbench [18], ARC [100], AGIEval [100], HellaSwag [100], MedMCQA [100], AQUA-RAT [100], LogiQA [100], LSAT-AR [100], LSAT-LR [100], LSAT-RC [100], SAT-English [100], SAT-Math [100], PersonalWAB [17], TravelPlanner+ [114] |
| Others | Pass@k, Executability, Productivity, Plausibility of the Story | Hong et al. [40], Zheng et al. [169] |
",
+ "bbox": [
+ 119,
+ 138,
+ 823,
+ 488
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "5 EVALUATION AND DATASET",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 510,
+ 338,
+ 522
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "In the evolving landscape of personalization, from RAG to advanced Agent-based systems, the evaluation of models relies heavily on diverse datasets and metrics tailored to specific tasks. This survey categorizes metrics into several key types: Textual Quality metrics (e.g., BLEU, ROUGE, METEOR) assess the fluency and coherence of generated outputs; Information Retrieval metrics (e.g., MAP, MRR, Recall) evaluate the accuracy and relevance of retrieved information; Classification metrics (e.g., Accuracy, F1) measure task-specific correctness; Regression metrics (e.g., MAE, RMSE) quantify prediction errors; and Other metrics (e.g., Fluency, Pass@k) address domain-specific or task-unique aspects like plausibility or executability. These metrics span pre-retrieval, retrieval, generation, and agent-based personalization approaches, reflecting their varied objectives. To provide a comprehensive overview, we compile an extensive list of datasets across these fields, as detailed in Table 2. These datasets, paired with their respective metrics, enable researchers to benchmark and refine personalized systems, from enhancing query rewriting to enabling autonomous agents in physical and virtual environments.",
+ "bbox": [
+ 114,
+ 531,
+ 821,
+ 719
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "6 CHALLENGES AND FUTURE DIRECTIONS",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 734,
+ 429,
+ 748
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Personalized RAG and agent-based systems still face several critical challenges that warrant further exploration. We list them as follows:",
+ "bbox": [
+ 114,
+ 756,
+ 818,
+ 787
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "- Balancing Personalization and Scalability: Integrating personalization data (such as preferences, history, and contextual signals) into RAG processes often increases computational complexity, making it difficult to maintain",
+ "bbox": [
+ 114,
+ 795,
+ 818,
+ 825
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 116,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 818,
+ 101
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "page_number",
+ "text": "17",
+ "bbox": [
+ 462,
+ 830,
+ 473,
+ 840
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "efficiency and scalability across large-scale systems. Future work could explore lightweight, adaptive embeddings and hybrid frameworks that seamlessly fuse user profiles with real-time contexts.",
+ "bbox": [
+ 186,
+ 121,
+ 880,
+ 152
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- Evaluating Personalization Effectively: Current metrics like BLEU, ROUGE, and human evaluations fall short in capturing the nuanced alignment of outputs with dynamic user preferences, lacking tailored measures for personalization efficacy. Developing specialized benchmarks and metrics that assess long-term user satisfaction and adaptability is crucial for real-world applicability.",
+ "- Preserving Privacy through Device-Cloud Collaboration: Personalized retrieval often involves processing sensitive user data, raising privacy concerns, especially with the increased global emphasis on data protection regulations, such as the European Union's General Data Protection Regulation (GDPR). Consequently, a promising approach is the collaborative integration of on-device small Language models which handle sensitive personal data locally, with cloud-based LLM, which provides broader contextual knowledge.",
+ "- Personalized Agent Planning: Current research on agent planning remains mainly in its early stages, with much of the work focusing on building foundational frameworks such as GUI agents [81] and the application of agents across diverse domains [131]. Notably, the incorporation of personalized approaches has yet to be widely adopted. Exploring how to integrate personalized support into existing frameworks to enhance user experience represents a promising and valuable direction for future investigation.",
+ "- Ensuring Ethical and Coherent Systems: Bias in data processing, privacy concerns in user profiling, and coherence across retrieval and generation stages remain unresolved. Future directions should prioritize ethical safeguards, privacy-preserving techniques, and cross-stage optimization to build trustworthy, unified personalized systems."
+ ],
+ "bbox": [
+ 174,
+ 156,
+ 883,
+ 448
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "text",
+ "text": "7 CONCLUSION",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 468,
+ 302,
+ 481
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "text",
+ "text": "In this paper, we explore the landscape of personalization from Retrieval-Augmented Generation (RAG) to advanced LLM-based Agents, detailing adaptations across pre-retrieval, retrieval, and generation stages while extending into agentic capabilities. By reviewing recent literature, datasets, and metrics, we highlight the progress and diversity in enhancing user satisfaction through tailored AI systems. However, challenges such as scalability, effective evaluation, and ethical concerns underscore the need for innovative solutions. Future research should focus on lightweight frameworks, specialized benchmarks, and privacy-preserving techniques to advance personalized AI. Relevant papers and resources are also compiled online for ease of future research.",
+ "bbox": [
+ 174,
+ 489,
+ 883,
+ 608
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "text",
+ "text": "REFERENCES",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 630,
+ 272,
+ 642
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[1] 2021. BERT: a review of applications in natural language processing and understanding. arXiv preprint arXiv:2103.11943 (2021).",
+ "[2] Mahyar Abbasian, Iman Azimi, Amir M Rahmani, and Ramesh Jain. 2023. Conversational health agents: A personalized llm-powered agent framework. arXiv preprint arXiv:2310.02374 (2023).",
+ "[3] Bashar Alhafni, Vivek Kulkarni, Dhruv Kumar, and Vipul Raheja. 2024. Personalized Text Generation with Fine-Grained Linguistic Control. In Proceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024). 88–101.",
+ "[4] Amazon. [n.d.]. Amazon Customer Review Dataset. Online dataset. https://nijianmo.github.io/amazon/",
+ "[5] Steven Au, Cameron J Dimacali, Ojasmitha Pedirappagari, Namyong Park, Franck Dernoncourt, Yu Wang, Nikos Kanakaris, Hanieh Deilamsalehy, Ryan A Rossi, and Nesreen K Ahmed. 2025. Personalized Graph-Based Retrieval for Large Language Models. arXiv preprint arXiv:2501.02157 (2025).",
+ "[6] Elias Bassani, Pranav Kasela, Alessandro Raganato, and Gabriella Pasi. 2022. A multi-domain benchmark for personalized search evaluation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3822-3827.",
+ "[7] Elias Bassani, Nicola Tonellotto, and Gabriella Pasi. 2023. Personalized query expansion with contextual word embeddings. ACM Transactions on Information Systems 42, 2 (2023), 1-35.",
+ "[8] Oliver Baumann and Mirco Schoenfeld. 2024. PSQE: Personalized Semantic Query Expansion for user-centric query disambiguation. (2024)."
+ ],
+ "bbox": [
+ 186,
+ 648,
+ 882,
+ 824
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 879,
+ 101
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "page_number",
+ "text": "18",
+ "bbox": [
+ 522,
+ 830,
+ 535,
+ 840
+ ],
+ "page_idx": 17
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[9] Matthias Bender, Tom Crecelius, Mouna Kacimi, Sebastian Michel, Thomas Neumann, Josiane Xavier Parreira, Ralf Schenkel, and Gerhard Weikum. 2008. Exploiting social relations for query expansion and result ranking. In 2008 IEEE 24th International Conference on Data Engineering Workshop. IEEE, 501-506.",
+ "[10] Marin Bertier, Rachid Guerraoui, Vincent Leroy, and Anne-Marie Kermarrec. 2009. Toward personalized query expansion. In Proceedings of the Second ACM EuroSys Workshop on Social Network Systems. 7-12.",
+ "[11] Keping Bi, Qingyao Ai, and W Bruce Croft. 2021. Learning a fine-grained review-based transformer model for personalized product search. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 123-132.",
+ "[12] Claudio Biancalana and Alessandro Micarelli. 2009. Social tagging in query expansion: A new way for personalized web search. In 2009 International Conference on Computational Science and Engineering, Vol. 4. IEEE, 1060-1065.",
+ "[13] Microsoft Bing. [n.d]. Bing Search Engine. https://www.bing.com",
+ "[14] Mohamed Reda Bouadjenek, Hakim Hacid, and Mokrane Bouzeghoub. 2019. Personalized social query expansion using social annotations. Transactions on Large-Scale Data-and Knowledge-Centered Systems XL (2019), 1-25.",
+ "[15] Mohamed Reda Bouadjenek, Hakim Hacid, Mokrane Bouzeghoub, and Johann Daigremont. 2011. Personalized social query expansion using social bookmarking systems. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval. 1113-1114.",
+ "[16] Domenico Bulfamante. 2023. Generative enterprise search with extensible knowledge base using ai. Ph.D. Dissertation. Politecnico di Torino.",
+ "[17] Hongru Cai, Yongqi Li, Wenjie Wang, ZHU Fengbin, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. [n. d]. Large Language Models Empowered Personalized Web Agents. In THE WEB CONFERENCE 2025.",
+ "[18] Hongzhan Chen, Hehong Chen, Ming Yan, Wenshen Xu, Xing Gao, Weizhou Shen, Xiaojun Quan, Chenliang Li, Ji Zhang, Fei Huang, et al. 2024. Socialbench: Sociality evaluation of role-playing conversational agents. arXiv preprint arXiv:2403.13679 (2024).",
+ "[19] Jiangjie Chen, Xintao Wang, Rui Xu, Siyu Yuan, Yikai Zhang, Wei Shi, Jian Xie, Shuang Li, Ruihan Yang, Tinghui Zhu, et al. 2024. From persona to personalization: A survey on role-playing language agents. arXiv preprint arXiv:2404.18231 (2024).",
+ "[20] Ruizhe Chen, Xiaotian Zhang, Meng Luo, Wenhao Chai, and Zuozhu Liu. 2024. Pad: Personalized alignment of llms at decoding-time. arXiv preprint arXiv:2410.04070 (2024).",
+ "[21] Weijie Chen, Ting Bai, Jinbo Su, Jian Luan, Wei Liu, and Chuan Shi. 2024. Kg-retriever: Efficient knowledge indexing for retrieval-augmented large language models. arXiv preprint arXiv:2412.05547 (2024).",
+ "[22] Zheng Chen, Ziyan Jiang, Fan Yang, Eunah Cho, Xing Fan, Xiaojiang Huang, Yanbin Lu, and Aram Galstyan. 2023. Graph meets LLM: A novel approach to collaborative filtering for robust conversational understanding. arXiv preprint arXiv:2305.14449 (2023).",
+ "[23] Hao Cheng, Shuo Wang, Wensheng Lu, Wei Zhang, Mingyang Zhou, Kezhong Lu, and Hao Liao. 2023. Explainable recommendation with personalized review retrieval and aspect learning. arXiv preprint arXiv:2306.12657 (2023).",
+ "[24] Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. 2023. Adapting language models to compress contexts. arXiv preprint arXiv:2305.14788 (2023).",
+ "[25] Eunah Cho, Ziyan Jiang, Jie Hao, Zheng Chen, Saurabh Gupta, Xing Fan, and Chenlei Guo. 2021. Personalized search-based query rewrite system for conversational ai. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI. 179-188.",
+ "[26]Hyundong Cho, Karishma Sharma, Nicolaas Jedema, Leonardo FR Ribeiro, Alessandro Moschitti, Ravi Krishnan, and Jonathan May. 2025. TuningFree Personalized Alignment via Trial-Error-Explain In-Context Learning. arXiv preprint arXiv:2502.08972 (2025).",
+ "[27] Yanqi Dai, Huanran Hu, Lei Wang, Shengjie Jin, Xu Chen, and Zhiwu Lu. 2024. Mmrole: A comprehensive framework for developing and evaluating multimodal role-playing agents. arXiv preprint arXiv:2408.04203 (2024).",
+ "[28] Wenlong Deng, Christos Thrampoulidis, and Xiaoxiao Li. 2024. Unlocking the potential of prompt-tuning in bridging generalized and personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6087-6097.",
+ "[29] Yang Deng, Yaliang Li, Wenxuan Zhang, Bolin Ding, and Wai Lam. 2022. Toward personalized answer generation in e-commerce via multiperspective preference modeling. ACM Transactions on Information Systems (TOIS) 40, 4 (2022), 1-28.",
+ "[30] Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The Faiss library. (2024). arXiv:2401.08281 [cs.LG]",
+ "[31] ESPN. [n.d.]. ESPN Sports Statistics Dataset. Online dataset.",
+ "[32] 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.",
+ "[33] Byron J Gao, David C Anastasiu, and Xing Jiang. 2010. Utilizing user-input contextual terms for query disambiguation. In Coling 2010: Posters. 329-337.",
+ "[34] Jingsheng Gao, Yixin Lian, Ziyi Zhou, Yuzhuo Fu, and Baoyuan Wang. 2023. LiveChat: A large-scale personalized dialogue dataset automatically constructed from live streaming. arXiv preprint arXiv:2306.08401 (2023).",
+ "[35] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. 2023. Retrievalaugmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 2 (2023).",
+ "[36] Google. [n.d.]. Google Search. https://www.google.com"
+ ],
+ "bbox": [
+ 122,
+ 123,
+ 821,
+ 814
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 116,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "page_number",
+ "text": "19",
+ "bbox": [
+ 462,
+ 830,
+ 475,
+ 840
+ ],
+ "page_idx": 18
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[37] Jia-Chen Gu, Hui Liu, Zhen-Hua Ling, Quan Liu, Zhigang Chen, and Xiaodan Zhu. 2021. Partner matters! an empirical study on fusing personas for personalized response selection in retrieval-based chatbots. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 565-574.",
+ "[38] Jie Hao, Yang Liu, Xing Fan, Saurabh Gupta, Saleh Soltan, Rakesh Chada, Pradeep Natarajan, Chenlei Guo, and Gokhan Tur. 2022. CGF: Constrained generation framework for query rewriting in conversational AI. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track. 475-483.",
+ "[39] Nicola Henze, Peter Dolog, and Wolfgang Nejdl. 2004. Reasoning and ontologies for personalized e-learning in the semantic web. Journal of Educational Technology & Society 7, 4 (2004), 82-97.",
+ "[40] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. 2023. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352 3, 4 (2023), 6.",
+ "[41] WANG Hongru, Minda Hu, Yang Deng, Rui Wang, Fei Mi, Weichao Wang, Yasheng Wang, Wai-Chung Kwan, Irwin King, and Kam-Fai Wong. [n. d]. Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues. In The 2023 Conference on Empirical Methods in Natural Language Processing.",
+ "[42] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR 1, 2 (2022), 3.",
+ "[43] Qiushi Huang, Shuai Fu, Xubo Liu, Wenwu Wang, Tom Ko, Yu Zhang, and Lilian Tang. 2024. Learning retrieval augmentation for personalized dialogue generation. arXiv preprint arXiv:2406.18847 (2024).",
+ "[44] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. 2022. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning. PMLR, 9118-9147.",
+ "[45] Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of LLM agents: A survey. arXiv preprint arXiv:2402.02716 (2024).",
+ "[46] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. 2023. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653 (2023).",
+ "[47] Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, and Prithviraj Ammanabrolu. 2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv preprint arXiv:2310.11564 (2023).",
+ "[48] Pengyue Jia, Yiding Liu, Xiangyu Zhao, Xiaopeng Li, Changying Hao, Shuaiqiang Wang, and Dawei Yin. 2024. MILL: Mutual Verification with Large Language Models for Zero-Shot Query Expansion. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2498-2518.",
+ "[49] Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Zhang, and Yixin Zhu. 2023. Evaluating and inducing personality in pre-trained language models. Advances in Neural Information Processing Systems 36 (2023), 10622-10643.",
+ "[50] Hideaki Joko, Shubham Chatterjee, Andrew Ramsay, Arjen P De Vries, Jeff Dalton, and Faegheh Hasibi. 2024. Doing personal laps: Llm-augmented dialogue construction for personalized multi-session conversational search. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 796-806.",
+ "[51] Wang-Cheng Kang, Jianmo Ni, Nikhil Mehta, Maheswaran Sathiamoorthy, Lichan Hong, Ed Chi, and Derek Zhiyuan Cheng. 2023. Do llms understand user preferences? evaluating llms on user rating prediction. arXiv preprint arXiv:2305.06474 (2023).",
+ "[52] Manojkumar Rangasamy Kannadasan and Grigor Aslanyan. 2019. Personalized query auto-completion through a lightweight representation of the user context. arXiv preprint arXiv:1905.01386 (2019).",
+ "[53] Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Greg Corrado, Laszlo Lukacs, Marina Ganea, Peter Young, and Vivek Ramavajjala. 2016. Smart Reply: Automated Response Suggestion for Email. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (San Francisco, California, USA) (KDD '16). Association for Computing Machinery, New York, NY, USA, 955-964. https://doi.org/10.1145/2939672.2939801",
+ "[54] Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. 2024. Reinforcement learning for optimizing rag for domain chatbots. arXiv preprint arXiv:2401.06800 (2024).",
+ "[55] Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, et al. 2024. Longlamp: A benchmark for personalized long-form text generation. arXiv preprint arXiv:2407.11016 (2024).",
+ "[56] Brenden Lake and Marco Baroni. 2018. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning. PMLR, 2873-2882.",
+ "[57] Cheng Li, Mingyang Zhang, Qiao zhu Mei, Weize Kong, and Michael Bendersky. 2024. Learning to rewrite prompts for personalized text generation. In Proceedings of the ACM Web Conference 2024. 3367-3378.",
+ "[58] Changhao Li, Yuchen Zhuang, Rushi Qiang, Haotian Sun, Hanjun Dai, Chao Zhang, and Bo Dai. 2024. Matryoshka: Learning to Drive Black-Box LLMs with LLMs. arXiv preprint arXiv:2410.20749 (2024).",
+ "[59] Lei Li, Yongfeng Zhang, and Li Chen. 2023. Personalized prompt learning for explainable recommendation. ACM Transactions on Information Systems 41, 4 (2023), 1-26.",
+ "[60] Sen Li, Fuyu Lv, Taiwei Jin, Guiyang Li, Yukun Zheng, Tao Zhuang, Qingwen Liu, Xiaoyi Zeng, James Kwok, and Qianli Ma. 2022. Query rewriting in taobao search. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3262-3271."
+ ],
+ "bbox": [
+ 181,
+ 123,
+ 880,
+ 827
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 880,
+ 101
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "page_number",
+ "text": "20",
+ "bbox": [
+ 522,
+ 830,
+ 534,
+ 840
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[61] Xiaopeng Li, Lixin Su, Pengyue Jia, Xiangyu Zhao, Suqi Cheng, Junfeng Wang, and Dawei Yin. 2023. Agent4ranking: Semantic robust ranking via personalized query rewriting using multi-agent llm. arXiv preprint arXiv:2312.15450 (2023).",
+ "[62] Xinyu Li, Ruiyang Zhou, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133 (2024).",
+ "[63] Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, et al. 2024. Personal llm agents: Insights and survey about the capability, efficiency and security. arXiv preprint arXiv:2401.05459 (2024).",
+ "[64] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281 (2023).",
+ "[65] Ruixue Lian, Sixing Lu, Clint Solomon, Gustavo Aguilar, Pragaash Ponnusamy, Jialong Han, Chengyuan Ma, and Chenlei Guo. 2023. PersonalTM: Transformer memory for personalized retrieval. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2256-2260.",
+ "[66] Shan-Mu Lin and Chuen-Min Huang. 2006. Personalized optimal search in local query expansion. In Proceedings of the 18th Conference on Computational Linguistics and Speech Processing. 221-236.",
+ "[67] Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is chatgpt a good recommender? a preliminary study. arXiv preprint arXiv:2304.10149 (2023).",
+ "[68] Jiahong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Jieming Zhu, Minda Hu, Menglin Yang, and Irwin King. 2025. A Survey of Personalized Large Language Models: Progress and Future Directions. arXiv preprint arXiv:2502.11528 (2025).",
+ "[69] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics 12 (2024), 157-173.",
+ "[70] Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024. Once: Boosting content-based recommendation with both open-and closed-source large language models. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining. 452-461.",
+ "[71] Shuai Liu, Hyundong J Cho, Marjorie Freedman, Xuezhe Ma, and Jonathan May. 2023. RECAP: retrieval-enhanced context-aware prefix encoder for personalized dialogue response generation. arXiv preprint arXiv:2306.07206 (2023).",
+ "[72] Tyler Lu and Craig Boutilier. 2011. Budgeted social choice: From consensus to personalized decision making. In *IJCAI*, Vol. 11, 280-286.",
+ "[73] Zhengyi Ma, Zhicheng Dou, Yutao Zhu, Hanxun Zhong, and Ji-Rong Wen. 2021. One chatbot per person: Creating personalized chatbots based on implicit user profiles. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 555-564.",
+ "[74] Aman Madaan, Niket Tandon, Peter Clark, and Yiming Yang. 2022. Memory-assisted prompt editing to improve GPT-3 after deployment. arXiv preprint arXiv:2201.06009 (2022).",
+ "[75] Shengyu Mao, Xiaohan Wang, Mengru Wang, Yong Jiang, Pengjun Xie, Fei Huang, and Ningyu Zhang. 2024. Editing Personality for Large Language Models. In CCF International Conference on Natural Language Processing and Chinese Computing. Springer, 241-254.",
+ "[76] Yuren Mao, Xuemei Dong, Wenyi Xu, Yunjun Gao, Bin Wei, and Ying Zhang. 2024. Fit-rag: black-box rag with factual information and token reduction. arXiv preprint arXiv:2403.14374 (2024).",
+ "[77] Puneet Mathur, Zhe Liu, Ke Li, Yingyi Ma, Gil Keren, Zeeshan Ahmed, Dinesh Manocha, and Xuedong Zhang. 2023. Personal: Language model personalization via domain-distributed span aggregated k-nearest n-gram retrieval augmentation. In Findings of the Association for Computational Linguistics: EMNLP 2023. 11314-11328.",
+ "[78] Fatemehsadat Mireshghallah, Vaishnavi Shrivastava, Milad Shokouhi, Taylor Berg-Kirkpatrick, Robert Sim, and Dimitrios Dimitriadis. 2021. Identifier: Implicit user representations for simple and effective personalized sentiment analysis. arXiv preprint arXiv:2110.00135 (2021).",
+ "[79] Philippe Mulhem, Nawal Ould Amer, and Mathias Gery. 2016. Axiomatic term-based personalized query expansion using bookmarking system. In International Conference on Database and Expert Systems Applications. Springer, 235-243.",
+ "[80] Sheshera Mysore, Zhuoran Lu, Mengting Wan, Longqi Yang, Steve Menezes, Tina Baghaee, Emmanuel Barajas Gonzalez, Jennifer Neville, and Tara Safavi. 2023. Pearl: Personalizing large language model writing assistants with generation-calibrated retrievers. arXiv preprint arXiv:2311.09180 (2023).",
+ "[81] Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, et al. 2024. Gui agents: A survey. arXiv preprint arXiv:2412.13501 (2024).",
+ "[82] Duy A Nguyen, Rishi Kesav Mohan, Van Yang, Pritom Saha Akash, and Kevin Chen-Chuan Chang. 2025. RL-based Query Rewriting with Distilled LLM for online E-Commerce Systems. arXiv preprint arXiv:2501.18056 (2025).",
+ "[83] Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 188-197.",
+ "[84] Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O'Banion, and Jun Xie. 2024. User-llm: Efficient llm contextualization with user embeddings. arXiv preprint arXiv:2402.13598 (2024).",
+ "[85] Douglas Oard, William Webber, David Kirsch, and Sergey Golitsynski. 2015. Avocado research email collection. Philadelphia: Linguistic Data Consortium (2015).",
+ "[86] U.S. National Library of Medicine. [n.d.]. PubMed: A Free Resource for Biomedical Literature. https://pubmed.ncbi.nlm.nih.gov/",
+ "[87] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology. 1-22."
+ ],
+ "bbox": [
+ 122,
+ 123,
+ 821,
+ 827
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 116,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "page_number",
+ "text": "21",
+ "bbox": [
+ 462,
+ 830,
+ 473,
+ 840
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[88] Greg Pass, Abdur Chowdhury, and Cayley Torgeson. 2006. A picture of search. In Proceedings of the 1st International Conference on Scalable Information Systems (Hong Kong) (InfoScale '06). Association for Computing Machinery, New York, NY, USA, 1-es. https://doi.org/10.1145/1146847.1146848",
+ "[89] Vadim Igorevich Pavliukevich, Alina Khasanovna Zherdeva, Olesya Vladimirovna Makhnytkina, and Dmitriy Viktorovich Dyrmovskiy. [n. d.]. Improving RAG with LoRA finetuning for persona text generation. ([n. d.]).",
+ "[90] Dan Peng, Zhihui Fu, and Jun Wang. 2024. Pocketllm: Enabling on-device fine-tuning for personalized llms. arXiv preprint arXiv:2407.01031 (2024).",
+ "[91] Qiyao Peng, Hongtao Liu, Hongyan Xu, Qing Yang, Minglai Shao, and Wenjun Wang. 2024. Review-LLM: Harnessing Large Language Models for Personalized Review Generation. arXiv:2407.07487 [cs.CL] https://arxiv.org/abs/2407.07487",
+ "[92] Hongjin Qian, Zhicheng Dou, Yutao Zhu, Yueyuan Ma, and Ji-Rong Wen. 2021. Learning implicit user profile for personalized retrieval-based chatbot. In proceedings of the 30th ACM international conference on Information & Knowledge Management. 1467-1477.",
+ "[93] Hongjin Qian, Xiahe Li, Hanxun Zhong, Yu Guo, Yueyuan Ma, Yutao Zhu, Zhanliang Liu, Zhicheng Dou, and Ji-Rong Wen. 2021. Pchatbot: a large-scale dataset for personalized chatbot. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 2470-2477.",
+ "[94] Xiaoru Qu, Yifan Wang, Zhao Li, and Jun Gao. 2024. Graph-enhanced prompt learning for personalized review generation. Data Science and Engineering 9, 3 (2024), 309-324.",
+ "[95] A. Rajaraman and J.D. Ullman. 2011. Mining of Massive Datasets. Cambridge University Press. https://books.google.co.uk/books?id=OefRhZyYOb0C",
+ "[96] Yiting Ran, Xintao Wang, Rui Xu, Xinfeng Yuan, Jiaqing Liang, Deqing Yang, and Yanghua Xiao. 2024. Capturing minds, not just words: Enhancing role-playing language models with personality-indicative data. arXiv preprint arXiv:2406.18921 (2024).",
+ "[97] Chandan K. Reddy, Lluis Marquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopadhyay, Arnab Biswas, Anlu Xing, and Karthik Subbian. 2022. Shopping Queries Dataset: A Large-Scale ESCI Benchmark for Improving Product Search. (2022). arXiv:2206.06588",
+ "[98] Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084 (2019).",
+ "[99] Ruiyang Ren, Peng Qiu, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Hua Wu, Ji-Rong Wen, and Haifeng Wang. 2024. Bases: Large-scale web search user simulation with large language model based agents. arXiv preprint arXiv:2402.17505 (2024).",
+ "[100] Matthew Renze and Erhan Guven. 2024. Self-reflection in llm agents: Effects on problem-solving performance. arXiv preprint arXiv:2405.06682 (2024).",
+ "[101] 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).",
+ "[102] Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval 3, 4 (2009), 333-389.",
+ "[103] 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.",
+ "[104] Alireza Salemi, Cheng Li, Mingyang Zhang, Qiao zhu Mei, Weize Kong, Tao Chen, Zhuowan Li, Michael Bendersky, and Hamed Zamani. 2025. Reasoning-Enhanced Self-Training for Long-Form Personalized Text Generation. arXiv preprint arXiv:2501.04167 (2025).",
+ "[105] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 7370-7392.",
+ "[106] Alireza Salemi and Hamed Zamani. 2024. Learning to Rank for Multiple Retrieval-Augmented Models through Iterative Utility Maximization. arXiv preprint arXiv:2410.09942 (2024).",
+ "[107] Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect?. In International Conference on Machine Learning. PMLR, 29971-30004.",
+ "[108] Rossano Schifanella, Alain Barrat, Ciro Cattuto, Benjamin Markines, and Filippo Menczer. 2010. Folks in folksonomies: social link prediction from shared metadata. In Proceedings of the third ACM international conference on Web search and data mining. 271-280.",
+ "[109] Noor Shaker, Georgios Yannakakis, and Julian Togelius. 2010. Towards automatic personalized content generation for platform games. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, Vol. 6. 63-68.",
+ "[110] Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2023. Character-llm: A trainable agent for role-playing. arXiv preprint arXiv:2310.10158 (2023).",
+ "[111] Jocelyn Shen, Joel Mire, Hae Won Park, Cynthia Breazeal, and Maarten Sap. 2024. HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs. arXiv preprint arXiv:2405.17633 (2024).",
+ "[112] Yunxiao Shi, Xing Zi, Zijing Shi, Haimin Zhang, Qiang Wu, and Min Xu. 2024. Eragent: Enhancing retrieval-augmented language models with improved accuracy, efficiency, and personalization. arXiv preprint arXiv:2405.06683 (2024).",
+ "[113] Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. 2025. Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG. arXiv preprint arXiv:2501.09136 (2025).",
+ "[114] Harmanpreet Singh, Nikhil Verma, Yixiao Wang, Manasa Bharadwaj, Homa Fashandi, Kevin Ferreira, and Chul Lee. 2024. Personal Large Language Model Agents: A Case Study on Tailored Travel Planning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 486-514.",
+ "[115] Shamane Siriwardhana, Rivindu Weerasekera, Elliott Wen, Tharindu Kaluarachchi, Rajib Rana, and Suranga Nanayakkara. 2023. Improving the domain adaptation of retrieval augmented generation (RAG) models for open domain question answering. Transactions of the Association for Computational Linguistics 11 (2023), 1-17.",
+ "[116] Mingyang Song and Mao Zheng. 2024. A Survey of Query Optimization in Large Language Models. arXiv preprint arXiv:2412.17558 (2024)."
+ ],
+ "bbox": [
+ 179,
+ 123,
+ 880,
+ 827
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 880,
+ 101
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "page_number",
+ "text": "22",
+ "bbox": [
+ 522,
+ 830,
+ 535,
+ 839
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[117] Spotify. 2023. Annoy: Approximate Nearest Neighbors in C++/Python. https://github.com/spotify/annoy",
+ "[118] Stuck_In_the Matrix. 2015. Reddit Public Comments (2007-10 through 2015-05). (2015). https://www.reddit.com/r/datasets/comments/3bxlg7/i_have EVERY_publicly-available Reddit_COMMENT/",
+ "[119] Lei Sun, Jinming Zhao, and Qin Jin. 2024. Revealing Personality Traits: A New Benchmark Dataset for Explanable Personality Recognition on Dialogues. arXiv preprint arXiv:2409.19723 (2024).",
+ "[120] Zhaoxuan Tan, Zheyuan Liu, and Meng Jiang. 2024. Personalized pieces: Efficient personalized large language models through collaborative efforts. arXiv preprint arXiv:2406.10471 (2024).",
+ "[121] 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 preprint arXiv:2402.04401 (2024).",
+ "[122] Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2025. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. arXiv:2402.04401 [cs.CL] https://arxiv.org/abs/2402.04401",
+ "[123] Quan Tu, Shilong Fan, Zihang Tian, and Rui Yan. 2024. Charactereval: A Chinese benchmark for role-playing conversational agent evaluation. arXiv preprint arXiv:2401.01275 (2024).",
+ "[124] Cornell University. [n.d.]. arXiv: An Open Access Repository for Research. https://arxiv.org/",
+ "[125] Hemanth Vemuri, Sheshansh Agrawal, Shivam Mittal, Deepak Saini, Akshay Soni, Abhinav V Sambasivan, Wenhao Lu, Yajun Wang, Mehul Parsana, Purushottam Kar, et al. 2023. Personalized retrieval over millions of items. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1014-1022.",
+ "[126] Bryan Wang, Gang Li, and Yang Li. 2023. Enabling conversational interaction with mobile ui using large language models. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1-17.",
+ "[127] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023).",
+ "[128] Hongru Wang, Wenyu Huang, Yang Deng, Rui Wang, Zezhong Wang, Yufei Wang, Fei Mi, Jeff Z Pan, and Kam-Fai Wong. 2024. Unims-rag: A unified multi-source retrieval-augmented generation for personalized dialogue systems. arXiv preprint arXiv:2401.13256 (2024).",
+ "[129] Hongru Wang, Rui Wang, Fei Mi, Yang Deng, Zezhong Wang, Bin Liang, Ruifeng Xu, and Kam-Fai Wong. 2023. Cue-CoT: Chain-of-thought prompting for responding to in-depth dialogue questions with LLMs. arXiv preprint arXiv:2305.11792 (2023).",
+ "[130] Jian Wang, Yi Cheng, Dongding Lin, Chak Tou Leong, and Wenjie Li. 2023. Target-oriented proactive dialogue systems with personalization: Problem formulation and dataset curation. arXiv preprint arXiv:2310.07397 (2023).",
+ "[131] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2024. A survey on large language model based autonomous agents. Frontiers of Computer Science 18, 6 (2024), 186345.",
+ "[132] Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query expansion with large language models. arXiv preprint arXiv:2303.07678 (2023).",
+ "[133] Lei Wang, Jingsen Zhang, Hao Yang, Zhiyuan Chen, Jiakai Tang, Zeyu Zhang, Xu Chen, Yankai Lin, Ruihua Song, Wayne Xin Zhao, et al. 2023. User behavior simulation with large language model based agents. arXiv preprint arXiv:2306.02552 (2023).",
+ "[134] Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, et al. 2023. Incharacter: Evaluating personality fidelity in role-playing agents through psychological interviews. arXiv preprint arXiv:2310.17976 (2023).",
+ "[135] Yixiao Wang, Homa Fashandi, and Kevin Ferreira. 2024. Investigating the Personality Consistency in Quantized Role-Playing Dialogue Agents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 239–255.",
+ "[136] Yu Wang, Yifan Gao, Xiusi Chen, Haoming Jiang, Shiyang Li, Jingfeng Yang, Qingyu Yin, Zheng Li, Xian Li, Bing Yin, et al. [n.d.]. MEMORYLLM: Towards Self-Updatable Large Language Models. In Forty-first International Conference on Machine Learning.",
+ "[137] Zheng Wang, Zhongyang Li, Zeren Jiang, Dandan Tu, and Wei Shi. 2024. Crafting Personalized Agents through Retrieval-Augmented Generation on Editable Memory Graphs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 4891-4906.",
+ "[138] Zijie J Wang and Duen Horng Chau. 2024. MeMemo: On-device Retrieval Augmentation for Private and Personalized Text Generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2765-2770.",
+ "[139] Zekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, et al. 2023. Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. arXiv preprint arXiv:2310.00746 (2023).",
+ "[140] Tianxin Wei, Bowen Jin, Ruirui Li, Hansi Zeng, Zhengyang Wang, Jianhui Sun, Qingyu Yin, Hanqing Lu, Suhang Wang, Jingrui He, et al. 2024. Towards unified multi-modal personalization: Large vision-language models for generative recommendation and beyond. arXiv preprint arXiv:2403.10667 (2024).",
+ "[141] Robert Wetzker, Carsten Zimmermann, and Christian Bauchage. 2008. Analyzing social bookmarking systems: A del. icio. us cookbook. In Proceedings of the ECAI 2008 Mining Social Data Workshop. 26-30.",
+ "[142] Stanisław Wozniak, Bartlomiej Koptyra, Arkadiusz Janz, Przemysław Kazienko, and Jan Kocón. 2024. Personalized large language models. arXiv preprint arXiv:2402.09269 (2024).",
+ "[143] Junde Wu, Jiayuan Zhu, Yunli Qi, Jingkun Chen, Min Xu, Filippo Menolascina, and Vicente Grau. 2024. Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187 (2024).",
+ "[144] Xuan Wu, Dong Zhou, Yu Xu, and Seamus Lawless. 2017. Personalized query expansion utilizing multi-relational social data. In 2017 12th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP). IEEE, 65-70."
+ ],
+ "bbox": [
+ 117,
+ 123,
+ 820,
+ 825
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 114,
+ 90,
+ 348,
+ 102
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "page_number",
+ "text": "23",
+ "bbox": [
+ 462,
+ 830,
+ 473,
+ 840
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[145] Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender Systems. 12-22.",
+ "[146] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences 68, 2 (2025), 121101.",
+ "[147] Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. 641-649.",
+ "[148] Huatao Xu, Liying Han, Qirui Yang, Mo Li, and Mani Srivastava. 2024. Penetrative ai: Making llms comprehend the physical world. In Proceedings of the 25th International Workshop on Mobile Computing Systems and Applications. 1-7.",
+ "[149] Hongyan Xu, Hongtao Liu, Pengfei Jiao, and Wenjun Wang. 2021. Transformer reasoning network for personalized review summarization. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1452-1461.",
+ "[150] Xinchao Xu, Zhibin Gou, Wenquan Wu, Zheng-Yu Niu, Hua Wu, Haifeng Wang, and Shihang Wang. 2022. Long time no see! open-domain conversation with long-term persona memory. arXiv preprint arXiv:2203.05797 (2022).",
+ "[151] Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, and Tat-Seng Chua. 2025. Personalized Generation In Large Model Era: A Survey. arXiv preprint arXiv:2503.02614 (2025).",
+ "[152] Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. 2024. Personalized federated continual learning via multi-granularity prompt. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4023-4034.",
+ "[153] Xiaoyan Yu, Tongxu Luo, Yifan Wei, Fangyu Lei, Yiming Huang, Hao Peng, and Liehuang Zhu. 2024. Neeko: Leveraging dynamic lora for efficient multi-character role-playing agent. arXiv preprint arXiv:2402.13717 (2024).",
+ "[154] Xinfeng Yuan, Siyu Yuan, Yuhan Cui, Tianhe Lin, Xintao Wang, Rui Xu, Jiangjie Chen, and Deqing Yang. 2024. Evaluating character understanding of large language models via character profiling from fictional works. arXiv preprint arXiv:2404.12726 (2024).",
+ "[155] Hansi Zeng, Surya Kallumadi, Zaid Alibadi, Rodrigo Nogueira, and Hamed Zamani. 2023. A personalized dense retrieval framework for unified information access. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 121-130.",
+ "[156] Saber Zerhoudi and Michael Granitzer. 2024. PersonaRAG: Enhancing Retrieval-Augmented Generation Systems with User-Centric Agents. arXiv preprint arXiv:2407.09394 (2024).",
+ "[157] Han Zhang, Songlin Wang, Kang Zhang, Zhiling Tang, Yunjiang Jiang, Yun Xiao, Weipeng Yan, and Wen-Yun Yang. 2020. Towards personalized and semantic retrieval: An end-to-end solution for e-commerce search via embedding learning. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2407-2416.",
+ "[158] Jiarui Zhang. 2024. Guided profile generation improves personalization with llms. arXiv preprint arXiv:2409.13093 (2024).",
+ "[159] Jesse Zhang, Jiahui Zhang, Karl Pertsch, Ziyi Liu, Xiang Ren, Minsuk Chang, Shao-Hua Sun, and Joseph J Lim. [n.d.]. Bootstrap Your Own Skills: Learning to Solve New Tasks with Large Language Model Guidance. In 7th Annual Conference on Robot Learning.",
+ "[160] Kai Zhang, Yangyang Kang, Fubang Zhao, and Xiaozhong Liu. 2023. LLM-based medical assistant personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023).",
+ "[161] Kaiyan Zhang, Jianyu Wang, Ermo Hua, Biqing Qi, Ning Ding, and Bowen Zhou. 2024. Cogenesis: A framework collaborating large and small language models for secure context-aware instruction following. arXiv preprint arXiv:2403.03129 (2024).",
+ "[162] Kai Zhang, Fubang Zhao, Yangyang Kang, and Xiaozhong Liu. 2023. Memory-augmented llm personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023).",
+ "[163] Wenlin Zhang, Chuhan Wu, Xiangyang Li, Yuhao Wang, Kuicai Dong, Yichao Wang, Xinyi Dai, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2025. LLMTreeRec: Unleashing the Power of Large Language Models for Cold-Start Recommendations. In Proceedings of the 31st International Conference on Computational Linguistics. 886-896.",
+ "[164] Yanyue Zhang, Yulan He, and Deyu Zhou. 2025. Rehearse With User: Personalized Opinion Summarization via Role-Playing based on Large Language Models. arXiv preprint arXiv:2503.00449 (2025).",
+ "[165] You Zhang, Jin Wang, Liang-Chih Yu, Dan Xu, and Xuejie Zhang. 2024. Personalized LoRA for human-centered text understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 1958-19596.",
+ "[166] Yabin Zhang, Wenhui Yu, Erhan Zhang, Xu Chen, Lantao Hu, Peng Jiang, and Kun Gai. 2024. Recgpt: Generative personalized prompts for sequential recommendation via chatgpt training paradigm. arXiv preprint arXiv:2404.08675 (2024).",
+ "[167] Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501 (2024).",
+ "[168] Zhehao Zhang, Ryan A Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, et al. 2024. Personalization of large language models: A survey. arXiv preprint arXiv:2411.00027 (2024).",
+ "[169] Yi Zheng, Chongyang Ma, Kanle Shi, and Haibin Huang. 2023. Agents meet okr: An object and key results driven agent system with hierarchical self-collaboration and self-evaluation. arXiv preprint arXiv:2311.16542 (2023).",
+ "[170] Hanxun Zhong, Zhicheng Dou, Yutao Zhu, Hongjin Qian, and Ji-Rong Wen. 2022. Less is more: Learning to refine dialogue history for personalized dialogue generation. arXiv preprint arXiv:2204.08128 (2022).",
+ "[171] Wanjun Zhong, Duyu Tang, Jiahai Wang, Jian Yin, and Nan Duan. 2021. UserAdapter: Few-shot user learning in sentiment analysis. In Findings of the Association for Computational Linguistics: ACL-JJCNLP 2021. 1484-1488."
+ ],
+ "bbox": [
+ 176,
+ 123,
+ 880,
+ 825
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 176,
+ 90,
+ 459,
+ 102
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "header",
+ "text": "X. Li and P. Jia, et al.",
+ "bbox": [
+ 779,
+ 90,
+ 880,
+ 101
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "page_number",
+ "text": "24",
+ "bbox": [
+ 522,
+ 830,
+ 535,
+ 840
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[172] Dong Zhou, Séamus Lawless, and Vincent Wade. 2012. Improving search via personalized query expansion using social media. Information retrieval 15 (2012), 218-242.",
+ "[173] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. 2022. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022).",
+ "[174] Yujia Zhou, Qiannan Zhu, Jiajie Jin, and Zhicheng Dou. 2024. Cognitive personalized search integrating large language models with an efficient memory mechanism. In Proceedings of the ACM Web Conference 2024. 1464-1473.",
+ "[175] Yuchen Zhuang, Haotian Sun, Yue Yu, Rushi Qiang, Qifan Wang, Chao Zhang, and Bo Dai. [n.d.]. Hydra: Model factorization framework for black-box llm personalization, 2024. URL https://arxiv.org/abs/2406.02888 ([n.d.])."
+ ],
+ "bbox": [
+ 117,
+ 123,
+ 821,
+ 223
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "text",
+ "text": "Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009",
+ "bbox": [
+ 116,
+ 238,
+ 503,
+ 250
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "header",
+ "text": "A Survey of Personalization: From RAG to Agent",
+ "bbox": [
+ 116,
+ 90,
+ 346,
+ 102
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "header",
+ "text": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY",
+ "bbox": [
+ 539,
+ 90,
+ 821,
+ 101
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "page_number",
+ "text": "25",
+ "bbox": [
+ 462,
+ 830,
+ 473,
+ 840
+ ],
+ "page_idx": 24
+ }
+]
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10147/26499afc-f1b2-4507-8b62-4adc6ac17e5f_model.json b/data/2025/2504_10xxx/2504.10147/26499afc-f1b2-4507-8b62-4adc6ac17e5f_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..6bae7212863b8d4d15d5aeea4d579ddbf29f8b10
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/26499afc-f1b2-4507-8b62-4adc6ac17e5f_model.json
@@ -0,0 +1,4650 @@
+[
+ [
+ {
+ "type": "aside_text",
+ "bbox": [
+ 0.023,
+ 0.273,
+ 0.061,
+ 0.709
+ ],
+ "angle": 270,
+ "content": "arXiv:2504.10147v1 [cs.IR] 14 Apr 2025"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.117,
+ 0.551,
+ 0.136
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.149,
+ 0.495,
+ 0.167
+ ],
+ "angle": 0,
+ "content": "XIAOPENG LI*, City University of Hong Kong, Hong Kong"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.17,
+ 0.497,
+ 0.186
+ ],
+ "angle": 0,
+ "content": "PENGYUE JIA*, City University of Hong Kong, Hong Kong"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.19,
+ 0.822,
+ 0.207
+ ],
+ "angle": 0,
+ "content": "DERONG XU, City University of Hong Kong, Hong Kong and University of Science and Technology of China, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.211,
+ 0.445,
+ 0.227
+ ],
+ "angle": 0,
+ "content": "YI WEN, City University of Hong Kong, Hong Kong"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.232,
+ 0.773,
+ 0.248
+ ],
+ "angle": 0,
+ "content": "YINGYI ZHANG, City University of Hong Kong, Hong Kong and Dalian University of Technology, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.252,
+ 0.516,
+ 0.269
+ ],
+ "angle": 0,
+ "content": "WENLIN ZHANG, City University of Hong Kong, Hong Kong"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.273,
+ 0.502,
+ 0.289
+ ],
+ "angle": 0,
+ "content": "WANYU WANG, City University of Hong Kong, Hong Kong"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.293,
+ 0.438,
+ 0.308
+ ],
+ "angle": 0,
+ "content": "YICHAO WANG, Noah's Ark Lab, Huawei, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.314,
+ 0.456,
+ 0.329
+ ],
+ "angle": 0,
+ "content": "ZHAOCHENG DU, Noah's Ark Lab, Huawei, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.334,
+ 0.435,
+ 0.349
+ ],
+ "angle": 0,
+ "content": "XIANGYANG LI, Noah's Ark Lab, Huawei, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.355,
+ 0.422,
+ 0.37
+ ],
+ "angle": 0,
+ "content": "YONG LIU, Noah's Ark Lab, Huawei, Singapore"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.375,
+ 0.436,
+ 0.39
+ ],
+ "angle": 0,
+ "content": "HUIFENG GUO, Noah's Ark Lab, Huawei, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.394,
+ 0.451,
+ 0.41
+ ],
+ "angle": 0,
+ "content": "RUIMING TANG†, Noah's Ark Lab, Huawei, China"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.116,
+ 0.416,
+ 0.519,
+ 0.432
+ ],
+ "angle": 0,
+ "content": "XIANGYU ZHAO†, City University of Hong Kong, Hong Kong"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.443,
+ 0.823,
+ 0.598
+ ],
+ "angle": 0,
+ "content": "Personalization has become an essential capability in modern AI systems, enabling customized interactions that align with individual user preferences, contexts, and goals. Recent research has increasingly concentrated on Retrieval-Augmented Generation (RAG) frameworks and their evolution into more advanced agent-based architectures within personalized settings to enhance user satisfaction. Building on this foundation, this survey systematically examines personalization across the three core stages of RAG: pre-retrieval, retrieval, and generation. Beyond RAG, we further extend its capabilities into the realm of Personalized LLM-based Agents, which enhance traditional RAG systems with agentic functionalities, including user understanding, personalized planning and execution, and dynamic generation. For both personalization in RAG and agent-based personalization, we provide formal definitions, conduct a comprehensive review of recent literature, and summarize key datasets and evaluation metrics. Additionally, we discuss fundamental challenges, limitations, and promising research directions in this evolving field. Relevant papers and resources are continuously updated at the Github Repo| Field | Sub-field | Subsub-field | Papers |
| Pre-retrieval | Query Rewriting | Learning to Personalized Query Rewrite | CLE-QR [60], CGF [38], PEARL [80] |
| LLM to Personalized Query Rewrite | Least-to-Most Prompting [173], ERAGent [112], CoPS [174], Agent4Ranking [61], FIG [22], BASES [99] |
| Query Expansion | Tagging-based query expansion | Gossiple [10], Biancalana and Micarelli [12], SoQuES [15], Zhou et al. [172] |
| Else | Lin and Huang [66], Bender et al. [9], Axiomatic PQEC [79], WE-LM [144], PSQE [14], PQEWC [7] |
| Others | Bobo [33], Kannadasan and Aslanyan [52], PSQE [8] |
| Retrieval | Indexing | PEARL [80], KG-Retriever [21], EMG-RAG [137], PGraphRAG [5] |
| Retrieval | Dense Retrieval | MeMemo [138], RECAP [71], LAPDOG [43], Gu et al. [37], PersonalLM [77], UIA [155], XPERT [125], DPSR [157], RTM [11], Pearl [80], MemPrompt [74], EERRA [23], MALP [160], USER-LLM [84], PER-PCS [120] |
| Sparse Retrieval | OPPU [121], PAG [101], Au et al. [5], UniMS-RAG [128], Deng et al. [29], |
| Prompt-based Retrieval | LAPS [50], UniIMP [140], Shen et al. [111] |
| Others | Salem et al. [103], PersonalITM [65], Zhang et al. [165] |
| Post-retrieval | PersonaRAG [156], Pavliukevich et al. [89], UniMS-RAG [128], Salemi and Zamani [106], Zhang et al. [164], AutoCompressors [24], FIT-RAG [76] |
| Generation | Generation from Explicit Preferences | Direct Prompting | P² [49], Character Profiling [154] OpinionQA [107], Kang et al. [51], Liu et al. [67], Cue-CoT [129], TICL [26] |
| Profile-Augmented Prompting | GPG [158], Richardson et al. [101], ONCE [70], LLMTreeRec [163], KAR [145], Matryoshka [58] |
| Personalized-Prompt Prompting | Li et al. [57], RecGPT [166], PEPLER-D [59], GRAPA [94], SGPT [28], PFCL [152] |
| Generation from Implicit Preferences | Fine-tuning-Based Methods | PLoRA [165], LM-P [142], MiLP [165], OPPU [122], PER-PCS [120], Review-LLM [91], UserIdentifier [78], UserAdapter [171], HYDRA [175], PocketLLM [90], CoGenesis [161] |
| Reinforcement Learning-Based Methods | P-RLHF [62], P-SOUPS [47], PAD [20], REST-PG [104], Salemi et al. [103], RewrimerSIRI [57],Kulkarni et al. [54] |
| From RAG to Agent | Personalized Understanding | In user-profile understanding | Xu et al. [148], Abbasian et al. [2], |
| In agent's role understanding | RoleLLM [139], Character-LLM [110], Wang et al. [134], |
| In agent's user-role joint understanding | SocialBench [18], Dai et al. [27], Ran et al. [96], Wang et al. [126], Tu et al. [123], Necko [153] |
| Personalized Planning and Execution | Memory Management | EMG-RAG [137], Park et al. [87], Abbasian et al. [2], RecAgent [133], TravelPlanner+ [114], PersonalWAB [17], VOYAGER [127], MemoeryLLM [136] |
| Tool and API Calling | VOYAGER [127], Zhang et al. [159], PUMA [17], Wang et al. [126], PenetrativeAI [148], Huang et al. [44], [87], MetaGPT [40], OKR-Agent [169] |
| Personalized Generation | Alignment with User Fact | Character-LLM [110], Wang et al. [135], Dai et al. [27] |
| Alignment with User Preferences | Wang et al. [139], Ran et al. [96], Wang et al. [134], Chen et al. [18] |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.531,
+ 0.88,
+ 0.56
+ ],
+ "angle": 0,
+ "content": "- User Historical Content: Implicit personalization derived from user-generated content, such as chat history, emails, reviews, and social media interactions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.565,
+ 0.872,
+ 0.579
+ ],
+ "angle": 0,
+ "content": "- Persona-Based User Simulation: The use of LLM-based agents to simulate and generate personalized interactions."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.176,
+ 0.531,
+ 0.88,
+ 0.579
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.588,
+ 0.88,
+ 0.62
+ ],
+ "angle": 0,
+ "content": "Integrating this personalized information at various stages of the RAG and agent workflows enables dynamic alignment with human preferences, thereby making responses more user-centric and adaptive."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.176,
+ 0.638,
+ 0.461,
+ 0.65
+ ],
+ "angle": 0,
+ "content": "3 HOW TO ADOPT PERSONALIZATION"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.66,
+ 0.712,
+ 0.674
+ ],
+ "angle": 0,
+ "content": "We define the process of introducing personalization within the RAG pipeline as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.413,
+ 0.685,
+ 0.881,
+ 0.699
+ ],
+ "angle": 0,
+ "content": "\\[\ng = \\mathcal {G} (\\mathcal {R} (Q (q, p), C, p), \\text {p r o m p t}, p, \\theta) \\tag {1}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.709,
+ 0.882,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "where \\( p \\) denotes personalized information, and the process unfolds in three steps. In the pre-retrieval phase, query processing \\( (Q) \\) refines the query \\( q \\) using personalized information, such as through query rewriting or expansion. During the retrieval phase, the retriever \\( (\\mathcal{R}) \\) leverages \\( p \\) to fetch relevant documents from the corpus \\( (C) \\). Finally, in the generation phase, the retrieved information, combined with \\( p \\) and structured using the given prompt, id fed into the generator \\( (\\mathcal{G}) \\) with parameter \\( \\theta \\) to produce the final response \\( g \\). It is evident that personalized information directly influences multiple stages of the RAG pipeline. In this survey, we consider the agent system as a specialized application of the RAG framework, where personalization is incorporated in a manner similar to the RAG framework."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.525,
+ 0.832,
+ 0.534,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "4"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.116,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.116,
+ 0.125,
+ 0.818,
+ 0.384
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.316,
+ 0.392,
+ 0.62,
+ 0.405
+ ],
+ "angle": 0,
+ "content": "Fig. 2. Overview of the personalized pre-retrieval stage."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.426,
+ 0.416,
+ 0.439
+ ],
+ "angle": 0,
+ "content": "4 WHERE TO ADOPT PERSONALIZATION"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.115,
+ 0.448,
+ 0.24,
+ 0.46
+ ],
+ "angle": 0,
+ "content": "4.1 Pre-retrieval"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.47,
+ 0.823,
+ 0.536
+ ],
+ "angle": 0,
+ "content": "4.1.1 Definition. Pre-retrieval is a crucial step in information retrieval systems, where the original user query is enhanced or modified before the retrieval process to improve the relevance and quality of the search results, as shown in Figure 2. This process often incorporates additional contextual or personalized information to better align the query with the user's intent. The process can be formalized as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.427,
+ 0.546,
+ 0.822,
+ 0.561
+ ],
+ "angle": 0,
+ "content": "\\[\nq ^ {*} = Q (q, p) \\tag {2}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.571,
+ 0.821,
+ 0.601
+ ],
+ "angle": 0,
+ "content": "where \\( p \\) and \\( q \\) denote the personalized information and original query, and \\( q^{*} \\) is the optimized query after query reformulation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.615,
+ 0.822,
+ 0.698
+ ],
+ "angle": 0,
+ "content": "4.1.2 Query Rewriting. Query rewriting in RAG at the pre-retrieval stage refers to the process of reformulating user queries to enhance retrieval effectiveness by improving relevance, disambiguating intent, or incorporating contextual information before retrieving documents from an external knowledge source. The literature on personalized query rewriting can be broadly classified into two primary categories: (1) Direct Personalized Query Rewriting and (2) Auxiliary Personalized Query Rewriting."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.71,
+ 0.823,
+ 0.828
+ ],
+ "angle": 0,
+ "content": "(1). Direct Personalized Query Rewriting. The first category focuses on personalized query rewriting by using direct models. For example, Cho et al. [25] presents a personalized search-based query rewrite system for conversational AI that addresses user-specific semantic and phonetic errors. Nguyen et al. [82] apply reinforcement learning techniques to improve query rewriting in online e-commerce systems, leveraging distilled LLMs for personalized performance. CLE-QR [60] explores query rewriting in Taobao's search engine to enhance user satisfaction through customized query adaptation. CGF [38] introduces a constrained generation framework that allows for more flexible and personalized query rewriting in conversational AI. Li et al. [57] investigate learning methods to rewrite prompts for personalized"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.465,
+ 0.832,
+ 0.472,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "5"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.88,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.122,
+ 0.882,
+ 0.171
+ ],
+ "angle": 0,
+ "content": "text generation, improving the relevance and engagement of AI-generated content. Additionally, PEARL [80] discusses personalizing large language model-based writing assistants through the integration of generation-calibrated retrievers, enhancing AI-generated content."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.186,
+ 0.885,
+ 0.374
+ ],
+ "angle": 0,
+ "content": "(2). Auxiliary Personalized Query Rewriting. The second category emphasizes personalized query rewriting by using auxiliary mechanisms, such as retrieval, reasoning strategies, and external memory. Zhou et al. [173] propose a least-to-most prompting strategy that aids in complex reasoning within LLMs, which can be adapted for personalized text generation. ERAGent [112] enhance retrieval-augmented LLMs to improve personalization, efficiency, and accuracy, indirectly supporting personalized query rewriting for content generation. CoPS [174] integrate LLMs with memory mechanisms to create more personalized search experiences, which also influences content generation through better query understanding. Further, Agent4Ranking [61] employs multi-agent LLMs to perform semantic robust ranking, including personalized query rewriting to improve search rankings. FIG [22] combine graph-based methods with LLMs to query rewrite, improving personalized content generation and conversational interactions. Lastly, BASES [99] employ LLM-based agents to simulate large-scale web search user interactions, contributing to the development of personalized query rewriting strategies for content generation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.389,
+ 0.884,
+ 0.492
+ ],
+ "angle": 0,
+ "content": "4.1.3 Query Expansion. Query expansion enhances retrieval systems by expanding a user's original query with additional terms, synonyms, or refined structure to better capture intent. This improves the relevance and scope of retrieved documents. Recent advancements in LLMs have reinvigorated this field [46, 48, 132], leveraging their comprehension and generation abilities to expand queries using encoded knowledge or external retrieval, with notable success. Personalized query expansion, a subset, incorporates user-specific data to tailor results, boosting performance and customizing the search experience."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.506,
+ 0.884,
+ 0.658
+ ],
+ "angle": 0,
+ "content": "(1). Tagging-based Query Expansion. By 2009, studies began incorporating tagging information to enhance personalized query expansion. For instance, Gossle [10] introduced the TagMap and TagRank algorithms, which dynamically selected tags from personalized networks constructed using the cosine similarity of user-item tag distances, improving recall performance. Similarly, Biancalana and Micarelli [12] recorded user queries and visited URLs, leveraging social bookmarking to extract relevant tags and build a personalized three-dimensional co-occurrence matrix. Based on this, multiple semantically categorized expanded queries were generated to better reflect user interests. Further advancements include SoQuES [15], which integrated tag semantic similarity with social proximity, and a graph-based approach [172] that utilized Tag-Topic models and pseudo-relevance feedback for term weighting, tailoring the expansion process to individual user preferences."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.673,
+ 0.882,
+ 0.828
+ ],
+ "angle": 0,
+ "content": "(2). Else. Apart from tagging-based techniques, early research on Personalized Query Expansion primarily focused on modeling user personalization based on search history [66], social networks, or preferences derived from friendship networks [9]. The Axiomatic PQEC framework [79] formalized expansion rules using both local (user behavior-driven) and social (network-driven) strategies. In 2017, WE-LM [144] advanced this paradigm by modeling multi-relational networks with word embeddings across tag-word relationships, refining associations through affinity graphs. Later, PSQE [14] further improved tagging-based methods using utf-iuf user profiling, integrating a tag similarity graph with user profiles in the online phase to compute expansion terms relevant to user interests in real-time, achieving dynamic personalized expansion. In addition, PQEWC [7] leveraged clustering and contextual word embeddings to optimize query expansions dynamically."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.175,
+ 0.506,
+ 0.884,
+ 0.828
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.525,
+ 0.832,
+ 0.534,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "6"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.116,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.119,
+ 0.118,
+ 0.822,
+ 0.282
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.328,
+ 0.289,
+ 0.609,
+ 0.302
+ ],
+ "angle": 0,
+ "content": "Fig. 3. Overview of the personalized retrieval stage."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.333,
+ 0.823,
+ 0.417
+ ],
+ "angle": 0,
+ "content": "4.1.4 Others. Besides query rewriting and query expansion, other personalized query-related research focuses on areas like query disambiguation and query auto-completion [116]. Bobo [33] allows users to input contextual terms reflecting their domain knowledge. In 2019, a method [52] applied fastText embeddings from recent queries to rank candidates. In addition, PSQE [8] employed synthetic user profiles from Wikipedia and word2vec embeddings for query disambiguation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.431,
+ 0.822,
+ 0.497
+ ],
+ "angle": 0,
+ "content": "4.1.5 Discussion. While both query rewriting and query expansion aim to align user input with system understanding to enhance retrieval quality, their roles in personalization differ in fundamental ways. Understanding the distinct operational characteristics and application scenarios of each technique is essential for designing effective personalized retrieval systems. The key takeaways are listed as follows:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.116,
+ 0.507,
+ 0.82,
+ 0.539
+ ],
+ "angle": 0,
+ "content": "- Query rewriting is most beneficial when the original query is ambiguous, underspecified, or misaligned with retrieval intents, particularly in conversational or multi-turn settings."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.116,
+ 0.541,
+ 0.82,
+ 0.573
+ ],
+ "angle": 0,
+ "content": "- Query expansion is most effective when the original query is relevant but incomplete - i.e., when it needs to be semantically broadened to cover additional relevant concepts."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.116,
+ 0.507,
+ 0.82,
+ 0.573
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.116,
+ 0.592,
+ 0.216,
+ 0.604
+ ],
+ "angle": 0,
+ "content": "4.2 Retrieval"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.613,
+ 0.82,
+ 0.68
+ ],
+ "angle": 0,
+ "content": "4.2.1 Definition. The retrieval process involves finding the most relevant documents \\( D^{*} \\) from a corpus \\( C \\) based on a query \\( q^{*} \\), as shown in Figure 3. To incorporate personalization, additional user-specific information \\( p \\) is integrated into the retrieval function \\( \\mathcal{R} \\). This allows the retrieval process to tailor the selected documents to align with individual user preferences or contexts, thereby enhancing the relevance and personalization of the generated outputs."
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.415,
+ 0.689,
+ 0.822,
+ 0.706
+ ],
+ "angle": 0,
+ "content": "\\[\nD ^ {*} = \\mathcal {R} (q ^ {*}, C, p) \\tag {3}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.715,
+ 0.82,
+ 0.764
+ ],
+ "angle": 0,
+ "content": "In the retrieval process, personalization can primarily be introduced by focusing on three steps: indexing, retrieval, and post-retrieval. These steps ensure efficient and accurate retrieval of relevant documents or knowledge, while tailoring the process to individual user preferences. Below, we provide a detailed explanation of each step."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.778,
+ 0.822,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "4.2.2 Indexing. Indexing organizes knowledge base data into a structured format to facilitate efficient retrieval. Within the RAG pipeline, documents are either chunked or entirely encoded into representations before being integrated into searchable systems [30, 117]. Conventional encoding methods employ either sparse encoding techniques (e.g.,"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.464,
+ 0.832,
+ 0.473,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "7"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.88,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.122,
+ 0.881,
+ 0.154
+ ],
+ "angle": 0,
+ "content": "TF-IDF [95], BM25 [102]) or dense encoding approaches leveraging pre-trained models, such as BERT [1], Siamese Encoders [98], or LLM-based encoders [64, 147]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.157,
+ 0.885,
+ 0.274
+ ],
+ "angle": 0,
+ "content": "To introduce personalization at the indexing stage, PEARL [80] generates user embeddings by encoding personal history data with models like DeBERTa. These embeddings are subsequently clustered to create personalized shared indices. Other approaches integrate knowledge graphs into indexing to enhance retrieval performance. For example, KG-R retriever [21] employs a Hierarchical Index Graph, consisting of a knowledge graph layer and a collaborative document layer, to improve RAG retrieval. EMG-RAG [137] incorporates personalized memory within an editable knowledge graph, enabling dynamic retrieval. Similarly, PGraphRAG [5] leverages user-centric knowledge graphs to enhance personalization in retrieval tasks."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.289,
+ 0.884,
+ 0.338
+ ],
+ "angle": 0,
+ "content": "4.2.3 Retrieval. The Retrieval step matches a user query with the indexed knowledge base to fetch relevant candidates. It can be broadly categorized into four different types: (1) Dense Retrieval, (2) Sparse Retrieval, (3) Prompt-based Retrieval, and (4) Others."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.352,
+ 0.885,
+ 0.539
+ ],
+ "angle": 0,
+ "content": "(1). Dense Retrieval. Dense retrieval methods often use vector embeddings and similarity metrics (e.g., cosine similarity) and achieve personalization by encoding user preferences, context, or interactions into query or document embeddings, enabling tailored results through similarity-based matching. For instance, MeMemo [138] retrieves personalized information by matching user-specific embeddings with document vectors, focusing on private, on-device text generation. Similarly, RECAP [71] and LAPDOG [43] enhance personalized dialogue generation by encoding queries and user profiles as dense vectors and retrieving top-N results, ensuring user-specific context drives the responses. In chatbots, Gu et al. [37] integrates conversational context and user profiles to align retrieved responses with user personas. PersonalM [77] employs group-wise contrastive learning, training its retrieval model to align user queries with domain-specific text fragments, thereby improving personalization. UIA [155] employs dual encoders to retrieve documents tailored to user preferences. XPERT [125] incorporates temporal events and user interactions into embeddings, enabling large-scale retrieval across millions of items."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.542,
+ 0.887,
+ 0.66
+ ],
+ "angle": 0,
+ "content": "Dense retrieval also enhances specific applications like e-commerce, medical assistance, and language models. DPSR [157] and RTM [11] encode user queries and product information to personalize product searches dynamically. Pearl [80] and MemPrompt [74] retrieve personalized content by leveraging historical user data and memory-assisted mechanisms. EERRA [23] uses review embeddings as dense queries for recommendations. In medical assistance, MALP [160] and User-LLM [84] integrate short- and long-term user interactions into embeddings for contextualized, personalized responses. Finally, PER-PCS [120] retrieves relevant information using individual user histories, enhancing the personalization capabilities of large language models."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.674,
+ 0.887,
+ 0.828
+ ],
+ "angle": 0,
+ "content": "(2). Sparse Retrieval. Sparse retrieval methods often rely on term-based matching (e.g., BM25) and apply personalization by assigning higher weights to terms or keywords that are more relevant to the user. OPPU [121] uses the BM25 algorithm to select the k most relevant records from the user's historical data for the current query. Similarly, PAG [101] incorporates user input and profiles to enhance summarization and retrieval, aligning sparse representations with personalization objectives for large language models. Au et al. [5] uses BM25 search algorithms to find entries related to the target user or neighboring users through the graph structure. UniMS-RAG [128] combines sparse and dense retrieval by leveraging multi-source knowledge, such as dialogue context and user images, to refine personalized responses in dialogue systems. Lastly, Deng et al. [29] apply sparse retrieval to support fact-based queries, considering user queries and preferences to enhance answer generation for e-commerce applications."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.525,
+ 0.832,
+ 0.535,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "8"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.116,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.122,
+ 0.823,
+ 0.309
+ ],
+ "angle": 0,
+ "content": "(3). Prompt-based Retrieval. Prompt-based retrieval leverages prompts to guide retrieval from the model or external sources and introduces personalization by crafting user-specific prompts that guide the retrieval process. These prompts may include explicit user preferences, historical interactions, or detailed instructions that reflect the user's unique requirements. By embedding this personalized context directly into the prompt, the retrieval process can dynamically adjust to capture and return results that are most relevant to the user. LAPS [50] focuses on multi-session conversational search by storing user preferences and dialogue context, then using prompts to retrieve relevant information tailored to the user's biases and categories of interest. UniMP [140] employs user interaction histories as input to prompt-based retrieval, enabling personalized recommendations for multi-modal tasks, such as vision-language applications, by aligning prompts with user behavioral data. In contrast, Shen et al. [111] explores the use of LLMs to extract empathy and narrative styles from user-provided stories, but this work primarily focuses on style extraction and does not explicitly involve a retrieval component."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.322,
+ 0.824,
+ 0.458
+ ],
+ "angle": 0,
+ "content": "(4). Others. Reinforcement learning-based retrieval personalizes the process by optimizing retrieval policies based on user feedback, learning user preferences over time to adjust strategies. Salemi et al. [103] combines models like BM25, RbR, and dense retrieval, refining them with reinforcement learning (RL) and knowledge distillation (KD) to adapt to user profiles for personalized outputs. Parameter-based retrieval leverages pre-trained model parameters to implicitly store and retrieve user-specific information, allowing direct retrieval from the model without traditional indices. PersonalTM [65] generates document identifiers (Document IDs) using a Transformer model, encoding query, history, and document relationships into its parameters for personalization. Similarly, Zhang et al. [165] uses parameterized representations to integrate user queries and histories, tailoring responses to individual preferences."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.115,
+ 0.122,
+ 0.824,
+ 0.458
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.471,
+ 0.821,
+ 0.52
+ ],
+ "angle": 0,
+ "content": "4.2.4 Post-retrieval. Current Post-Retrieval methods primarily focus on refining retrieved documents or responses to improve relevance and coherence, current methodologies could be categorized into three parts (1) Re-ranking, (2) Summarization, and (3) Compression."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.533,
+ 0.824,
+ 0.669
+ ],
+ "angle": 0,
+ "content": "(1). Re-ranking. Re-ranking enhances personalized content generation by prioritizing more relevant documents at the top. PersonaRAG [156] extends RAG by integrating user-centric agents, such as the Live Session Agent and the Document Ranking Agent, to refine document ranking and improve overall performance. Pavliukevich et al. [89] propose a cross-encoder BERT model for re-ranking external knowledge within a personalized context. UniMS-RAG [128] introduces a scoring mechanism that evaluates retrieved documents and outputs by optimizing the retriever. Besides, it includes an evidence attention mask, enabling re-ranking during inference and applying it to personalized datasets. Salemi and Zamani [106] present an iterative approach to optimizing ranking results based on the expectation-maximization algorithm, with performance validated in personalized scenarios."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.682,
+ 0.821,
+ 0.731
+ ],
+ "angle": 0,
+ "content": "(2). Summarization. Summarization refers to the process of summarizing retrieved information to enhance performance. For instance, Zhang et al. [164] introduced a role-playing agent system to summarize retrieved history in order to improve the final Personalized Opinion Summarization process."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.744,
+ 0.822,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "(3). Compression. Compression involves condensing embeddings or retrieved content to enhance efficiency and effectiveness. Approaches like AutoCompressor [24] compress contextual embeddings into shorter semantic representations, and FIT-RAG [76] introduces a self-knowledge recognizer along with a sub-document-level token reduction mechanism to minimize tokens within RAG pipeline. However, few studies have specifically explored personalized fields, highlighting a promising direction for future research."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.115,
+ 0.533,
+ 0.824,
+ 0.827
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.465,
+ 0.832,
+ 0.473,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "9"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.88,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.122,
+ 0.885,
+ 0.432
+ ],
+ "angle": 0,
+ "content": "4.2.5 Discussion. Indexing, retrieval, and post-retrieval methods each play a critical role in ensuring efficient and personalized information processing, with specific applications and trade-offs. Indexing focuses on organizing knowledge bases for efficient retrieval, using techniques such as sparse encoding methods like TF-IDF and BM25, which are efficient but limited in understanding semantics, and dense encoding methods like BERT and DeBERTa, which provide better semantic understanding but require significant computational resources. These methods are widely used in tasks like question answering and personalized recommendation systems. Retrieval involves matching user queries with relevant documents and can be categorized into dense retrieval, which provides high semantic understanding and personalization but is computationally expensive; sparse retrieval, which is efficient and interpretable but less capable of handling semantics; prompt-based retrieval, which is highly flexible and adaptable to user needs but requires careful engineering of prompts; and advanced methods like reinforcement learning-based approaches, which dynamically adapt to user feedback but are complex to implement. This step is essential in applications like personalized dialogue systems, search engines, and e-commerce. Post-retrieval methods refine retrieved results to enhance relevance and coherence through re-ranking, which improves personalization and prioritizes relevant content but increases computational overhead; summarization, which simplifies complex information for better user understanding but risks losing critical details; and compression, which reduces computational costs by condensing information but remains underexplored in personalized contexts. Together, these methods provide a comprehensive pipeline for delivering efficient, relevant, and personalized outputs, balancing their strengths in semantic understanding, relevance, and flexibility with challenges related to computational costs and implementation complexity."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.176,
+ 0.448,
+ 0.29,
+ 0.459
+ ],
+ "angle": 0,
+ "content": "4.3 Generation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.468,
+ 0.884,
+ 0.518
+ ],
+ "angle": 0,
+ "content": "4.3.1 Definition. Personalized generation incorporates user-specific retrieved documents \\( D^{*} \\), task-specific prompt prompt, and user preference information \\( p \\) via the generator \\( \\mathcal{G} \\) parameterized by \\( \\theta \\) to produce tailored content \\( g^{*} \\) aligned with individual preference, where the flow is shown in Figure 4. The generation process can be formulated as"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.448,
+ 0.527,
+ 0.881,
+ 0.543
+ ],
+ "angle": 0,
+ "content": "\\[\ng ^ {*} = \\mathcal {G} (D ^ {*}, \\text {p r o m p t}, p, \\theta). \\tag {4}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.553,
+ 0.884,
+ 0.636
+ ],
+ "angle": 0,
+ "content": "Personalized generation can be achieved by incorporating explicit and implicit preferences. Explicit preference-driven methodologies utilize direct input signals (e.g., \\( D^{*} \\), prompt, and \\( p \\)), to tailor outputs to specific user preferences. Conversely, implicit preference-encoded approaches embed personalized information within the parameters \\( \\theta \\) of the generator model, during training, thereby facilitating preference alignment without the necessity for explicit runtime inputs."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.648,
+ 0.884,
+ 0.75
+ ],
+ "angle": 0,
+ "content": "4.3.2 Generation from Explicit Preferences. Integrating explicit preferences into LLMs facilitates personalized content generation. Explicit preference information encompasses user demographic information (e.g., age, occupation, gender, location), user behavior sequences (reflecting historical behavioral patterns), and user historical output texts (capturing writing style and tone preferences). The injection of explicit preferences for personalized generation can be categorized into three types: (1) Direct-integrated Prompting, (2) Summary-augmented Prompting, and (3) Adaptive Prompting."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.762,
+ 0.884,
+ 0.829
+ ],
+ "angle": 0,
+ "content": "(1). Direct-integrated Prompting. Integrating user explicit preferences into language models through prompting enables the prediction of users' intent and behavioral patterns, facilitating personalized content generation. For instance, \\(\\mathrm{P}^2\\) [49], Character Profiling [154], and OpinionQA [107] integrate personalized data into LLMs through prompting for role-playing task, thereby aligning the model's responses with specified user profiles. Kang et al. [51] and Liu et al. [67]"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.523,
+ 0.832,
+ 0.536,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "10"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.116,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.137,
+ 0.121,
+ 0.803,
+ 0.293
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.32,
+ 0.31,
+ 0.616,
+ 0.324
+ ],
+ "angle": 0,
+ "content": "Fig. 4. Overview of the personalized generation stage."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.35,
+ 0.823,
+ 0.435
+ ],
+ "angle": 0,
+ "content": "integrate interaction histories into LLMs via prompting to predict user rating for candidate items. Cue-CoT [129] employs chain-of-thought reasoning to infer user needs from contextual cues, enabling personalized responses to in-depth dialogue questions. Additionally, TICL [26] proposes a trial-and-error framework that critiques initial LLM-generated responses, derives explanations and integrates these negative examples into prompts to improve personalization alignment."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.447,
+ 0.824,
+ 0.601
+ ],
+ "angle": 0,
+ "content": "(2). Summary-augmented Prompting. Direct integration of personalized information via prompting struggles with ambiguous intent signals: Lengthy interaction histories introduce noise that obscures critical behavioral patterns [69], while sparse behavioral data lacks sufficient context for LLMs to derive meaningful user preferences. To address these issues, recent approaches focus on summarizing user personalized intents and integrating them into prompts. For instance, GPG [158] extracts key user habits and preferences from personal contexts, enabling fine-grained personalization. Similarly, LLMs are employed to generate task-specific summaries of user preferences, enhancing retrieval-augmented personalized generation capabilities [101]. In recommendation systems, ONCE [70], LLMTreeRec [163], and KAR [145] leverage historical user-item interactions to summarize user preferences. Furthermore, Matryoshka [58] generates user preference summaries by dynamically retrieving and synthesizing historical data."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.613,
+ 0.824,
+ 0.766
+ ],
+ "angle": 0,
+ "content": "(3). Adaptive Prompting. Manually designing personalized prompts demands both expert knowledge and significant labor, motivating the development of automated methods for personalized prompt generation. For example, Li et al. [57] trains a personalized prompt rewriter via supervised and reinforcement learning. RecGPT [166] and PEPLER-D [59] leverage prompt tuning to generate personalized prompts, enhancing sequential and explainable recommendations, respectively. GRAPA [94] integrates semantic and collaborative signals from user-item interaction graphs with graph neural networks to generate context-aware personalized prompts. SGPT [28] employs prompt tuning to jointly model common and group-specific patterns, bridging generalized and personalized federated learning paradigms. Furthermore, PFCL [152] achieves multi-granularity human preference modeling: coarse-grained prompts distill shared knowledge, while fine-grained prompts adapt to individual user characteristics."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.114,
+ 0.447,
+ 0.824,
+ 0.766
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.779,
+ 0.824,
+ 0.828
+ ],
+ "angle": 0,
+ "content": "4.3.3 Generation from Implicit Preferences. Unlike explicit preference modeling, which captures user preferences through textual input, implicit preference-based methods incorporate personalization through internal parameters. This personalization is achieved either through Parameter-Efficient Fine-tuning (PEFT) techniques, such as LoRA [42],"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.463,
+ 0.832,
+ 0.475,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "11"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.881,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.122,
+ 0.884,
+ 0.155
+ ],
+ "angle": 0,
+ "content": "or reinforcement learning-based approaches for preference alignment [20, 57]. Based on these strategies, we classify existing methods into two categories: (1) Fine-tuning-Based Methods and (2) Reinforcement Learning-Based Methods."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.167,
+ 0.885,
+ 0.32
+ ],
+ "angle": 0,
+ "content": "(1). Fine-tuning Based Methods. For fine-tuning methods, LoRA is the most widely adopted since it is resource-efficient and enables rapid adaptation without compromising model performance. PLoRA [165] introduces a personalized knowledge integration framework that combines task-specific LoRA with user-specific knowledge. Similarly, LM-P [142] personalizes information via LoRA by incorporating User ID as a personalization factor. MiLP [165] employs Bayesian optimization to determine the optimal personalization injection configuration, including LoRA settings, to effectively capture and utilize user-specific information. OPPU [122] and PER-PCS [120] follow a similar approach, leveraging user history data for fine-tuning LoRA-based personalization. However, PER-PCS differs by incorporating a gating module that selects the appropriate LoRA, enabling fine-grained personalization. Additionally, Review-LLM [91] integrates LoRA for supervised fine-tuning in the task of personalized review generation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.323,
+ 0.884,
+ 0.424
+ ],
+ "angle": 0,
+ "content": "Beyond LoRA-based approaches, alternative pipelines have been proposed for personalized generation. UserIdentifier [78] introduces a user-specific identifier, significantly reducing training costs while enhancing personalized demonstration. UserAdapter [171] proposes user-independent prefix embeddings, leveraging prefix tuning for personalization. Meanwhile, HYDRA [175] achieves implicit personalization by training user-specific headers. Recently, researchers have also explored fine-tuning personalized model on edge devices [90] and collaborative learning between small and large language models to enable more personalized generation [161]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.438,
+ 0.884,
+ 0.695
+ ],
+ "angle": 0,
+ "content": "(2). Reinforcement Learning Based Methods. Apart from fine-tuning based methods, recent research has explored reinforcement learning based techniques to personalize text generation by aligning outputs with user preferences. P-RLHF [62] has been proposed to jointly learn a user-specific and reward model to enable text generation that aligns with a user's styles or criteria. P-SOUPS [47] models multiple user preferences as a Multi-Objective Reinforcement Learning (MORL) problem, decomposing preferences into multiple dimensions, each trained independently. PAD [20] aligns text generation with human preferences during inference by utilizing token-level personalized rewards to guide the decoding process. REST-PG [104] introduces a framework that trains large language models to reason over personal data during response generation. This approach first generates reasoning paths to enhance the LLM's reasoning ability and then employs Expectation-Maximization Reinforced Self-Training to iteratively refine the model based on its high-reward outputs. Additionally, Salemi et al. [103] incorporate reinforcement learning into the RAG pipeline to improve retrieval accuracy, thereby enhancing the personalization of generated content. Other applications include RewriterSIRI [57], which has been introduced to generate text via RL-based personalized prompt rewriting using API-based LLMs, and Kulkarni et al. [54], who explore the use of reinforcement learning to optimize RAG for improving the relevance and coherence of chatbot responses in specialized domains, ultimately enhancing user satisfaction and engagement."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.709,
+ 0.884,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "4.3.4 Discussion. Personalized generation can be adopted via both explicit and implicit preference injection, yet they exhibit distinct characteristics that make them suitable for different scenarios. In explicit preference-based generation, personalization is clearly defined through user profile descriptions, contextual information, and similar inputs, which are incorporated into generators via prompts. A key advantage of this approach is explainability, as the personalized information is explicitly provided and easily traceable. Despite leveraging provided preferences and internal knowledge, explicit preference injection's personalization is constrained by model capabilities and irrelevant information interference. In contrast, implicit preference-based generation internalizes personalized information into"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.523,
+ 0.832,
+ 0.536,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "12"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.116,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.123,
+ 0.119,
+ 0.818,
+ 0.491
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.262,
+ 0.502,
+ 0.677,
+ 0.516
+ ],
+ "angle": 0,
+ "content": "Fig. 5. Overview of transition from personalized RAG to personalized agent."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.556,
+ 0.823,
+ 0.622
+ ],
+ "angle": 0,
+ "content": "the generator's parameters through scene-specific personalized data, thereby adapting the model for more fine-grained personalization. However, these methods typically incur substantial training and computational costs, as they require fine-tuning the generator's internal parameters. Therefore, selecting between these approaches should be guided by the specific application scenario and resource constraints."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.116,
+ 0.643,
+ 0.285,
+ 0.657
+ ],
+ "angle": 0,
+ "content": "4.4 From RAG to Agent"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.114,
+ 0.664,
+ 0.823,
+ 0.749
+ ],
+ "angle": 0,
+ "content": "4.4.1 Definition. A personalized LLM-based agent is a system designed to dynamically incorporate user context, memory, and external tools or APIs to support highly personalized and goal-oriented interactions [19, 45, 146], and solve problems in a goal-oriented manner [63, 113]. From the previously introduced stages of RAG, we observe that the evolution of personalized RAG reveals a structural convergence with agent architectures. We analyze them from three key perspectives:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.116,
+ 0.759,
+ 0.822,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "- Personalized Understanding: This phase within the agent parallels the query understanding and rewriting process of RAG as outlined in Section 4.1. However, it extends beyond static semantic parsing by incorporating dynamic user profiling [139] and role modeling [110]. This integration enables the agent to dynamically align interactions with implicit user preferences, facilitating personalized responses and task-specific adaptations [96]."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.463,
+ 0.832,
+ 0.475,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "13"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.881,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.122,
+ 0.88,
+ 0.205
+ ],
+ "angle": 0,
+ "content": "- Personalized Planning and Execution: This phase in agents mirrors RAG's retrieval process in Section 4.2 yet it advances beyond static document retrieval by incorporating real-time memory management [87] and sophisticated tool and API calling [127]. This approach ensures the dynamic alignment of external knowledge with personalized constraints, such as integrating medical history in healthcare agents [2], to deliver context-aware and user-specific outcomes."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.209,
+ 0.881,
+ 0.293
+ ],
+ "angle": 0,
+ "content": "- Personalized Generation: This phase in agents mirrors RAG's generative process in Section 4.3 but transcends static template-based generation by integrating user preference and fact alignment. Agents dynamically enforce user preferences and ensure fact consistency through role-specific mechanisms (e.g., social adaptability in conversational agents [2]), enabling outputs to evolve in harmony with personalized and situational constraints rather than relying solely on predefined generative frameworks."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.176,
+ 0.122,
+ 0.881,
+ 0.293
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.301,
+ 0.881,
+ 0.384
+ ],
+ "angle": 0,
+ "content": "In general we frame agent architectures as \"personalized RAG++\", where persistent memory [137] replaces static indexes, and tool APIs [17] serve as dynamic knowledge connectors, enabling complicated, human-aligned interactions beyond one-shot retrieval, as shown in Figure 5. This progression highlights that as RAG systems incorporate deeper personalization—requiring user-state tracking, adaptive tool usage, and context-aware generation, they inherently adopt agent-like capabilities."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.399,
+ 0.881,
+ 0.482
+ ],
+ "angle": 0,
+ "content": "4.4.2 Personalized Understanding. Personalized understanding refers to an agent's ability to accurately interpret user inputs by integrating user intent recognition and contextual analysis. This process ensures interactions that are both meaningful and contextually appropriate. The rationale behind this classification lies in its capacity to address three core aspects of understanding: recognizing user intent, analyzing context, and leveraging user profiles. Each of these aspects plays a distinct role in improving the agent's performance."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.496,
+ 0.88,
+ 0.614
+ ],
+ "angle": 0,
+ "content": "(1). User-profile Understanding. In user-profile understanding, an agent's personalized ability primarily depends on its capacity to accurately model and understand the user's preferences, context, and intentions. Xu et al. [148] proposes a framework in which LLMs are designed to understand the physical world, thereby facilitating a deeper connection between the agent and its environment, which is essential for accurate task execution. Abbasian et al. [2] further expands this understanding by emphasizing the importance of personalization in health agents, where the user's profile directly influences the behavior and decisions of the agent. This user understanding is foundational to ensuring that the AI agent performs tasks in a way that aligns with individual user needs."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.629,
+ 0.88,
+ 0.764
+ ],
+ "angle": 0,
+ "content": "(2). Role Understanding. In agent's role understanding, the role of the agent within these environments is also crucial. Recent studies focus on enhancing role-playing capabilities within LLMs. Wang et al. [139] introduce RoleLLM, a benchmark that aims to elicit and refine the role-playing abilities of LLMs, demonstrating how role understanding influences agent performance in conversational tasks. Similarly, Shao et al. [110] present Character-LLM, a trainable agent framework for role-playing, which tailors its responses based on predefined roles. Wang et al. [134] introduce a method for evaluating personality fidelity in role-playing agents through psychological interviews, aiming to enhance the realism and consistency of AI-driven characters. This role understanding allows for more contextually appropriate interactions, increasing the relevance and utility of AI agents across various applications."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.779,
+ 0.881,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "(3). User-role Joint Understanding. In agent's user-role joint understanding, the intersection of user and role understanding is explored through frameworks that evaluate and enhance the social and personality aspects of LLMs. SocialBench Chen et al. [18] provides a sociality evaluation framework for role-playing agents. Dai et al. [27], and"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.175,
+ 0.496,
+ 0.881,
+ 0.827
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.523,
+ 0.832,
+ 0.536,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "14"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.116,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.122,
+ 0.823,
+ 0.205
+ ],
+ "angle": 0,
+ "content": "[96] extend this by incorporating multi-modal data and personality-indicative information, respectively, which allows agents to better adapt to both user and role understanding in dynamic environments. Furthermore, Wang et al. [126] offers a perspective on how role and environment understanding can improve user experience. Tu et al. [123] contribute by providing a benchmark specifically for evaluating role-playing agents in the Chinese context, adding a cultural dimension to role understanding. Finally, Neeko [153] further advances role-based interactions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.221,
+ 0.824,
+ 0.306
+ ],
+ "angle": 0,
+ "content": "4.4.3 Personalized Planning and Execution. Personalized planning and execution refer to the process of designing and implementing strategies or actions that are specifically tailored to an individual's unique context, and goals [44, 87, 114, 159]. It requires agents to dynamically integrate long-term memory, real-time reasoning, and external tool utilization [40, 41, 169], as demonstrated in healthcare decision support [2] and travel planning scenarios [17]. We analyze two fundamental components that enable this personalization in the following."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.32,
+ 0.824,
+ 0.473
+ ],
+ "angle": 0,
+ "content": "(1). Memory Management. Effective memory systems allow agents to integrate users' historical preferences, behavioral patterns, and contextual habits, enhancing their ability to make planning and tailor interactions to user-specific needs [17, 127, 136]. The EMG-RAG framework [137] combines editable memory graphs with retrieval-augmented generation to maintain dynamic user profiles, while Park et al. [87] implements memory streams and periodic reflection mechanisms to simulate human-like behavior. In healthcare applications, Abbasian et al. [2] integrates multimodal user data through specialized memory modules to optimize treatment recommendations. For recommendation systems, RecAgent [133] employs hierarchical memory structures to model user interaction patterns across multiple domains. Recent advances like TravelPlanner+ [114] demonstrate how memory-augmented LLMs achieve higher relevance in personalized itinerary generation compared to generic planners."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.488,
+ 0.824,
+ 0.642
+ ],
+ "angle": 0,
+ "content": "(2). Tool and API Calling. The integration of external tools expands agents' capabilities beyond pure linguistic reasoning, enabling agents to interact with users and perform personalized tasks [17, 126, 127, 148, 159]. For instance, VOYAGER [127] establishes a paradigm for lifelong skill acquisition through automatic API curriculum learning and skill library construction. In robotics, Zhang et al. [159] develops a bootstrapping framework where LLMs guide robots in tool-mediated skill discovery, enabling a high success rate in novel object manipulation tasks. The PUMA framework [17] demonstrates how personalized web agents can achieve performance gains in e-commerce tasks through adaptive API orchestration. For mobile interaction, Wang et al. [126] implements few-shot tool learning to handle diverse UI operations with minimal training data. These approaches highlight the importance of tool grounding mechanisms [44] that translate linguistic plans into executable API sequences while maintaining personalization constraints."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.115,
+ 0.32,
+ 0.824,
+ 0.642
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.644,
+ 0.824,
+ 0.744
+ ],
+ "angle": 0,
+ "content": "This synthesis highlights that modern agent systems achieve enhanced personalization through two primary strategies: 1) Memory-augmented architectures, which leverage editable memory graphs [137], reflection mechanisms [87], and hierarchical memory structures [133] to dynamically adapt to user preferences across various domains; and 2) Tool and API integration, which expand agent capabilities by balancing generalization with specialization. Future work may explore improving the contextual relevance and adaptability of memory systems while optimizing real-time tool interaction for seamless task execution."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.76,
+ 0.824,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "4.4.4 Personalized Generation. Based on the foundation of personalized planning and execution mechanisms, which enable agents to adapt strategies to user-specific contexts [44, 159], the next critical concern lies in personalized generation. This capability ensures that generated outputs not only align with factual correctness but also resonate with users' unique preferences, personality traits, and situational needs. Personalized generation bridges the gap between"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.463,
+ 0.832,
+ 0.475,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "15"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.882,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.122,
+ 0.882,
+ 0.155
+ ],
+ "angle": 0,
+ "content": "adaptive reasoning and human-aligned outcomes, allowing agents to produce contextually relevant and emotionally appropriate responses."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.177,
+ 0.885,
+ 0.332
+ ],
+ "angle": 0,
+ "content": "(1). Alignment with User Fact. Alignment with User Fact emphasizes the accuracy, consistency, and factual grounding of personalized responses, ensuring they remain trustworthy across diverse user interactions. This is particularly challenging in personalized agents, where maintaining character authenticity while avoiding hallucinations requires balancing creativity with factual adherence. Recent advances address these challenges through improved training frameworks and evaluation metrics. For instance, Character-LLM [110] integrates memory-augmented architectures to reduce hallucinations while preserving character-specific traits. Wang et al. [135] investigate quantization effects on personality consistency in edge-deployed agents and stabilize outputs under computational constraints. Dai et al. [27] ensures multimodal consistency (text-image) in role-playing. These works highlight the importance of architectural innovations and rigorous evaluation in achieving reliability."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.354,
+ 0.887,
+ 0.44
+ ],
+ "angle": 0,
+ "content": "(2). Alignment with User Preferences. Alignment with user preferences ensures that generated outputs reflect individualized personalities, values, and interaction styles. This requires agents to dynamically interpret implicit user cues and adapt responses accordingly. Wang et al. [139] benchmarks role-specific alignment. Ran et al. [96] improves personality fidelity via psychological scale datasets. Wang et al. [134] quantifies alignment via psychological interviews. Chen et al. [18] evaluates social adaptability in conversations."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.175,
+ 0.177,
+ 0.887,
+ 0.44
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.462,
+ 0.882,
+ 0.493
+ ],
+ "angle": 0,
+ "content": "4.4.5 Discussion. The architectural evolution from RAG to personalized agents introduces significant advancements in human-AI interaction but also surfaces critical challenges that warrant further investigation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.496,
+ 0.885,
+ 0.614
+ ],
+ "angle": 0,
+ "content": "Personalized Understanding, while enabling interpretation of user intent and context, faces limitations in real-time adaptability and generalization. Current approaches like RoleLLM [139] and Character-LLM [110] demonstrate robust role-specific comprehension but struggle with dynamic user state tracking, particularly when handling evolving preferences or multi-session interactions. Furthermore, cultural specificity in benchmarks like CharacterEval [123] reveals gaps in global applicability, as agents trained on region-specific data often fail to generalize across diverse sociocultural contexts. Future work could explore hybrid architectures that combine continuous learning mechanisms with privacy-preserving federated learning to address these adaptability constraints while maintaining user trust."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.618,
+ 0.885,
+ 0.737
+ ],
+ "angle": 0,
+ "content": "Personalized Planning and Execution, achieves remarkable task specialization through memory management and tool integration, yet suffers from scalability issues in complex environments. While frameworks like EMG-RAG [137] and VOYAGER [127] effectively manage user-specific constraints, their reliance on predefined API taxonomies limits emergent tool discovery in novel scenarios. The \"cold-start\" problem persists in domains requiring rapid skill acquisition, as seen in healthcare applications [2], where delayed API responses can compromise decision-making efficacy. A promising direction involves developing meta-reasoning architectures that dynamically prioritize memory recall versus tool invocation based on situational urgency and confidence thresholds."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.739,
+ 0.885,
+ 0.824
+ ],
+ "angle": 0,
+ "content": "Personalized Generation balances factual accuracy with preference alignment but risks over-fitting, where excessive finetuning to user profiles may reinforce cognitive biases. Techniques address surface-level alignment but lack mechanisms for ethical boundary detection. For instance, agents might inadvertently propagate harmful stereotypes when mirroring user preferences without critical oversight. Future systems could integrate value-aligned reinforcement learning with human-in-the-loop validation to preserve authenticity while preventing detrimental customization."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.523,
+ 0.832,
+ 0.536,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "16"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.117,
+ 0.091,
+ 0.347,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.82,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.297,
+ 0.119,
+ 0.64,
+ 0.132
+ ],
+ "angle": 0,
+ "content": "Table 2. Datasets and metrics for personalized RAG and Agent."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.12,
+ 0.139,
+ 0.825,
+ 0.489
+ ],
+ "angle": 0,
+ "content": "| Field | Metrics Category | Metrics | Datasets |
| Pre-retrieval | Textual Quality | BLEU, ROUGE, EM | Avocado Research Email Collection [57, 85], Amazon review [57, 83], Reddit comments[57, 118], Amazon ESCI dataset[82, 97], PIP |
| Information Retrieval | MAP, MRR, NDCG, Precision, Recall, RBP | AOL[88, 174], WARRIORS[99], Personalized Results Re-Ranking benchmark [6], delicio.us [9, 15, 144, 172], Flickr [9, 108], CiteULike [10, 14], LRDP [12], Delicious [141], Bibsonomy [79], Wikipedia [8, 33] |
| Classification | Accuracy, Macro-F1 | SCAN [56, 173], AITA WORKSM[53, 80], Robust04 [61] |
| Others | XEntropy, PMS, Image-Align, PQEC, ProfOverlap | Amazon ESCI dataset[82, 97], PIP, Bibsonomy [79] |
| Retrieval | Textual Quality | BLEU, ROUGE, Dis, PPL | TOPDIAL [130], Pchatbot [93], DuLemon [150] |
| Information Retrieval | Recall, MRR, Precision, F1 | LiveChat [34], Pchatbot [93], DuLemon [150] |
| Classification | Accuracy, Succ | TOPDIAL [130], PersonalityEvid [119], DuLemon [150], PersonalityEdit [75] |
| Others | Fluency, Coherence, Plausibility, ES, DD, TPEI, PAE | PersonalityEvid [119], PersonalityEdit [75] |
| Generation | Textual Quality | BLEU, ROUGE, Dis, PPL, METEOR | LaMP [105], Long LaMP [55], Dulemon [150], PGraphRAG [5], AmazonQA/Products [29], Reddit [170], MedicalDialogue [162] |
| Classification | Accuracy, F1, Persona F1 | LaMP [105], Long LaMP [55], Dulemon [150], AmazonQA/Products [29], Reddit [170], MedicalDialogue [162] |
| Regression | MAE, RMSE | LaMP [105], Long LaMP [55], PGraphRAG [5] |
| Others | Fluency, Mean Success Rate, Median Relative Improvements | Personalized-Gen [3] |
| Agent | Textual Quality | BLEU, ROUGE, METEOR, CIDer, EM, Fluency, Coherence, Instruction Adherence, Consistency related metrics | RICO [126], RoleBench [139], Shao et al. [110], Socialbench [18], MMRole-Data [27], ROLEPERSONALITY [96], ChatHarui [134], Character-LLM-Data [153], Knowledge Behind Persona [41], Wang et al. [137], Wang et al. [135], Zheng et al. [169] |
| Information Retrieval | Recall, F1, Precision | Knowledge Behind Persona [41] |
| Classification | Accuracy, Failure Rate, Classification Accuracy, Preference Rate, Correctness | MIT-BIH Arrhythmia Database [148], VirtualHome [44], Socialbench [18], ARC [100], AGIEval [100], HellaSwag [100], MedMCQA [100], AQUA-RAT [100], LogiQA [100], LSAT-AR [100], LSAT-LR [100], LSAT-RC [100], SAT-English [100], SAT-Math [100], PersonalWAB [17], TravelPlanner+ [114] |
| Others | Pass@k, Executability, Productivity, Plausibility of the Story | Hong et al. [40], Zheng et al. [169] |
"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.116,
+ 0.511,
+ 0.339,
+ 0.523
+ ],
+ "angle": 0,
+ "content": "5 EVALUATION AND DATASET"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.532,
+ 0.823,
+ 0.72
+ ],
+ "angle": 0,
+ "content": "In the evolving landscape of personalization, from RAG to advanced Agent-based systems, the evaluation of models relies heavily on diverse datasets and metrics tailored to specific tasks. This survey categorizes metrics into several key types: Textual Quality metrics (e.g., BLEU, ROUGE, METEOR) assess the fluency and coherence of generated outputs; Information Retrieval metrics (e.g., MAP, MRR, Recall) evaluate the accuracy and relevance of retrieved information; Classification metrics (e.g., Accuracy, F1) measure task-specific correctness; Regression metrics (e.g., MAE, RMSE) quantify prediction errors; and Other metrics (e.g., Fluency, Pass@k) address domain-specific or task-unique aspects like plausibility or executability. These metrics span pre-retrieval, retrieval, generation, and agent-based personalization approaches, reflecting their varied objectives. To provide a comprehensive overview, we compile an extensive list of datasets across these fields, as detailed in Table 2. These datasets, paired with their respective metrics, enable researchers to benchmark and refine personalized systems, from enhancing query rewriting to enabling autonomous agents in physical and virtual environments."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.116,
+ 0.736,
+ 0.431,
+ 0.749
+ ],
+ "angle": 0,
+ "content": "6 CHALLENGES AND FUTURE DIRECTIONS"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.115,
+ 0.757,
+ 0.82,
+ 0.789
+ ],
+ "angle": 0,
+ "content": "Personalized RAG and agent-based systems still face several critical challenges that warrant further exploration. We list them as follows:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.116,
+ 0.796,
+ 0.82,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "- Balancing Personalization and Scalability: Integrating personalization data (such as preferences, history, and contextual signals) into RAG processes often increases computational complexity, making it difficult to maintain"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.463,
+ 0.832,
+ 0.475,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "17"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.88,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.187,
+ 0.122,
+ 0.882,
+ 0.154
+ ],
+ "angle": 0,
+ "content": "efficiency and scalability across large-scale systems. Future work could explore lightweight, adaptive embeddings and hybrid frameworks that seamlessly fuse user profiles with real-time contexts."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.157,
+ 0.885,
+ 0.224
+ ],
+ "angle": 0,
+ "content": "- Evaluating Personalization Effectively: Current metrics like BLEU, ROUGE, and human evaluations fall short in capturing the nuanced alignment of outputs with dynamic user preferences, lacking tailored measures for personalization efficacy. Developing specialized benchmarks and metrics that assess long-term user satisfaction and adaptability is crucial for real-world applicability."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.226,
+ 0.884,
+ 0.31
+ ],
+ "angle": 0,
+ "content": "- Preserving Privacy through Device-Cloud Collaboration: Personalized retrieval often involves processing sensitive user data, raising privacy concerns, especially with the increased global emphasis on data protection regulations, such as the European Union's General Data Protection Regulation (GDPR). Consequently, a promising approach is the collaborative integration of on-device small Language models which handle sensitive personal data locally, with cloud-based LLM, which provides broader contextual knowledge."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.312,
+ 0.884,
+ 0.395
+ ],
+ "angle": 0,
+ "content": "- Personalized Agent Planning: Current research on agent planning remains mainly in its early stages, with much of the work focusing on building foundational frameworks such as GUI agents [81] and the application of agents across diverse domains [131]. Notably, the incorporation of personalized approaches has yet to be widely adopted. Exploring how to integrate personalized support into existing frameworks to enhance user experience represents a promising and valuable direction for future investigation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.399,
+ 0.884,
+ 0.449
+ ],
+ "angle": 0,
+ "content": "- Ensuring Ethical and Coherent Systems: Bias in data processing, privacy concerns in user profiling, and coherence across retrieval and generation stages remain unresolved. Future directions should prioritize ethical safeguards, privacy-preserving techniques, and cross-stage optimization to build trustworthy, unified personalized systems."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.176,
+ 0.157,
+ 0.885,
+ 0.449
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.177,
+ 0.469,
+ 0.303,
+ 0.482
+ ],
+ "angle": 0,
+ "content": "7 CONCLUSION"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.175,
+ 0.49,
+ 0.884,
+ 0.609
+ ],
+ "angle": 0,
+ "content": "In this paper, we explore the landscape of personalization from Retrieval-Augmented Generation (RAG) to advanced LLM-based Agents, detailing adaptations across pre-retrieval, retrieval, and generation stages while extending into agentic capabilities. By reviewing recent literature, datasets, and metrics, we highlight the progress and diversity in enhancing user satisfaction through tailored AI systems. However, challenges such as scalability, effective evaluation, and ethical concerns underscore the need for innovative solutions. Future research should focus on lightweight frameworks, specialized benchmarks, and privacy-preserving techniques to advance personalized AI. Relevant papers and resources are also compiled online for ease of future research."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.177,
+ 0.631,
+ 0.274,
+ 0.643
+ ],
+ "angle": 0,
+ "content": "REFERENCES"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.187,
+ 0.649,
+ 0.804,
+ 0.661
+ ],
+ "angle": 0,
+ "content": "[1] 2021. BERT: a review of applications in natural language processing and understanding. arXiv preprint arXiv:2103.11943 (2021)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.187,
+ 0.662,
+ 0.882,
+ 0.685
+ ],
+ "angle": 0,
+ "content": "[2] Mahyar Abbasian, Iman Azimi, Amir M Rahmani, and Ramesh Jain. 2023. Conversational health agents: A personalized llm-powered agent framework. arXiv preprint arXiv:2310.02374 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.187,
+ 0.687,
+ 0.881,
+ 0.711
+ ],
+ "angle": 0,
+ "content": "[3] Bashar Alhafni, Vivek Kulkarni, Dhruv Kumar, and Vipul Raheja. 2024. Personalized Text Generation with Fine-Grained Linguistic Control. In Proceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024). 88–101."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.188,
+ 0.712,
+ 0.698,
+ 0.724
+ ],
+ "angle": 0,
+ "content": "[4] Amazon. [n.d.]. Amazon Customer Review Dataset. Online dataset. https://nijianmo.github.io/amazon/"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.188,
+ 0.726,
+ 0.883,
+ 0.76
+ ],
+ "angle": 0,
+ "content": "[5] Steven Au, Cameron J Dimacali, Ojasmitha Pedirappagari, Namyong Park, Franck Dernoncourt, Yu Wang, Nikos Kanakaris, Hanieh Deilamsalehy, Ryan A Rossi, and Nesreen K Ahmed. 2025. Personalized Graph-Based Retrieval for Large Language Models. arXiv preprint arXiv:2501.02157 (2025)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.188,
+ 0.762,
+ 0.881,
+ 0.786
+ ],
+ "angle": 0,
+ "content": "[6] Elias Bassani, Pranav Kasela, Alessandro Raganato, and Gabriella Pasi. 2022. A multi-domain benchmark for personalized search evaluation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3822-3827."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.188,
+ 0.788,
+ 0.881,
+ 0.811
+ ],
+ "angle": 0,
+ "content": "[7] Elias Bassani, Nicola Tonellotto, and Gabriella Pasi. 2023. Personalized query expansion with contextual word embeddings. ACM Transactions on Information Systems 42, 2 (2023), 1-35."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.187,
+ 0.813,
+ 0.862,
+ 0.825
+ ],
+ "angle": 0,
+ "content": "[8] Oliver Baumann and Mirco Schoenfeld. 2024. PSQE: Personalized Semantic Query Expansion for user-centric query disambiguation. (2024)."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.187,
+ 0.649,
+ 0.883,
+ 0.825
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.523,
+ 0.832,
+ 0.536,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "18"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.117,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.128,
+ 0.124,
+ 0.822,
+ 0.16
+ ],
+ "angle": 0,
+ "content": "[9] Matthias Bender, Tom Crecelius, Mouna Kacimi, Sebastian Michel, Thomas Neumann, Josiane Xavier Parreira, Ralf Schenkel, and Gerhard Weikum. 2008. Exploiting social relations for query expansion and result ranking. In 2008 IEEE 24th International Conference on Data Engineering Workshop. IEEE, 501-506."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.161,
+ 0.823,
+ 0.186
+ ],
+ "angle": 0,
+ "content": "[10] Marin Bertier, Rachid Guerraoui, Vincent Leroy, and Anne-Marie Kermarrec. 2009. Toward personalized query expansion. In Proceedings of the Second ACM EuroSys Workshop on Social Network Systems. 7-12."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.187,
+ 0.822,
+ 0.211
+ ],
+ "angle": 0,
+ "content": "[11] Keping Bi, Qingyao Ai, and W Bruce Croft. 2021. Learning a fine-grained review-based transformer model for personalized product search. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 123-132."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.212,
+ 0.822,
+ 0.236
+ ],
+ "angle": 0,
+ "content": "[12] Claudio Biancalana and Alessandro Micarelli. 2009. Social tagging in query expansion: A new way for personalized web search. In 2009 International Conference on Computational Science and Engineering, Vol. 4. IEEE, 1060-1065."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.237,
+ 0.461,
+ 0.249
+ ],
+ "angle": 0,
+ "content": "[13] Microsoft Bing. [n.d]. Bing Search Engine. https://www.bing.com"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.25,
+ 0.822,
+ 0.274
+ ],
+ "angle": 0,
+ "content": "[14] Mohamed Reda Bouadjenek, Hakim Hacid, and Mokrane Bouzeghoub. 2019. Personalized social query expansion using social annotations. Transactions on Large-Scale Data-and Knowledge-Centered Systems XL (2019), 1-25."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.275,
+ 0.822,
+ 0.311
+ ],
+ "angle": 0,
+ "content": "[15] Mohamed Reda Bouadjenek, Hakim Hacid, Mokrane Bouzeghoub, and Johann Daigremont. 2011. Personalized social query expansion using social bookmarking systems. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval. 1113-1114."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.312,
+ 0.795,
+ 0.324
+ ],
+ "angle": 0,
+ "content": "[16] Domenico Bulfamante. 2023. Generative enterprise search with extensible knowledge base using ai. Ph.D. Dissertation. Politecnico di Torino."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.326,
+ 0.822,
+ 0.349
+ ],
+ "angle": 0,
+ "content": "[17] Hongru Cai, Yongqi Li, Wenjie Wang, ZHU Fengbin, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. [n. d]. Large Language Models Empowered Personalized Web Agents. In THE WEB CONFERENCE 2025."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.35,
+ 0.822,
+ 0.374
+ ],
+ "angle": 0,
+ "content": "[18] Hongzhan Chen, Hehong Chen, Ming Yan, Wenshen Xu, Xing Gao, Weizhou Shen, Xiaojun Quan, Chenliang Li, Ji Zhang, Fei Huang, et al. 2024. Socialbench: Sociality evaluation of role-playing conversational agents. arXiv preprint arXiv:2403.13679 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.376,
+ 0.822,
+ 0.4
+ ],
+ "angle": 0,
+ "content": "[19] Jiangjie Chen, Xintao Wang, Rui Xu, Siyu Yuan, Yikai Zhang, Wei Shi, Jian Xie, Shuang Li, Ruihan Yang, Tinghui Zhu, et al. 2024. From persona to personalization: A survey on role-playing language agents. arXiv preprint arXiv:2404.18231 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.401,
+ 0.822,
+ 0.425
+ ],
+ "angle": 0,
+ "content": "[20] Ruizhe Chen, Xiaotian Zhang, Meng Luo, Wenhao Chai, and Zuozhu Liu. 2024. Pad: Personalized alignment of llms at decoding-time. arXiv preprint arXiv:2410.04070 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.426,
+ 0.822,
+ 0.45
+ ],
+ "angle": 0,
+ "content": "[21] Weijie Chen, Ting Bai, Jinbo Su, Jian Luan, Wei Liu, and Chuan Shi. 2024. Kg-retriever: Efficient knowledge indexing for retrieval-augmented large language models. arXiv preprint arXiv:2412.05547 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.451,
+ 0.822,
+ 0.475
+ ],
+ "angle": 0,
+ "content": "[22] Zheng Chen, Ziyan Jiang, Fan Yang, Eunah Cho, Xing Fan, Xiaojiang Huang, Yanbin Lu, and Aram Galstyan. 2023. Graph meets LLM: A novel approach to collaborative filtering for robust conversational understanding. arXiv preprint arXiv:2305.14449 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.476,
+ 0.822,
+ 0.5
+ ],
+ "angle": 0,
+ "content": "[23] Hao Cheng, Shuo Wang, Wensheng Lu, Wei Zhang, Mingyang Zhou, Kezhong Lu, and Hao Liao. 2023. Explainable recommendation with personalized review retrieval and aspect learning. arXiv preprint arXiv:2306.12657 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.501,
+ 0.822,
+ 0.525
+ ],
+ "angle": 0,
+ "content": "[24] Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. 2023. Adapting language models to compress contexts. arXiv preprint arXiv:2305.14788 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.527,
+ 0.822,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "[25] Eunah Cho, Ziyan Jiang, Jie Hao, Zheng Chen, Saurabh Gupta, Xing Fan, and Chenlei Guo. 2021. Personalized search-based query rewrite system for conversational ai. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI. 179-188."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.552,
+ 0.822,
+ 0.576
+ ],
+ "angle": 0,
+ "content": "[26]Hyundong Cho, Karishma Sharma, Nicolaas Jedema, Leonardo FR Ribeiro, Alessandro Moschitti, Ravi Krishnan, and Jonathan May. 2025. TuningFree Personalized Alignment via Trial-Error-Explain In-Context Learning. arXiv preprint arXiv:2502.08972 (2025)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.577,
+ 0.822,
+ 0.601
+ ],
+ "angle": 0,
+ "content": "[27] Yanqi Dai, Huanran Hu, Lei Wang, Shengjie Jin, Xu Chen, and Zhiwu Lu. 2024. Mmrole: A comprehensive framework for developing and evaluating multimodal role-playing agents. arXiv preprint arXiv:2408.04203 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.602,
+ 0.822,
+ 0.626
+ ],
+ "angle": 0,
+ "content": "[28] Wenlong Deng, Christos Thrampoulidis, and Xiaoxiao Li. 2024. Unlocking the potential of prompt-tuning in bridging generalized and personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6087-6097."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.627,
+ 0.822,
+ 0.651
+ ],
+ "angle": 0,
+ "content": "[29] Yang Deng, Yaliang Li, Wenxuan Zhang, Bolin Ding, and Wai Lam. 2022. Toward personalized answer generation in e-commerce via multiperspective preference modeling. ACM Transactions on Information Systems (TOIS) 40, 4 (2022), 1-28."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.652,
+ 0.822,
+ 0.676
+ ],
+ "angle": 0,
+ "content": "[30] Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The Faiss library. (2024). arXiv:2401.08281 [cs.LG]"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.677,
+ 0.433,
+ 0.689
+ ],
+ "angle": 0,
+ "content": "[31] ESPN. [n.d.]. ESPN Sports Statistics Dataset. Online dataset."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.69,
+ 0.822,
+ 0.726
+ ],
+ "angle": 0,
+ "content": "[32] 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": [
+ 0.124,
+ 0.728,
+ 0.822,
+ 0.751
+ ],
+ "angle": 0,
+ "content": "[33] Byron J Gao, David C Anastasiu, and Xing Jiang. 2010. Utilizing user-input contextual terms for query disambiguation. In Coling 2010: Posters. 329-337."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.753,
+ 0.822,
+ 0.777
+ ],
+ "angle": 0,
+ "content": "[34] Jingsheng Gao, Yixin Lian, Ziyi Zhou, Yuzhuo Fu, and Baoyuan Wang. 2023. LiveChat: A large-scale personalized dialogue dataset automatically constructed from live streaming. arXiv preprint arXiv:2306.08401 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.778,
+ 0.822,
+ 0.802
+ ],
+ "angle": 0,
+ "content": "[35] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. 2023. Retrievalaugmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 2 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.803,
+ 0.411,
+ 0.815
+ ],
+ "angle": 0,
+ "content": "[36] Google. [n.d.]. Google Search. https://www.google.com"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.124,
+ 0.124,
+ 0.823,
+ 0.815
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.463,
+ 0.832,
+ 0.476,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "19"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.881,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.124,
+ 0.882,
+ 0.16
+ ],
+ "angle": 0,
+ "content": "[37] Jia-Chen Gu, Hui Liu, Zhen-Hua Ling, Quan Liu, Zhigang Chen, and Xiaodan Zhu. 2021. Partner matters! an empirical study on fusing personas for personalized response selection in retrieval-based chatbots. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 565-574."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.161,
+ 0.882,
+ 0.198
+ ],
+ "angle": 0,
+ "content": "[38] Jie Hao, Yang Liu, Xing Fan, Saurabh Gupta, Saleh Soltan, Rakesh Chada, Pradeep Natarajan, Chenlei Guo, and Gokhan Tur. 2022. CGF: Constrained generation framework for query rewriting in conversational AI. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track. 475-483."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.199,
+ 0.882,
+ 0.223
+ ],
+ "angle": 0,
+ "content": "[39] Nicola Henze, Peter Dolog, and Wolfgang Nejdl. 2004. Reasoning and ontologies for personalized e-learning in the semantic web. Journal of Educational Technology & Society 7, 4 (2004), 82-97."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.224,
+ 0.882,
+ 0.249
+ ],
+ "angle": 0,
+ "content": "[40] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. 2023. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352 3, 4 (2023), 6."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.25,
+ 0.882,
+ 0.286
+ ],
+ "angle": 0,
+ "content": "[41] WANG Hongru, Minda Hu, Yang Deng, Rui Wang, Fei Mi, Weichao Wang, Yasheng Wang, Wai-Chung Kwan, Irwin King, and Kam-Fai Wong. [n. d]. Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues. In The 2023 Conference on Empirical Methods in Natural Language Processing."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.287,
+ 0.882,
+ 0.312
+ ],
+ "angle": 0,
+ "content": "[42] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR 1, 2 (2022), 3."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.313,
+ 0.882,
+ 0.337
+ ],
+ "angle": 0,
+ "content": "[43] Qiushi Huang, Shuai Fu, Xubo Liu, Wenwu Wang, Tom Ko, Yu Zhang, and Lilian Tang. 2024. Learning retrieval augmentation for personalized dialogue generation. arXiv preprint arXiv:2406.18847 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.338,
+ 0.882,
+ 0.362
+ ],
+ "angle": 0,
+ "content": "[44] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. 2022. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning. PMLR, 9118-9147."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.363,
+ 0.882,
+ 0.387
+ ],
+ "angle": 0,
+ "content": "[45] Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of LLM agents: A survey. arXiv preprint arXiv:2402.02716 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.388,
+ 0.882,
+ 0.412
+ ],
+ "angle": 0,
+ "content": "[46] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. 2023. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.413,
+ 0.882,
+ 0.449
+ ],
+ "angle": 0,
+ "content": "[47] Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, and Prithviraj Ammanabrolu. 2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv preprint arXiv:2310.11564 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.45,
+ 0.882,
+ 0.488
+ ],
+ "angle": 0,
+ "content": "[48] Pengyue Jia, Yiding Liu, Xiangyu Zhao, Xiaopeng Li, Changying Hao, Shuaiqiang Wang, and Dawei Yin. 2024. MILL: Mutual Verification with Large Language Models for Zero-Shot Query Expansion. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2498-2518."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.489,
+ 0.882,
+ 0.513
+ ],
+ "angle": 0,
+ "content": "[49] Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Zhang, and Yixin Zhu. 2023. Evaluating and inducing personality in pre-trained language models. Advances in Neural Information Processing Systems 36 (2023), 10622-10643."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.514,
+ 0.882,
+ 0.55
+ ],
+ "angle": 0,
+ "content": "[50] Hideaki Joko, Shubham Chatterjee, Andrew Ramsay, Arjen P De Vries, Jeff Dalton, and Faegheh Hasibi. 2024. Doing personal laps: Llm-augmented dialogue construction for personalized multi-session conversational search. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 796-806."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.551,
+ 0.882,
+ 0.576
+ ],
+ "angle": 0,
+ "content": "[51] Wang-Cheng Kang, Jianmo Ni, Nikhil Mehta, Maheswaran Sathiamoorthy, Lichan Hong, Ed Chi, and Derek Zhiyuan Cheng. 2023. Do llms understand user preferences? evaluating llms on user rating prediction. arXiv preprint arXiv:2305.06474 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.577,
+ 0.882,
+ 0.601
+ ],
+ "angle": 0,
+ "content": "[52] Manojkumar Rangasamy Kannadasan and Grigor Aslanyan. 2019. Personalized query auto-completion through a lightweight representation of the user context. arXiv preprint arXiv:1905.01386 (2019)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.602,
+ 0.882,
+ 0.651
+ ],
+ "angle": 0,
+ "content": "[53] Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Greg Corrado, Laszlo Lukacs, Marina Ganea, Peter Young, and Vivek Ramavajjala. 2016. Smart Reply: Automated Response Suggestion for Email. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (San Francisco, California, USA) (KDD '16). Association for Computing Machinery, New York, NY, USA, 955-964. https://doi.org/10.1145/2939672.2939801"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.652,
+ 0.882,
+ 0.676
+ ],
+ "angle": 0,
+ "content": "[54] Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. 2024. Reinforcement learning for optimizing rag for domain chatbots. arXiv preprint arXiv:2401.06800 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.677,
+ 0.882,
+ 0.702
+ ],
+ "angle": 0,
+ "content": "[55] Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, et al. 2024. Longlamp: A benchmark for personalized long-form text generation. arXiv preprint arXiv:2407.11016 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.703,
+ 0.882,
+ 0.727
+ ],
+ "angle": 0,
+ "content": "[56] Brenden Lake and Marco Baroni. 2018. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning. PMLR, 2873-2882."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.728,
+ 0.882,
+ 0.752
+ ],
+ "angle": 0,
+ "content": "[57] Cheng Li, Mingyang Zhang, Qiao zhu Mei, Weize Kong, and Michael Bendersky. 2024. Learning to rewrite prompts for personalized text generation. In Proceedings of the ACM Web Conference 2024. 3367-3378."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.753,
+ 0.882,
+ 0.777
+ ],
+ "angle": 0,
+ "content": "[58] Changhao Li, Yuchen Zhuang, Rushi Qiang, Haotian Sun, Hanjun Dai, Chao Zhang, and Bo Dai. 2024. Matryoshka: Learning to Drive Black-Box LLMs with LLMs. arXiv preprint arXiv:2410.20749 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.778,
+ 0.882,
+ 0.802
+ ],
+ "angle": 0,
+ "content": "[59] Lei Li, Yongfeng Zhang, and Li Chen. 2023. Personalized prompt learning for explainable recommendation. ACM Transactions on Information Systems 41, 4 (2023), 1-26."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.803,
+ 0.882,
+ 0.828
+ ],
+ "angle": 0,
+ "content": "[60] Sen Li, Fuyu Lv, Taiwei Jin, Guiyang Li, Yukun Zheng, Tao Zhuang, Qingwen Liu, Xiaoyi Zeng, James Kwok, and Qianli Ma. 2022. Query rewriting in taobao search. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3262-3271."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.183,
+ 0.124,
+ 0.882,
+ 0.828
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.523,
+ 0.832,
+ 0.535,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "20"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.117,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.123,
+ 0.124,
+ 0.822,
+ 0.148
+ ],
+ "angle": 0,
+ "content": "[61] Xiaopeng Li, Lixin Su, Pengyue Jia, Xiangyu Zhao, Suqi Cheng, Junfeng Wang, and Dawei Yin. 2023. Agent4ranking: Semantic robust ranking via personalized query rewriting using multi-agent llm. arXiv preprint arXiv:2312.15450 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.123,
+ 0.149,
+ 0.822,
+ 0.173
+ ],
+ "angle": 0,
+ "content": "[62] Xinyu Li, Ruiyang Zhou, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.175,
+ 0.821,
+ 0.199
+ ],
+ "angle": 0,
+ "content": "[63] Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, et al. 2024. Personal llm agents: Insights and survey about the capability, efficiency and security. arXiv preprint arXiv:2401.05459 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.2,
+ 0.821,
+ 0.224
+ ],
+ "angle": 0,
+ "content": "[64] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.225,
+ 0.822,
+ 0.26
+ ],
+ "angle": 0,
+ "content": "[65] Ruixue Lian, Sixing Lu, Clint Solomon, Gustavo Aguilar, Pragaash Ponnusamy, Jialong Han, Chengyuan Ma, and Chenlei Guo. 2023. PersonalTM: Transformer memory for personalized retrieval. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2256-2260."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.262,
+ 0.822,
+ 0.286
+ ],
+ "angle": 0,
+ "content": "[66] Shan-Mu Lin and Chuen-Min Huang. 2006. Personalized optimal search in local query expansion. In Proceedings of the 18th Conference on Computational Linguistics and Speech Processing. 221-236."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.288,
+ 0.822,
+ 0.311
+ ],
+ "angle": 0,
+ "content": "[67] Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is chatgpt a good recommender? a preliminary study. arXiv preprint arXiv:2304.10149 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.313,
+ 0.822,
+ 0.337
+ ],
+ "angle": 0,
+ "content": "[68] Jiahong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Jieming Zhu, Minda Hu, Menglin Yang, and Irwin King. 2025. A Survey of Personalized Large Language Models: Progress and Future Directions. arXiv preprint arXiv:2502.11528 (2025)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.338,
+ 0.822,
+ 0.362
+ ],
+ "angle": 0,
+ "content": "[69] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics 12 (2024), 157-173."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.363,
+ 0.822,
+ 0.387
+ ],
+ "angle": 0,
+ "content": "[70] Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024. Once: Boosting content-based recommendation with both open-and closed-source large language models. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining. 452-461."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.388,
+ 0.822,
+ 0.412
+ ],
+ "angle": 0,
+ "content": "[71] Shuai Liu, Hyundong J Cho, Marjorie Freedman, Xuezhe Ma, and Jonathan May. 2023. RECAP: retrieval-enhanced context-aware prefix encoder for personalized dialogue response generation. arXiv preprint arXiv:2306.07206 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.413,
+ 0.783,
+ 0.425
+ ],
+ "angle": 0,
+ "content": "[72] Tyler Lu and Craig Boutilier. 2011. Budgeted social choice: From consensus to personalized decision making. In *IJCAI*, Vol. 11, 280-286."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.427,
+ 0.822,
+ 0.45
+ ],
+ "angle": 0,
+ "content": "[73] Zhengyi Ma, Zhicheng Dou, Yutao Zhu, Hanxun Zhong, and Ji-Rong Wen. 2021. One chatbot per person: Creating personalized chatbots based on implicit user profiles. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 555-564."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.451,
+ 0.822,
+ 0.475
+ ],
+ "angle": 0,
+ "content": "[74] Aman Madaan, Niket Tandon, Peter Clark, and Yiming Yang. 2022. Memory-assisted prompt editing to improve GPT-3 after deployment. arXiv preprint arXiv:2201.06009 (2022)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.476,
+ 0.822,
+ 0.5
+ ],
+ "angle": 0,
+ "content": "[75] Shengyu Mao, Xiaohan Wang, Mengru Wang, Yong Jiang, Pengjun Xie, Fei Huang, and Ningyu Zhang. 2024. Editing Personality for Large Language Models. In CCF International Conference on Natural Language Processing and Chinese Computing. Springer, 241-254."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.502,
+ 0.822,
+ 0.525
+ ],
+ "angle": 0,
+ "content": "[76] Yuren Mao, Xuemei Dong, Wenyi Xu, Yunjun Gao, Bin Wei, and Ying Zhang. 2024. Fit-rag: black-box rag with factual information and token reduction. arXiv preprint arXiv:2403.14374 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.527,
+ 0.822,
+ 0.563
+ ],
+ "angle": 0,
+ "content": "[77] Puneet Mathur, Zhe Liu, Ke Li, Yingyi Ma, Gil Keren, Zeeshan Ahmed, Dinesh Manocha, and Xuedong Zhang. 2023. Personal: Language model personalization via domain-distributed span aggregated k-nearest n-gram retrieval augmentation. In Findings of the Association for Computational Linguistics: EMNLP 2023. 11314-11328."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.565,
+ 0.822,
+ 0.588
+ ],
+ "angle": 0,
+ "content": "[78] Fatemehsadat Mireshghallah, Vaishnavi Shrivastava, Milad Shokouhi, Taylor Berg-Kirkpatrick, Robert Sim, and Dimitrios Dimitriadis. 2021. Identifier: Implicit user representations for simple and effective personalized sentiment analysis. arXiv preprint arXiv:2110.00135 (2021)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.59,
+ 0.822,
+ 0.613
+ ],
+ "angle": 0,
+ "content": "[79] Philippe Mulhem, Nawal Ould Amer, and Mathias Gery. 2016. Axiomatic term-based personalized query expansion using bookmarking system. In International Conference on Database and Expert Systems Applications. Springer, 235-243."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.615,
+ 0.822,
+ 0.65
+ ],
+ "angle": 0,
+ "content": "[80] Sheshera Mysore, Zhuoran Lu, Mengting Wan, Longqi Yang, Steve Menezes, Tina Baghaee, Emmanuel Barajas Gonzalez, Jennifer Neville, and Tara Safavi. 2023. Pearl: Personalizing large language model writing assistants with generation-calibrated retrievers. arXiv preprint arXiv:2311.09180 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.653,
+ 0.822,
+ 0.676
+ ],
+ "angle": 0,
+ "content": "[81] Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, et al. 2024. Gui agents: A survey. arXiv preprint arXiv:2412.13501 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.677,
+ 0.822,
+ 0.701
+ ],
+ "angle": 0,
+ "content": "[82] Duy A Nguyen, Rishi Kesav Mohan, Van Yang, Pritom Saha Akash, and Kevin Chen-Chuan Chang. 2025. RL-based Query Rewriting with Distilled LLM for online E-Commerce Systems. arXiv preprint arXiv:2501.18056 (2025)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.703,
+ 0.822,
+ 0.739
+ ],
+ "angle": 0,
+ "content": "[83] Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 188-197."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.741,
+ 0.822,
+ 0.765
+ ],
+ "angle": 0,
+ "content": "[84] Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O'Banion, and Jun Xie. 2024. User-llm: Efficient llm contextualization with user embeddings. arXiv preprint arXiv:2402.13598 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.766,
+ 0.822,
+ 0.789
+ ],
+ "angle": 0,
+ "content": "[85] Douglas Oard, William Webber, David Kirsch, and Sergey Golitsynski. 2015. Avocado research email collection. Philadelphia: Linguistic Data Consortium (2015)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.79,
+ 0.74,
+ 0.802
+ ],
+ "angle": 0,
+ "content": "[86] U.S. National Library of Medicine. [n.d.]. PubMed: A Free Resource for Biomedical Literature. https://pubmed.ncbi.nlm.nih.gov/"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.124,
+ 0.804,
+ 0.822,
+ 0.828
+ ],
+ "angle": 0,
+ "content": "[87] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology. 1-22."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.123,
+ 0.124,
+ 0.822,
+ 0.828
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.463,
+ 0.832,
+ 0.475,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "21"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.881,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.184,
+ 0.124,
+ 0.882,
+ 0.148
+ ],
+ "angle": 0,
+ "content": "[88] Greg Pass, Abdur Chowdhury, and Cayley Torgeson. 2006. A picture of search. In Proceedings of the 1st International Conference on Scalable Information Systems (Hong Kong) (InfoScale '06). Association for Computing Machinery, New York, NY, USA, 1-es. https://doi.org/10.1145/1146847.1146848"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.184,
+ 0.149,
+ 0.882,
+ 0.174
+ ],
+ "angle": 0,
+ "content": "[89] Vadim Igorevich Pavliukevich, Alina Khasanovna Zherdeva, Olesya Vladimirovna Makhnytkina, and Dmitriy Viktorovich Dyrmovskiy. [n. d.]. Improving RAG with LoRA finetuning for persona text generation. ([n. d.])."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.184,
+ 0.174,
+ 0.878,
+ 0.186
+ ],
+ "angle": 0,
+ "content": "[90] Dan Peng, Zhihui Fu, and Jun Wang. 2024. Pocketllm: Enabling on-device fine-tuning for personalized llms. arXiv preprint arXiv:2407.01031 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.187,
+ 0.881,
+ 0.211
+ ],
+ "angle": 0,
+ "content": "[91] Qiyao Peng, Hongtao Liu, Hongyan Xu, Qing Yang, Minglai Shao, and Wenjun Wang. 2024. Review-LLM: Harnessing Large Language Models for Personalized Review Generation. arXiv:2407.07487 [cs.CL] https://arxiv.org/abs/2407.07487"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.212,
+ 0.881,
+ 0.237
+ ],
+ "angle": 0,
+ "content": "[92] Hongjin Qian, Zhicheng Dou, Yutao Zhu, Yueyuan Ma, and Ji-Rong Wen. 2021. Learning implicit user profile for personalized retrieval-based chatbot. In proceedings of the 30th ACM international conference on Information & Knowledge Management. 1467-1477."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.238,
+ 0.881,
+ 0.273
+ ],
+ "angle": 0,
+ "content": "[93] Hongjin Qian, Xiahe Li, Hanxun Zhong, Yu Guo, Yueyuan Ma, Yutao Zhu, Zhanliang Liu, Zhicheng Dou, and Ji-Rong Wen. 2021. Pchatbot: a large-scale dataset for personalized chatbot. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 2470-2477."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.274,
+ 0.881,
+ 0.299
+ ],
+ "angle": 0,
+ "content": "[94] Xiaoru Qu, Yifan Wang, Zhao Li, and Jun Gao. 2024. Graph-enhanced prompt learning for personalized review generation. Data Science and Engineering 9, 3 (2024), 309-324."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.3,
+ 0.881,
+ 0.312
+ ],
+ "angle": 0,
+ "content": "[95] A. Rajaraman and J.D. Ullman. 2011. Mining of Massive Datasets. Cambridge University Press. https://books.google.co.uk/books?id=OefRhZyYOb0C"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.313,
+ 0.881,
+ 0.337
+ ],
+ "angle": 0,
+ "content": "[96] Yiting Ran, Xintao Wang, Rui Xu, Xinfeng Yuan, Jiaqing Liang, Deqing Yang, and Yanghua Xiao. 2024. Capturing minds, not just words: Enhancing role-playing language models with personality-indicative data. arXiv preprint arXiv:2406.18921 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.338,
+ 0.881,
+ 0.362
+ ],
+ "angle": 0,
+ "content": "[97] Chandan K. Reddy, Lluis Marquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopadhyay, Arnab Biswas, Anlu Xing, and Karthik Subbian. 2022. Shopping Queries Dataset: A Large-Scale ESCI Benchmark for Improving Product Search. (2022). arXiv:2206.06588"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.363,
+ 0.878,
+ 0.375
+ ],
+ "angle": 0,
+ "content": "[98] Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084 (2019)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.183,
+ 0.376,
+ 0.881,
+ 0.4
+ ],
+ "angle": 0,
+ "content": "[99] Ruiyang Ren, Peng Qiu, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Hua Wu, Ji-Rong Wen, and Haifeng Wang. 2024. Bases: Large-scale web search user simulation with large language model based agents. arXiv preprint arXiv:2402.17505 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.401,
+ 0.881,
+ 0.424
+ ],
+ "angle": 0,
+ "content": "[100] Matthew Renze and Erhan Guven. 2024. Self-reflection in llm agents: Effects on problem-solving performance. arXiv preprint arXiv:2405.06682 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.426,
+ 0.881,
+ 0.45
+ ],
+ "angle": 0,
+ "content": "[101] 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": [
+ 0.18,
+ 0.451,
+ 0.881,
+ 0.475
+ ],
+ "angle": 0,
+ "content": "[102] Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval 3, 4 (2009), 333-389."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.476,
+ 0.881,
+ 0.5
+ ],
+ "angle": 0,
+ "content": "[103] 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": [
+ 0.18,
+ 0.501,
+ 0.881,
+ 0.526
+ ],
+ "angle": 0,
+ "content": "[104] Alireza Salemi, Cheng Li, Mingyang Zhang, Qiao zhu Mei, Weize Kong, Tao Chen, Zhuowan Li, Michael Bendersky, and Hamed Zamani. 2025. Reasoning-Enhanced Self-Training for Long-Form Personalized Text Generation. arXiv preprint arXiv:2501.04167 (2025)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.527,
+ 0.881,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "[105] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 7370-7392."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.552,
+ 0.881,
+ 0.575
+ ],
+ "angle": 0,
+ "content": "[106] Alireza Salemi and Hamed Zamani. 2024. Learning to Rank for Multiple Retrieval-Augmented Models through Iterative Utility Maximization. arXiv preprint arXiv:2410.09942 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.577,
+ 0.881,
+ 0.601
+ ],
+ "angle": 0,
+ "content": "[107] Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect?. In International Conference on Machine Learning. PMLR, 29971-30004."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.602,
+ 0.881,
+ 0.626
+ ],
+ "angle": 0,
+ "content": "[108] Rossano Schifanella, Alain Barrat, Ciro Cattuto, Benjamin Markines, and Filippo Menczer. 2010. Folks in folksonomies: social link prediction from shared metadata. In Proceedings of the third ACM international conference on Web search and data mining. 271-280."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.627,
+ 0.881,
+ 0.651
+ ],
+ "angle": 0,
+ "content": "[109] Noor Shaker, Georgios Yannakakis, and Julian Togelius. 2010. Towards automatic personalized content generation for platform games. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, Vol. 6. 63-68."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.652,
+ 0.878,
+ 0.664
+ ],
+ "angle": 0,
+ "content": "[110] Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2023. Character-llm: A trainable agent for role-playing. arXiv preprint arXiv:2310.10158 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.665,
+ 0.881,
+ 0.689
+ ],
+ "angle": 0,
+ "content": "[111] Jocelyn Shen, Joel Mire, Hae Won Park, Cynthia Breazeal, and Maarten Sap. 2024. HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs. arXiv preprint arXiv:2405.17633 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.69,
+ 0.881,
+ 0.714
+ ],
+ "angle": 0,
+ "content": "[112] Yunxiao Shi, Xing Zi, Zijing Shi, Haimin Zhang, Qiang Wu, and Min Xu. 2024. Eragent: Enhancing retrieval-augmented language models with improved accuracy, efficiency, and personalization. arXiv preprint arXiv:2405.06683 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.715,
+ 0.881,
+ 0.739
+ ],
+ "angle": 0,
+ "content": "[113] Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. 2025. Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG. arXiv preprint arXiv:2501.09136 (2025)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.741,
+ 0.881,
+ 0.776
+ ],
+ "angle": 0,
+ "content": "[114] Harmanpreet Singh, Nikhil Verma, Yixiao Wang, Manasa Bharadwaj, Homa Fashandi, Kevin Ferreira, and Chul Lee. 2024. Personal Large Language Model Agents: A Case Study on Tailored Travel Planning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 486-514."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.778,
+ 0.881,
+ 0.814
+ ],
+ "angle": 0,
+ "content": "[115] Shamane Siriwardhana, Rivindu Weerasekera, Elliott Wen, Tharindu Kaluarachchi, Rajib Rana, and Suranga Nanayakkara. 2023. Improving the domain adaptation of retrieval augmented generation (RAG) models for open domain question answering. Transactions of the Association for Computational Linguistics 11 (2023), 1-17."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.18,
+ 0.816,
+ 0.854,
+ 0.828
+ ],
+ "angle": 0,
+ "content": "[116] Mingyang Song and Mao Zheng. 2024. A Survey of Query Optimization in Large Language Models. arXiv preprint arXiv:2412.17558 (2024)."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.18,
+ 0.124,
+ 0.882,
+ 0.828
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.523,
+ 0.832,
+ 0.536,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "22"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.116,
+ 0.091,
+ 0.349,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.124,
+ 0.644,
+ 0.136
+ ],
+ "angle": 0,
+ "content": "[117] Spotify. 2023. Annoy: Approximate Nearest Neighbors in C++/Python. https://github.com/spotify/annoy"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.137,
+ 0.821,
+ 0.16
+ ],
+ "angle": 0,
+ "content": "[118] Stuck_In_the Matrix. 2015. Reddit Public Comments (2007-10 through 2015-05). (2015). https://www.reddit.com/r/datasets/comments/3bxlg7/i_have EVERY_publicly-available Reddit_COMMENT/"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.162,
+ 0.821,
+ 0.185
+ ],
+ "angle": 0,
+ "content": "[119] Lei Sun, Jinming Zhao, and Qin Jin. 2024. Revealing Personality Traits: A New Benchmark Dataset for Explanable Personality Recognition on Dialogues. arXiv preprint arXiv:2409.19723 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.187,
+ 0.821,
+ 0.211
+ ],
+ "angle": 0,
+ "content": "[120] 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": [
+ 0.118,
+ 0.213,
+ 0.821,
+ 0.236
+ ],
+ "angle": 0,
+ "content": "[121] 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 preprint arXiv:2402.04401 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.237,
+ 0.821,
+ 0.261
+ ],
+ "angle": 0,
+ "content": "[122] Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2025. 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": [
+ 0.118,
+ 0.262,
+ 0.821,
+ 0.286
+ ],
+ "angle": 0,
+ "content": "[123] Quan Tu, Shilong Fan, Zihang Tian, and Rui Yan. 2024. Charactereval: A Chinese benchmark for role-playing conversational agent evaluation. arXiv preprint arXiv:2401.01275 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.287,
+ 0.574,
+ 0.299
+ ],
+ "angle": 0,
+ "content": "[124] Cornell University. [n.d.]. arXiv: An Open Access Repository for Research. https://arxiv.org/"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.3,
+ 0.821,
+ 0.336
+ ],
+ "angle": 0,
+ "content": "[125] Hemanth Vemuri, Sheshansh Agrawal, Shivam Mittal, Deepak Saini, Akshay Soni, Abhinav V Sambasivan, Wenhao Lu, Yajun Wang, Mehul Parsana, Purushottam Kar, et al. 2023. Personalized retrieval over millions of items. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1014-1022."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.337,
+ 0.821,
+ 0.362
+ ],
+ "angle": 0,
+ "content": "[126] Bryan Wang, Gang Li, and Yang Li. 2023. Enabling conversational interaction with mobile ui using large language models. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1-17."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.363,
+ 0.821,
+ 0.387
+ ],
+ "angle": 0,
+ "content": "[127] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.388,
+ 0.821,
+ 0.412
+ ],
+ "angle": 0,
+ "content": "[128] Hongru Wang, Wenyu Huang, Yang Deng, Rui Wang, Zezhong Wang, Yufei Wang, Fei Mi, Jeff Z Pan, and Kam-Fai Wong. 2024. Unims-rag: A unified multi-source retrieval-augmented generation for personalized dialogue systems. arXiv preprint arXiv:2401.13256 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.413,
+ 0.821,
+ 0.437
+ ],
+ "angle": 0,
+ "content": "[129] Hongru Wang, Rui Wang, Fei Mi, Yang Deng, Zezhong Wang, Bin Liang, Ruifeng Xu, and Kam-Fai Wong. 2023. Cue-CoT: Chain-of-thought prompting for responding to in-depth dialogue questions with LLMs. arXiv preprint arXiv:2305.11792 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.438,
+ 0.821,
+ 0.462
+ ],
+ "angle": 0,
+ "content": "[130] Jian Wang, Yi Cheng, Dongding Lin, Chak Tou Leong, and Wenjie Li. 2023. Target-oriented proactive dialogue systems with personalization: Problem formulation and dataset curation. arXiv preprint arXiv:2310.07397 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.463,
+ 0.821,
+ 0.487
+ ],
+ "angle": 0,
+ "content": "[131] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2024. A survey on large language model based autonomous agents. Frontiers of Computer Science 18, 6 (2024), 186345."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.488,
+ 0.811,
+ 0.5
+ ],
+ "angle": 0,
+ "content": "[132] Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query expansion with large language models. arXiv preprint arXiv:2303.07678 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.501,
+ 0.821,
+ 0.525
+ ],
+ "angle": 0,
+ "content": "[133] Lei Wang, Jingsen Zhang, Hao Yang, Zhiyuan Chen, Jiakai Tang, Zeyu Zhang, Xu Chen, Yankai Lin, Ruihua Song, Wayne Xin Zhao, et al. 2023. User behavior simulation with large language model based agents. arXiv preprint arXiv:2306.02552 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.527,
+ 0.821,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "[134] Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, et al. 2023. Incharacter: Evaluating personality fidelity in role-playing agents through psychological interviews. arXiv preprint arXiv:2310.17976 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.552,
+ 0.821,
+ 0.576
+ ],
+ "angle": 0,
+ "content": "[135] Yixiao Wang, Homa Fashandi, and Kevin Ferreira. 2024. Investigating the Personality Consistency in Quantized Role-Playing Dialogue Agents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 239–255."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.577,
+ 0.821,
+ 0.601
+ ],
+ "angle": 0,
+ "content": "[136] Yu Wang, Yifan Gao, Xiusi Chen, Haoming Jiang, Shiyang Li, Jingfeng Yang, Qingyu Yin, Zheng Li, Xian Li, Bing Yin, et al. [n.d.]. MEMORYLLM: Towards Self-Updatable Large Language Models. In Forty-first International Conference on Machine Learning."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.602,
+ 0.821,
+ 0.626
+ ],
+ "angle": 0,
+ "content": "[137] Zheng Wang, Zhongyang Li, Zeren Jiang, Dandan Tu, and Wei Shi. 2024. Crafting Personalized Agents through Retrieval-Augmented Generation on Editable Memory Graphs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 4891-4906."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.627,
+ 0.821,
+ 0.651
+ ],
+ "angle": 0,
+ "content": "[138] Zijie J Wang and Duen Horng Chau. 2024. MeMemo: On-device Retrieval Augmentation for Private and Personalized Text Generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2765-2770."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.652,
+ 0.821,
+ 0.688
+ ],
+ "angle": 0,
+ "content": "[139] Zekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, et al. 2023. Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. arXiv preprint arXiv:2310.00746 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.69,
+ 0.821,
+ 0.726
+ ],
+ "angle": 0,
+ "content": "[140] Tianxin Wei, Bowen Jin, Ruirui Li, Hansi Zeng, Zhengyang Wang, Jianhui Sun, Qingyu Yin, Hanqing Lu, Suhang Wang, Jingrui He, et al. 2024. Towards unified multi-modal personalization: Large vision-language models for generative recommendation and beyond. arXiv preprint arXiv:2403.10667 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.728,
+ 0.821,
+ 0.752
+ ],
+ "angle": 0,
+ "content": "[141] Robert Wetzker, Carsten Zimmermann, and Christian Bauchage. 2008. Analyzing social bookmarking systems: A del. icio. us cookbook. In Proceedings of the ECAI 2008 Mining Social Data Workshop. 26-30."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.753,
+ 0.821,
+ 0.777
+ ],
+ "angle": 0,
+ "content": "[142] Stanisław Wozniak, Bartlomiej Koptyra, Arkadiusz Janz, Przemysław Kazienko, and Jan Kocón. 2024. Personalized large language models. arXiv preprint arXiv:2402.09269 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.778,
+ 0.821,
+ 0.802
+ ],
+ "angle": 0,
+ "content": "[143] Junde Wu, Jiayuan Zhu, Yunli Qi, Jingkun Chen, Min Xu, Filippo Menolascina, and Vicente Grau. 2024. Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.803,
+ 0.821,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "[144] Xuan Wu, Dong Zhou, Yu Xu, and Seamus Lawless. 2017. Personalized query expansion utilizing multi-relational social data. In 2017 12th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP). IEEE, 65-70."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.118,
+ 0.124,
+ 0.821,
+ 0.827
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.463,
+ 0.832,
+ 0.475,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "23"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.177,
+ 0.091,
+ 0.46,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.781,
+ 0.091,
+ 0.881,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "X. Li and P. Jia, et al."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.124,
+ 0.882,
+ 0.16
+ ],
+ "angle": 0,
+ "content": "[145] Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender Systems. 12-22."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.177,
+ 0.161,
+ 0.882,
+ 0.186
+ ],
+ "angle": 0,
+ "content": "[146] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences 68, 2 (2025), 121101."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.187,
+ 0.881,
+ 0.211
+ ],
+ "angle": 0,
+ "content": "[147] Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. 641-649."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.212,
+ 0.881,
+ 0.237
+ ],
+ "angle": 0,
+ "content": "[148] Huatao Xu, Liying Han, Qirui Yang, Mo Li, and Mani Srivastava. 2024. Penetrative ai: Making llms comprehend the physical world. In Proceedings of the 25th International Workshop on Mobile Computing Systems and Applications. 1-7."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.238,
+ 0.881,
+ 0.261
+ ],
+ "angle": 0,
+ "content": "[149] Hongyan Xu, Hongtao Liu, Pengfei Jiao, and Wenjun Wang. 2021. Transformer reasoning network for personalized review summarization. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1452-1461."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.262,
+ 0.881,
+ 0.286
+ ],
+ "angle": 0,
+ "content": "[150] Xinchao Xu, Zhibin Gou, Wenquan Wu, Zheng-Yu Niu, Hua Wu, Haifeng Wang, and Shihang Wang. 2022. Long time no see! open-domain conversation with long-term persona memory. arXiv preprint arXiv:2203.05797 (2022)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.287,
+ 0.881,
+ 0.312
+ ],
+ "angle": 0,
+ "content": "[151] Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, and Tat-Seng Chua. 2025. Personalized Generation In Large Model Era: A Survey. arXiv preprint arXiv:2503.02614 (2025)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.313,
+ 0.881,
+ 0.337
+ ],
+ "angle": 0,
+ "content": "[152] Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. 2024. Personalized federated continual learning via multi-granularity prompt. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4023-4034."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.338,
+ 0.881,
+ 0.362
+ ],
+ "angle": 0,
+ "content": "[153] Xiaoyan Yu, Tongxu Luo, Yifan Wei, Fangyu Lei, Yiming Huang, Hao Peng, and Liehuang Zhu. 2024. Neeko: Leveraging dynamic lora for efficient multi-character role-playing agent. arXiv preprint arXiv:2402.13717 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.363,
+ 0.881,
+ 0.387
+ ],
+ "angle": 0,
+ "content": "[154] Xinfeng Yuan, Siyu Yuan, Yuhan Cui, Tianhe Lin, Xintao Wang, Rui Xu, Jiangjie Chen, and Deqing Yang. 2024. Evaluating character understanding of large language models via character profiling from fictional works. arXiv preprint arXiv:2404.12726 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.388,
+ 0.881,
+ 0.412
+ ],
+ "angle": 0,
+ "content": "[155] Hansi Zeng, Surya Kallumadi, Zaid Alibadi, Rodrigo Nogueira, and Hamed Zamani. 2023. A personalized dense retrieval framework for unified information access. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 121-130."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.413,
+ 0.881,
+ 0.437
+ ],
+ "angle": 0,
+ "content": "[156] Saber Zerhoudi and Michael Granitzer. 2024. PersonaRAG: Enhancing Retrieval-Augmented Generation Systems with User-Centric Agents. arXiv preprint arXiv:2407.09394 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.438,
+ 0.881,
+ 0.475
+ ],
+ "angle": 0,
+ "content": "[157] Han Zhang, Songlin Wang, Kang Zhang, Zhiling Tang, Yunjiang Jiang, Yun Xiao, Weipeng Yan, and Wen-Yun Yang. 2020. Towards personalized and semantic retrieval: An end-to-end solution for e-commerce search via embedding learning. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2407-2416."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.476,
+ 0.773,
+ 0.488
+ ],
+ "angle": 0,
+ "content": "[158] Jiarui Zhang. 2024. Guided profile generation improves personalization with llms. arXiv preprint arXiv:2409.13093 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.489,
+ 0.881,
+ 0.513
+ ],
+ "angle": 0,
+ "content": "[159] Jesse Zhang, Jiahui Zhang, Karl Pertsch, Ziyi Liu, Xiang Ren, Minsuk Chang, Shao-Hua Sun, and Joseph J Lim. [n.d.]. Bootstrap Your Own Skills: Learning to Solve New Tasks with Large Language Model Guidance. In 7th Annual Conference on Robot Learning."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.514,
+ 0.881,
+ 0.538
+ ],
+ "angle": 0,
+ "content": "[160] Kai Zhang, Yangyang Kang, Fubang Zhao, and Xiaozhong Liu. 2023. LLM-based medical assistant personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.539,
+ 0.881,
+ 0.563
+ ],
+ "angle": 0,
+ "content": "[161] Kaiyan Zhang, Jianyu Wang, Ermo Hua, Biqing Qi, Ning Ding, and Bowen Zhou. 2024. Cogenesis: A framework collaborating large and small language models for secure context-aware instruction following. arXiv preprint arXiv:2403.03129 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.565,
+ 0.881,
+ 0.588
+ ],
+ "angle": 0,
+ "content": "[162] Kai Zhang, Fubang Zhao, Yangyang Kang, and Xiaozhong Liu. 2023. Memory-augmented llm personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.589,
+ 0.881,
+ 0.626
+ ],
+ "angle": 0,
+ "content": "[163] Wenlin Zhang, Chuhan Wu, Xiangyang Li, Yuhao Wang, Kuicai Dong, Yichao Wang, Xinyi Dai, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2025. LLMTreeRec: Unleashing the Power of Large Language Models for Cold-Start Recommendations. In Proceedings of the 31st International Conference on Computational Linguistics. 886-896."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.627,
+ 0.881,
+ 0.651
+ ],
+ "angle": 0,
+ "content": "[164] Yanyue Zhang, Yulan He, and Deyu Zhou. 2025. Rehearse With User: Personalized Opinion Summarization via Role-Playing based on Large Language Models. arXiv preprint arXiv:2503.00449 (2025)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.652,
+ 0.881,
+ 0.676
+ ],
+ "angle": 0,
+ "content": "[165] You Zhang, Jin Wang, Liang-Chih Yu, Dan Xu, and Xuejie Zhang. 2024. Personalized LoRA for human-centered text understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 1958-19596."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.677,
+ 0.881,
+ 0.702
+ ],
+ "angle": 0,
+ "content": "[166] Yabin Zhang, Wenhui Yu, Erhan Zhang, Xu Chen, Lantao Hu, Peng Jiang, and Kun Gai. 2024. Recgpt: Generative personalized prompts for sequential recommendation via chatgpt training paradigm. arXiv preprint arXiv:2404.08675 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.703,
+ 0.881,
+ 0.727
+ ],
+ "angle": 0,
+ "content": "[167] Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.728,
+ 0.881,
+ 0.752
+ ],
+ "angle": 0,
+ "content": "[168] Zhehao Zhang, Ryan A Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, et al. 2024. Personalization of large language models: A survey. arXiv preprint arXiv:2411.00027 (2024)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.753,
+ 0.881,
+ 0.777
+ ],
+ "angle": 0,
+ "content": "[169] Yi Zheng, Chongyang Ma, Kanle Shi, and Haibin Huang. 2023. Agents meet okr: An object and key results driven agent system with hierarchical self-collaboration and self-evaluation. arXiv preprint arXiv:2311.16542 (2023)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.778,
+ 0.881,
+ 0.802
+ ],
+ "angle": 0,
+ "content": "[170] Hanxun Zhong, Zhicheng Dou, Yutao Zhu, Hongjin Qian, and Ji-Rong Wen. 2022. Less is more: Learning to refine dialogue history for personalized dialogue generation. arXiv preprint arXiv:2204.08128 (2022)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.178,
+ 0.803,
+ 0.881,
+ 0.827
+ ],
+ "angle": 0,
+ "content": "[171] Wanjun Zhong, Duyu Tang, Jiahai Wang, Jian Yin, and Nan Duan. 2021. UserAdapter: Few-shot user learning in sentiment analysis. In Findings of the Association for Computational Linguistics: ACL-JJCNLP 2021. 1484-1488."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.177,
+ 0.124,
+ 0.882,
+ 0.827
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.523,
+ 0.832,
+ 0.536,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "24"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.117,
+ 0.091,
+ 0.348,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "A Survey of Personalization: From RAG to Agent"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.54,
+ 0.091,
+ 0.822,
+ 0.102
+ ],
+ "angle": 0,
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.124,
+ 0.822,
+ 0.147
+ ],
+ "angle": 0,
+ "content": "[172] Dong Zhou, Séamus Lawless, and Vincent Wade. 2012. Improving search via personalized query expansion using social media. Information retrieval 15 (2012), 218-242."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.149,
+ 0.822,
+ 0.173
+ ],
+ "angle": 0,
+ "content": "[173] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. 2022. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022)."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.175,
+ 0.821,
+ 0.198
+ ],
+ "angle": 0,
+ "content": "[174] Yujia Zhou, Qiannan Zhu, Jiajie Jin, and Zhicheng Dou. 2024. Cognitive personalized search integrating large language models with an efficient memory mechanism. In Proceedings of the ACM Web Conference 2024. 1464-1473."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.118,
+ 0.2,
+ 0.821,
+ 0.224
+ ],
+ "angle": 0,
+ "content": "[175] Yuchen Zhuang, Haotian Sun, Yue Yu, Rushi Qiang, Qifan Wang, Chao Zhang, and Bo Dai. [n.d.]. Hydra: Model factorization framework for black-box llm personalization, 2024. URL https://arxiv.org/abs/2406.02888 ([n.d.])."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.118,
+ 0.124,
+ 0.822,
+ 0.224
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.117,
+ 0.239,
+ 0.504,
+ 0.251
+ ],
+ "angle": 0,
+ "content": "Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.463,
+ 0.832,
+ 0.475,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "25"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10147/26499afc-f1b2-4507-8b62-4adc6ac17e5f_origin.pdf b/data/2025/2504_10xxx/2504.10147/26499afc-f1b2-4507-8b62-4adc6ac17e5f_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..5d39baa179e1cdb0494c8964c6dceadb221fd009
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/26499afc-f1b2-4507-8b62-4adc6ac17e5f_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97329609b909a06225fc6f112666706d10e5b3ceb1d8907638249f0b3a618c65
+size 2003619
diff --git a/data/2025/2504_10xxx/2504.10147/full.md b/data/2025/2504_10xxx/2504.10147/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..949c90d427bd9450e1cd62d1363fd267071298b2
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/full.md
@@ -0,0 +1,460 @@
+# A Survey of Personalization: From RAG to Agent
+
+XIAOPENG LI*, City University of Hong Kong, Hong Kong
+
+PENGYUE JIA*, City University of Hong Kong, Hong Kong
+
+DERONG XU, City University of Hong Kong, Hong Kong and University of Science and Technology of China, China
+
+YI WEN, City University of Hong Kong, Hong Kong
+
+YINGYI ZHANG, City University of Hong Kong, Hong Kong and Dalian University of Technology, China
+
+WENLIN ZHANG, City University of Hong Kong, Hong Kong
+
+WANYU WANG, City University of Hong Kong, Hong Kong
+
+YICHAO WANG, Noah's Ark Lab, Huawei, China
+
+ZHAOCHENG DU, Noah's Ark Lab, Huawei, China
+
+XIANGYANG LI, Noah's Ark Lab, Huawei, China
+
+YONG LIU, Noah's Ark Lab, Huawei, Singapore
+
+HUIFENG GUO, Noah's Ark Lab, Huawei, China
+
+RUIMING TANG†, Noah's Ark Lab, Huawei, China
+
+XIANGYU ZHAO†, City University of Hong Kong, Hong Kong
+
+Personalization has become an essential capability in modern AI systems, enabling customized interactions that align with individual user preferences, contexts, and goals. Recent research has increasingly concentrated on Retrieval-Augmented Generation (RAG) frameworks and their evolution into more advanced agent-based architectures within personalized settings to enhance user satisfaction. Building on this foundation, this survey systematically examines personalization across the three core stages of RAG: pre-retrieval, retrieval, and generation. Beyond RAG, we further extend its capabilities into the realm of Personalized LLM-based Agents, which enhance traditional RAG systems with agentic functionalities, including user understanding, personalized planning and execution, and dynamic generation. For both personalization in RAG and agent-based personalization, we provide formal definitions, conduct a comprehensive review of recent literature, and summarize key datasets and evaluation metrics. Additionally, we discuss fundamental challenges, limitations, and promising research directions in this evolving field. Relevant papers and resources are continuously updated at the Github Repo| Field | Sub-field | Subsub-field | Papers |
| Pre-retrieval | Query Rewriting | Learning to Personalized Query Rewrite | CLE-QR [60], CGF [38], PEARL [80] |
| LLM to Personalized Query Rewrite | Least-to-Most Prompting [173], ERAGent [112], CoPS [174], Agent4Ranking [61], FIG [22], BASES [99] |
| Query Expansion | Tagging-based query expansion | Gossiple [10], Biancalana and Micarelli [12], SoQuES [15], Zhou et al. [172] |
| Else | Lin and Huang [66], Bender et al. [9], Axiomatic PQEC [79], WE-LM [144], PSQE [14], PQEWC [7] |
| Others | Bobo [33], Kannadasan and Aslanyan [52], PSQE [8] |
| Retrieval | Indexing | PEARL [80], KG-Retriever [21], EMG-RAG [137], PGraphRAG [5] |
| Retrieval | Dense Retrieval | MeMemo [138], RECAP [71], LAPDOG [43], Gu et al. [37], PersonalLM [77], UIA [155], XPERT [125], DPSR [157], RTM [11], Pearl [80], MemPrompt [74], EERRA [23], MALP [160], USER-LLM [84], PER-PCS [120] |
| Sparse Retrieval | OPPU [121], PAG [101], Au et al. [5], UniMS-RAG [128], Deng et al. [29], |
| Prompt-based Retrieval | LAPS [50], UniIMP [140], Shen et al. [111] |
| Others | Salem et al. [103], PersonalITM [65], Zhang et al. [165] |
| Post-retrieval | PersonaRAG [156], Pavliukevich et al. [89], UniMS-RAG [128], Salemi and Zamani [106], Zhang et al. [164], AutoCompressors [24], FIT-RAG [76] |
| Generation | Generation from Explicit Preferences | Direct Prompting | P² [49], Character Profiling [154] OpinionQA [107], Kang et al. [51], Liu et al. [67], Cue-CoT [129], TICL [26] |
| Profile-Augmented Prompting | GPG [158], Richardson et al. [101], ONCE [70], LLMTreeRec [163], KAR [145], Matryoshka [58] |
| Personalized-Prompt Prompting | Li et al. [57], RecGPT [166], PEPLER-D [59], GRAPA [94], SGPT [28], PFCL [152] |
| Generation from Implicit Preferences | Fine-tuning-Based Methods | PLoRA [165], LM-P [142], MiLP [165], OPPU [122], PER-PCS [120], Review-LLM [91], UserIdentifier [78], UserAdapter [171], HYDRA [175], PocketLLM [90], CoGenesis [161] |
| Reinforcement Learning-Based Methods | P-RLHF [62], P-SOUPS [47], PAD [20], REST-PG [104], Salemi et al. [103], RewrimerSIRI [57],Kulkarni et al. [54] |
| From RAG to Agent | Personalized Understanding | In user-profile understanding | Xu et al. [148], Abbasian et al. [2], |
| In agent's role understanding | RoleLLM [139], Character-LLM [110], Wang et al. [134], |
| In agent's user-role joint understanding | SocialBench [18], Dai et al. [27], Ran et al. [96], Wang et al. [126], Tu et al. [123], Necko [153] |
| Personalized Planning and Execution | Memory Management | EMG-RAG [137], Park et al. [87], Abbasian et al. [2], RecAgent [133], TravelPlanner+ [114], PersonalWAB [17], VOYAGER [127], MemoeryLLM [136] |
| Tool and API Calling | VOYAGER [127], Zhang et al. [159], PUMA [17], Wang et al. [126], PenetrativeAI [148], Huang et al. [44], [87], MetaGPT [40], OKR-Agent [169] |
| Personalized Generation | Alignment with User Fact | Character-LLM [110], Wang et al. [135], Dai et al. [27] |
| Alignment with User Preferences | Wang et al. [139], Ran et al. [96], Wang et al. [134], Chen et al. [18] |
+
+- User Historical Content: Implicit personalization derived from user-generated content, such as chat history, emails, reviews, and social media interactions.
+- Persona-Based User Simulation: The use of LLM-based agents to simulate and generate personalized interactions.
+
+Integrating this personalized information at various stages of the RAG and agent workflows enables dynamic alignment with human preferences, thereby making responses more user-centric and adaptive.
+
+# 3 HOW TO ADOPT PERSONALIZATION
+
+We define the process of introducing personalization within the RAG pipeline as follows:
+
+$$
+g = \mathcal {G} (\mathcal {R} (Q (q, p), C, p), \text {p r o m p t}, p, \theta) \tag {1}
+$$
+
+where $p$ denotes personalized information, and the process unfolds in three steps. In the pre-retrieval phase, query processing $(Q)$ refines the query $q$ using personalized information, such as through query rewriting or expansion. During the retrieval phase, the retriever $(\mathcal{R})$ leverages $p$ to fetch relevant documents from the corpus $(C)$ . Finally, in the generation phase, the retrieved information, combined with $p$ and structured using the given prompt, id fed into the generator $(\mathcal{G})$ with parameter $\theta$ to produce the final response $g$ . It is evident that personalized information directly influences multiple stages of the RAG pipeline. In this survey, we consider the agent system as a specialized application of the RAG framework, where personalization is incorporated in a manner similar to the RAG framework.
+
+
+Fig. 2. Overview of the personalized pre-retrieval stage.
+
+# 4 WHERE TO ADOPT PERSONALIZATION
+
+# 4.1 Pre-retrieval
+
+4.1.1 Definition. Pre-retrieval is a crucial step in information retrieval systems, where the original user query is enhanced or modified before the retrieval process to improve the relevance and quality of the search results, as shown in Figure 2. This process often incorporates additional contextual or personalized information to better align the query with the user's intent. The process can be formalized as follows:
+
+$$
+q ^ {*} = Q (q, p) \tag {2}
+$$
+
+where $p$ and $q$ denote the personalized information and original query, and $q^{*}$ is the optimized query after query reformulation.
+
+4.1.2 Query Rewriting. Query rewriting in RAG at the pre-retrieval stage refers to the process of reformulating user queries to enhance retrieval effectiveness by improving relevance, disambiguating intent, or incorporating contextual information before retrieving documents from an external knowledge source. The literature on personalized query rewriting can be broadly classified into two primary categories: (1) Direct Personalized Query Rewriting and (2) Auxiliary Personalized Query Rewriting.
+
+(1). Direct Personalized Query Rewriting. The first category focuses on personalized query rewriting by using direct models. For example, Cho et al. [25] presents a personalized search-based query rewrite system for conversational AI that addresses user-specific semantic and phonetic errors. Nguyen et al. [82] apply reinforcement learning techniques to improve query rewriting in online e-commerce systems, leveraging distilled LLMs for personalized performance. CLE-QR [60] explores query rewriting in Taobao's search engine to enhance user satisfaction through customized query adaptation. CGF [38] introduces a constrained generation framework that allows for more flexible and personalized query rewriting in conversational AI. Li et al. [57] investigate learning methods to rewrite prompts for personalized
+
+text generation, improving the relevance and engagement of AI-generated content. Additionally, PEARL [80] discusses personalizing large language model-based writing assistants through the integration of generation-calibrated retrievers, enhancing AI-generated content.
+
+(2). Auxiliary Personalized Query Rewriting. The second category emphasizes personalized query rewriting by using auxiliary mechanisms, such as retrieval, reasoning strategies, and external memory. Zhou et al. [173] propose a least-to-most prompting strategy that aids in complex reasoning within LLMs, which can be adapted for personalized text generation. ERAGent [112] enhance retrieval-augmented LLMs to improve personalization, efficiency, and accuracy, indirectly supporting personalized query rewriting for content generation. CoPS [174] integrate LLMs with memory mechanisms to create more personalized search experiences, which also influences content generation through better query understanding. Further, Agent4Ranking [61] employs multi-agent LLMs to perform semantic robust ranking, including personalized query rewriting to improve search rankings. FIG [22] combine graph-based methods with LLMs to query rewrite, improving personalized content generation and conversational interactions. Lastly, BASES [99] employ LLM-based agents to simulate large-scale web search user interactions, contributing to the development of personalized query rewriting strategies for content generation.
+
+4.1.3 Query Expansion. Query expansion enhances retrieval systems by expanding a user's original query with additional terms, synonyms, or refined structure to better capture intent. This improves the relevance and scope of retrieved documents. Recent advancements in LLMs have reinvigorated this field [46, 48, 132], leveraging their comprehension and generation abilities to expand queries using encoded knowledge or external retrieval, with notable success. Personalized query expansion, a subset, incorporates user-specific data to tailor results, boosting performance and customizing the search experience.
+
+(1). Tagging-based Query Expansion. By 2009, studies began incorporating tagging information to enhance personalized query expansion. For instance, Gossle [10] introduced the TagMap and TagRank algorithms, which dynamically selected tags from personalized networks constructed using the cosine similarity of user-item tag distances, improving recall performance. Similarly, Biancalana and Micarelli [12] recorded user queries and visited URLs, leveraging social bookmarking to extract relevant tags and build a personalized three-dimensional co-occurrence matrix. Based on this, multiple semantically categorized expanded queries were generated to better reflect user interests. Further advancements include SoQuES [15], which integrated tag semantic similarity with social proximity, and a graph-based approach [172] that utilized Tag-Topic models and pseudo-relevance feedback for term weighting, tailoring the expansion process to individual user preferences.
+(2). Else. Apart from tagging-based techniques, early research on Personalized Query Expansion primarily focused on modeling user personalization based on search history [66], social networks, or preferences derived from friendship networks [9]. The Axiomatic PQEC framework [79] formalized expansion rules using both local (user behavior-driven) and social (network-driven) strategies. In 2017, WE-LM [144] advanced this paradigm by modeling multi-relational networks with word embeddings across tag-word relationships, refining associations through affinity graphs. Later, PSQE [14] further improved tagging-based methods using utf-iuf user profiling, integrating a tag similarity graph with user profiles in the online phase to compute expansion terms relevant to user interests in real-time, achieving dynamic personalized expansion. In addition, PQEWC [7] leveraged clustering and contextual word embeddings to optimize query expansions dynamically.
+
+
+Fig. 3. Overview of the personalized retrieval stage.
+
+4.1.4 Others. Besides query rewriting and query expansion, other personalized query-related research focuses on areas like query disambiguation and query auto-completion [116]. Bobo [33] allows users to input contextual terms reflecting their domain knowledge. In 2019, a method [52] applied fastText embeddings from recent queries to rank candidates. In addition, PSQE [8] employed synthetic user profiles from Wikipedia and word2vec embeddings for query disambiguation.
+
+4.1.5 Discussion. While both query rewriting and query expansion aim to align user input with system understanding to enhance retrieval quality, their roles in personalization differ in fundamental ways. Understanding the distinct operational characteristics and application scenarios of each technique is essential for designing effective personalized retrieval systems. The key takeaways are listed as follows:
+
+- Query rewriting is most beneficial when the original query is ambiguous, underspecified, or misaligned with retrieval intents, particularly in conversational or multi-turn settings.
+- Query expansion is most effective when the original query is relevant but incomplete - i.e., when it needs to be semantically broadened to cover additional relevant concepts.
+
+# 4.2 Retrieval
+
+4.2.1 Definition. The retrieval process involves finding the most relevant documents $D^{*}$ from a corpus $C$ based on a query $q^{*}$ , as shown in Figure 3. To incorporate personalization, additional user-specific information $p$ is integrated into the retrieval function $\mathcal{R}$ . This allows the retrieval process to tailor the selected documents to align with individual user preferences or contexts, thereby enhancing the relevance and personalization of the generated outputs.
+
+$$
+D ^ {*} = \mathcal {R} (q ^ {*}, C, p) \tag {3}
+$$
+
+In the retrieval process, personalization can primarily be introduced by focusing on three steps: indexing, retrieval, and post-retrieval. These steps ensure efficient and accurate retrieval of relevant documents or knowledge, while tailoring the process to individual user preferences. Below, we provide a detailed explanation of each step.
+
+4.2.2 Indexing. Indexing organizes knowledge base data into a structured format to facilitate efficient retrieval. Within the RAG pipeline, documents are either chunked or entirely encoded into representations before being integrated into searchable systems [30, 117]. Conventional encoding methods employ either sparse encoding techniques (e.g.,
+
+TF-IDF [95], BM25 [102]) or dense encoding approaches leveraging pre-trained models, such as BERT [1], Siamese Encoders [98], or LLM-based encoders [64, 147].
+
+To introduce personalization at the indexing stage, PEARL [80] generates user embeddings by encoding personal history data with models like DeBERTa. These embeddings are subsequently clustered to create personalized shared indices. Other approaches integrate knowledge graphs into indexing to enhance retrieval performance. For example, KG-R retriever [21] employs a Hierarchical Index Graph, consisting of a knowledge graph layer and a collaborative document layer, to improve RAG retrieval. EMG-RAG [137] incorporates personalized memory within an editable knowledge graph, enabling dynamic retrieval. Similarly, PGraphRAG [5] leverages user-centric knowledge graphs to enhance personalization in retrieval tasks.
+
+4.2.3 Retrieval. The Retrieval step matches a user query with the indexed knowledge base to fetch relevant candidates. It can be broadly categorized into four different types: (1) Dense Retrieval, (2) Sparse Retrieval, (3) Prompt-based Retrieval, and (4) Others.
+
+(1). Dense Retrieval. Dense retrieval methods often use vector embeddings and similarity metrics (e.g., cosine similarity) and achieve personalization by encoding user preferences, context, or interactions into query or document embeddings, enabling tailored results through similarity-based matching. For instance, MeMemo [138] retrieves personalized information by matching user-specific embeddings with document vectors, focusing on private, on-device text generation. Similarly, RECAP [71] and LAPDOG [43] enhance personalized dialogue generation by encoding queries and user profiles as dense vectors and retrieving top-N results, ensuring user-specific context drives the responses. In chatbots, Gu et al. [37] integrates conversational context and user profiles to align retrieved responses with user personas. PersonalM [77] employs group-wise contrastive learning, training its retrieval model to align user queries with domain-specific text fragments, thereby improving personalization. UIA [155] employs dual encoders to retrieve documents tailored to user preferences. XPERT [125] incorporates temporal events and user interactions into embeddings, enabling large-scale retrieval across millions of items.
+
+Dense retrieval also enhances specific applications like e-commerce, medical assistance, and language models. DPSR [157] and RTM [11] encode user queries and product information to personalize product searches dynamically. Pearl [80] and MemPrompt [74] retrieve personalized content by leveraging historical user data and memory-assisted mechanisms. EERRA [23] uses review embeddings as dense queries for recommendations. In medical assistance, MALP [160] and User-LLM [84] integrate short- and long-term user interactions into embeddings for contextualized, personalized responses. Finally, PER-PCS [120] retrieves relevant information using individual user histories, enhancing the personalization capabilities of large language models.
+
+(2). Sparse Retrieval. Sparse retrieval methods often rely on term-based matching (e.g., BM25) and apply personalization by assigning higher weights to terms or keywords that are more relevant to the user. OPPU [121] uses the BM25 algorithm to select the k most relevant records from the user's historical data for the current query. Similarly, PAG [101] incorporates user input and profiles to enhance summarization and retrieval, aligning sparse representations with personalization objectives for large language models. Au et al. [5] uses BM25 search algorithms to find entries related to the target user or neighboring users through the graph structure. UniMS-RAG [128] combines sparse and dense retrieval by leveraging multi-source knowledge, such as dialogue context and user images, to refine personalized responses in dialogue systems. Lastly, Deng et al. [29] apply sparse retrieval to support fact-based queries, considering user queries and preferences to enhance answer generation for e-commerce applications.
+
+(3). Prompt-based Retrieval. Prompt-based retrieval leverages prompts to guide retrieval from the model or external sources and introduces personalization by crafting user-specific prompts that guide the retrieval process. These prompts may include explicit user preferences, historical interactions, or detailed instructions that reflect the user's unique requirements. By embedding this personalized context directly into the prompt, the retrieval process can dynamically adjust to capture and return results that are most relevant to the user. LAPS [50] focuses on multi-session conversational search by storing user preferences and dialogue context, then using prompts to retrieve relevant information tailored to the user's biases and categories of interest. UniMP [140] employs user interaction histories as input to prompt-based retrieval, enabling personalized recommendations for multi-modal tasks, such as vision-language applications, by aligning prompts with user behavioral data. In contrast, Shen et al. [111] explores the use of LLMs to extract empathy and narrative styles from user-provided stories, but this work primarily focuses on style extraction and does not explicitly involve a retrieval component.
+(4). Others. Reinforcement learning-based retrieval personalizes the process by optimizing retrieval policies based on user feedback, learning user preferences over time to adjust strategies. Salemi et al. [103] combines models like BM25, RbR, and dense retrieval, refining them with reinforcement learning (RL) and knowledge distillation (KD) to adapt to user profiles for personalized outputs. Parameter-based retrieval leverages pre-trained model parameters to implicitly store and retrieve user-specific information, allowing direct retrieval from the model without traditional indices. PersonalTM [65] generates document identifiers (Document IDs) using a Transformer model, encoding query, history, and document relationships into its parameters for personalization. Similarly, Zhang et al. [165] uses parameterized representations to integrate user queries and histories, tailoring responses to individual preferences.
+
+4.2.4 Post-retrieval. Current Post-Retrieval methods primarily focus on refining retrieved documents or responses to improve relevance and coherence, current methodologies could be categorized into three parts (1) Re-ranking, (2) Summarization, and (3) Compression.
+
+(1). Re-ranking. Re-ranking enhances personalized content generation by prioritizing more relevant documents at the top. PersonaRAG [156] extends RAG by integrating user-centric agents, such as the Live Session Agent and the Document Ranking Agent, to refine document ranking and improve overall performance. Pavliukevich et al. [89] propose a cross-encoder BERT model for re-ranking external knowledge within a personalized context. UniMS-RAG [128] introduces a scoring mechanism that evaluates retrieved documents and outputs by optimizing the retriever. Besides, it includes an evidence attention mask, enabling re-ranking during inference and applying it to personalized datasets. Salemi and Zamani [106] present an iterative approach to optimizing ranking results based on the expectation-maximization algorithm, with performance validated in personalized scenarios.
+(2). Summarization. Summarization refers to the process of summarizing retrieved information to enhance performance. For instance, Zhang et al. [164] introduced a role-playing agent system to summarize retrieved history in order to improve the final Personalized Opinion Summarization process.
+(3). Compression. Compression involves condensing embeddings or retrieved content to enhance efficiency and effectiveness. Approaches like AutoCompressor [24] compress contextual embeddings into shorter semantic representations, and FIT-RAG [76] introduces a self-knowledge recognizer along with a sub-document-level token reduction mechanism to minimize tokens within RAG pipeline. However, few studies have specifically explored personalized fields, highlighting a promising direction for future research.
+
+4.2.5 Discussion. Indexing, retrieval, and post-retrieval methods each play a critical role in ensuring efficient and personalized information processing, with specific applications and trade-offs. Indexing focuses on organizing knowledge bases for efficient retrieval, using techniques such as sparse encoding methods like TF-IDF and BM25, which are efficient but limited in understanding semantics, and dense encoding methods like BERT and DeBERTa, which provide better semantic understanding but require significant computational resources. These methods are widely used in tasks like question answering and personalized recommendation systems. Retrieval involves matching user queries with relevant documents and can be categorized into dense retrieval, which provides high semantic understanding and personalization but is computationally expensive; sparse retrieval, which is efficient and interpretable but less capable of handling semantics; prompt-based retrieval, which is highly flexible and adaptable to user needs but requires careful engineering of prompts; and advanced methods like reinforcement learning-based approaches, which dynamically adapt to user feedback but are complex to implement. This step is essential in applications like personalized dialogue systems, search engines, and e-commerce. Post-retrieval methods refine retrieved results to enhance relevance and coherence through re-ranking, which improves personalization and prioritizes relevant content but increases computational overhead; summarization, which simplifies complex information for better user understanding but risks losing critical details; and compression, which reduces computational costs by condensing information but remains underexplored in personalized contexts. Together, these methods provide a comprehensive pipeline for delivering efficient, relevant, and personalized outputs, balancing their strengths in semantic understanding, relevance, and flexibility with challenges related to computational costs and implementation complexity.
+
+# 4.3 Generation
+
+4.3.1 Definition. Personalized generation incorporates user-specific retrieved documents $D^{*}$ , task-specific prompt prompt, and user preference information $p$ via the generator $\mathcal{G}$ parameterized by $\theta$ to produce tailored content $g^{*}$ aligned with individual preference, where the flow is shown in Figure 4. The generation process can be formulated as
+
+$$
+g ^ {*} = \mathcal {G} (D ^ {*}, \text {p r o m p t}, p, \theta). \tag {4}
+$$
+
+Personalized generation can be achieved by incorporating explicit and implicit preferences. Explicit preference-driven methodologies utilize direct input signals (e.g., $D^{*}$ , prompt, and $p$ ), to tailor outputs to specific user preferences. Conversely, implicit preference-encoded approaches embed personalized information within the parameters $\theta$ of the generator model, during training, thereby facilitating preference alignment without the necessity for explicit runtime inputs.
+
+4.3.2 Generation from Explicit Preferences. Integrating explicit preferences into LLMs facilitates personalized content generation. Explicit preference information encompasses user demographic information (e.g., age, occupation, gender, location), user behavior sequences (reflecting historical behavioral patterns), and user historical output texts (capturing writing style and tone preferences). The injection of explicit preferences for personalized generation can be categorized into three types: (1) Direct-integrated Prompting, (2) Summary-augmented Prompting, and (3) Adaptive Prompting.
+
+(1). Direct-integrated Prompting. Integrating user explicit preferences into language models through prompting enables the prediction of users' intent and behavioral patterns, facilitating personalized content generation. For instance, $\mathrm{P}^2$ [49], Character Profiling [154], and OpinionQA [107] integrate personalized data into LLMs through prompting for role-playing task, thereby aligning the model's responses with specified user profiles. Kang et al. [51] and Liu et al. [67]
+
+
+Fig. 4. Overview of the personalized generation stage.
+
+integrate interaction histories into LLMs via prompting to predict user rating for candidate items. Cue-CoT [129] employs chain-of-thought reasoning to infer user needs from contextual cues, enabling personalized responses to in-depth dialogue questions. Additionally, TICL [26] proposes a trial-and-error framework that critiques initial LLM-generated responses, derives explanations and integrates these negative examples into prompts to improve personalization alignment.
+
+(2). Summary-augmented Prompting. Direct integration of personalized information via prompting struggles with ambiguous intent signals: Lengthy interaction histories introduce noise that obscures critical behavioral patterns [69], while sparse behavioral data lacks sufficient context for LLMs to derive meaningful user preferences. To address these issues, recent approaches focus on summarizing user personalized intents and integrating them into prompts. For instance, GPG [158] extracts key user habits and preferences from personal contexts, enabling fine-grained personalization. Similarly, LLMs are employed to generate task-specific summaries of user preferences, enhancing retrieval-augmented personalized generation capabilities [101]. In recommendation systems, ONCE [70], LLMTreeRec [163], and KAR [145] leverage historical user-item interactions to summarize user preferences. Furthermore, Matryoshka [58] generates user preference summaries by dynamically retrieving and synthesizing historical data.
+(3). Adaptive Prompting. Manually designing personalized prompts demands both expert knowledge and significant labor, motivating the development of automated methods for personalized prompt generation. For example, Li et al. [57] trains a personalized prompt rewriter via supervised and reinforcement learning. RecGPT [166] and PEPLER-D [59] leverage prompt tuning to generate personalized prompts, enhancing sequential and explainable recommendations, respectively. GRAPA [94] integrates semantic and collaborative signals from user-item interaction graphs with graph neural networks to generate context-aware personalized prompts. SGPT [28] employs prompt tuning to jointly model common and group-specific patterns, bridging generalized and personalized federated learning paradigms. Furthermore, PFCL [152] achieves multi-granularity human preference modeling: coarse-grained prompts distill shared knowledge, while fine-grained prompts adapt to individual user characteristics.
+
+4.3.3 Generation from Implicit Preferences. Unlike explicit preference modeling, which captures user preferences through textual input, implicit preference-based methods incorporate personalization through internal parameters. This personalization is achieved either through Parameter-Efficient Fine-tuning (PEFT) techniques, such as LoRA [42],
+
+or reinforcement learning-based approaches for preference alignment [20, 57]. Based on these strategies, we classify existing methods into two categories: (1) Fine-tuning-Based Methods and (2) Reinforcement Learning-Based Methods.
+
+(1). Fine-tuning Based Methods. For fine-tuning methods, LoRA is the most widely adopted since it is resource-efficient and enables rapid adaptation without compromising model performance. PLoRA [165] introduces a personalized knowledge integration framework that combines task-specific LoRA with user-specific knowledge. Similarly, LM-P [142] personalizes information via LoRA by incorporating User ID as a personalization factor. MiLP [165] employs Bayesian optimization to determine the optimal personalization injection configuration, including LoRA settings, to effectively capture and utilize user-specific information. OPPU [122] and PER-PCS [120] follow a similar approach, leveraging user history data for fine-tuning LoRA-based personalization. However, PER-PCS differs by incorporating a gating module that selects the appropriate LoRA, enabling fine-grained personalization. Additionally, Review-LLM [91] integrates LoRA for supervised fine-tuning in the task of personalized review generation.
+
+Beyond LoRA-based approaches, alternative pipelines have been proposed for personalized generation. UserIdentifier [78] introduces a user-specific identifier, significantly reducing training costs while enhancing personalized demonstration. UserAdapter [171] proposes user-independent prefix embeddings, leveraging prefix tuning for personalization. Meanwhile, HYDRA [175] achieves implicit personalization by training user-specific headers. Recently, researchers have also explored fine-tuning personalized model on edge devices [90] and collaborative learning between small and large language models to enable more personalized generation [161].
+
+(2). Reinforcement Learning Based Methods. Apart from fine-tuning based methods, recent research has explored reinforcement learning based techniques to personalize text generation by aligning outputs with user preferences. P-RLHF [62] has been proposed to jointly learn a user-specific and reward model to enable text generation that aligns with a user's styles or criteria. P-SOUPS [47] models multiple user preferences as a Multi-Objective Reinforcement Learning (MORL) problem, decomposing preferences into multiple dimensions, each trained independently. PAD [20] aligns text generation with human preferences during inference by utilizing token-level personalized rewards to guide the decoding process. REST-PG [104] introduces a framework that trains large language models to reason over personal data during response generation. This approach first generates reasoning paths to enhance the LLM's reasoning ability and then employs Expectation-Maximization Reinforced Self-Training to iteratively refine the model based on its high-reward outputs. Additionally, Salemi et al. [103] incorporate reinforcement learning into the RAG pipeline to improve retrieval accuracy, thereby enhancing the personalization of generated content. Other applications include RewriterSIRI [57], which has been introduced to generate text via RL-based personalized prompt rewriting using API-based LLMs, and Kulkarni et al. [54], who explore the use of reinforcement learning to optimize RAG for improving the relevance and coherence of chatbot responses in specialized domains, ultimately enhancing user satisfaction and engagement.
+
+4.3.4 Discussion. Personalized generation can be adopted via both explicit and implicit preference injection, yet they exhibit distinct characteristics that make them suitable for different scenarios. In explicit preference-based generation, personalization is clearly defined through user profile descriptions, contextual information, and similar inputs, which are incorporated into generators via prompts. A key advantage of this approach is explainability, as the personalized information is explicitly provided and easily traceable. Despite leveraging provided preferences and internal knowledge, explicit preference injection's personalization is constrained by model capabilities and irrelevant information interference. In contrast, implicit preference-based generation internalizes personalized information into
+
+
+Fig. 5. Overview of transition from personalized RAG to personalized agent.
+
+the generator's parameters through scene-specific personalized data, thereby adapting the model for more fine-grained personalization. However, these methods typically incur substantial training and computational costs, as they require fine-tuning the generator's internal parameters. Therefore, selecting between these approaches should be guided by the specific application scenario and resource constraints.
+
+# 4.4 From RAG to Agent
+
+4.4.1 Definition. A personalized LLM-based agent is a system designed to dynamically incorporate user context, memory, and external tools or APIs to support highly personalized and goal-oriented interactions [19, 45, 146], and solve problems in a goal-oriented manner [63, 113]. From the previously introduced stages of RAG, we observe that the evolution of personalized RAG reveals a structural convergence with agent architectures. We analyze them from three key perspectives:
+
+- Personalized Understanding: This phase within the agent parallels the query understanding and rewriting process of RAG as outlined in Section 4.1. However, it extends beyond static semantic parsing by incorporating dynamic user profiling [139] and role modeling [110]. This integration enables the agent to dynamically align interactions with implicit user preferences, facilitating personalized responses and task-specific adaptations [96].
+
+- Personalized Planning and Execution: This phase in agents mirrors RAG's retrieval process in Section 4.2 yet it advances beyond static document retrieval by incorporating real-time memory management [87] and sophisticated tool and API calling [127]. This approach ensures the dynamic alignment of external knowledge with personalized constraints, such as integrating medical history in healthcare agents [2], to deliver context-aware and user-specific outcomes.
+- Personalized Generation: This phase in agents mirrors RAG's generative process in Section 4.3 but transcends static template-based generation by integrating user preference and fact alignment. Agents dynamically enforce user preferences and ensure fact consistency through role-specific mechanisms (e.g., social adaptability in conversational agents [2]), enabling outputs to evolve in harmony with personalized and situational constraints rather than relying solely on predefined generative frameworks.
+
+In general we frame agent architectures as "personalized RAG++", where persistent memory [137] replaces static indexes, and tool APIs [17] serve as dynamic knowledge connectors, enabling complicated, human-aligned interactions beyond one-shot retrieval, as shown in Figure 5. This progression highlights that as RAG systems incorporate deeper personalization—requiring user-state tracking, adaptive tool usage, and context-aware generation, they inherently adopt agent-like capabilities.
+
+4.4.2 Personalized Understanding. Personalized understanding refers to an agent's ability to accurately interpret user inputs by integrating user intent recognition and contextual analysis. This process ensures interactions that are both meaningful and contextually appropriate. The rationale behind this classification lies in its capacity to address three core aspects of understanding: recognizing user intent, analyzing context, and leveraging user profiles. Each of these aspects plays a distinct role in improving the agent's performance.
+
+(1). User-profile Understanding. In user-profile understanding, an agent's personalized ability primarily depends on its capacity to accurately model and understand the user's preferences, context, and intentions. Xu et al. [148] proposes a framework in which LLMs are designed to understand the physical world, thereby facilitating a deeper connection between the agent and its environment, which is essential for accurate task execution. Abbasian et al. [2] further expands this understanding by emphasizing the importance of personalization in health agents, where the user's profile directly influences the behavior and decisions of the agent. This user understanding is foundational to ensuring that the AI agent performs tasks in a way that aligns with individual user needs.
+(2). Role Understanding. In agent's role understanding, the role of the agent within these environments is also crucial. Recent studies focus on enhancing role-playing capabilities within LLMs. Wang et al. [139] introduce RoleLLM, a benchmark that aims to elicit and refine the role-playing abilities of LLMs, demonstrating how role understanding influences agent performance in conversational tasks. Similarly, Shao et al. [110] present Character-LLM, a trainable agent framework for role-playing, which tailors its responses based on predefined roles. Wang et al. [134] introduce a method for evaluating personality fidelity in role-playing agents through psychological interviews, aiming to enhance the realism and consistency of AI-driven characters. This role understanding allows for more contextually appropriate interactions, increasing the relevance and utility of AI agents across various applications.
+(3). User-role Joint Understanding. In agent's user-role joint understanding, the intersection of user and role understanding is explored through frameworks that evaluate and enhance the social and personality aspects of LLMs. SocialBench Chen et al. [18] provides a sociality evaluation framework for role-playing agents. Dai et al. [27], and
+
+[96] extend this by incorporating multi-modal data and personality-indicative information, respectively, which allows agents to better adapt to both user and role understanding in dynamic environments. Furthermore, Wang et al. [126] offers a perspective on how role and environment understanding can improve user experience. Tu et al. [123] contribute by providing a benchmark specifically for evaluating role-playing agents in the Chinese context, adding a cultural dimension to role understanding. Finally, Neeko [153] further advances role-based interactions.
+
+4.4.3 Personalized Planning and Execution. Personalized planning and execution refer to the process of designing and implementing strategies or actions that are specifically tailored to an individual's unique context, and goals [44, 87, 114, 159]. It requires agents to dynamically integrate long-term memory, real-time reasoning, and external tool utilization [40, 41, 169], as demonstrated in healthcare decision support [2] and travel planning scenarios [17]. We analyze two fundamental components that enable this personalization in the following.
+
+(1). Memory Management. Effective memory systems allow agents to integrate users' historical preferences, behavioral patterns, and contextual habits, enhancing their ability to make planning and tailor interactions to user-specific needs [17, 127, 136]. The EMG-RAG framework [137] combines editable memory graphs with retrieval-augmented generation to maintain dynamic user profiles, while Park et al. [87] implements memory streams and periodic reflection mechanisms to simulate human-like behavior. In healthcare applications, Abbasian et al. [2] integrates multimodal user data through specialized memory modules to optimize treatment recommendations. For recommendation systems, RecAgent [133] employs hierarchical memory structures to model user interaction patterns across multiple domains. Recent advances like TravelPlanner+ [114] demonstrate how memory-augmented LLMs achieve higher relevance in personalized itinerary generation compared to generic planners.
+(2). Tool and API Calling. The integration of external tools expands agents' capabilities beyond pure linguistic reasoning, enabling agents to interact with users and perform personalized tasks [17, 126, 127, 148, 159]. For instance, VOYAGER [127] establishes a paradigm for lifelong skill acquisition through automatic API curriculum learning and skill library construction. In robotics, Zhang et al. [159] develops a bootstrapping framework where LLMs guide robots in tool-mediated skill discovery, enabling a high success rate in novel object manipulation tasks. The PUMA framework [17] demonstrates how personalized web agents can achieve performance gains in e-commerce tasks through adaptive API orchestration. For mobile interaction, Wang et al. [126] implements few-shot tool learning to handle diverse UI operations with minimal training data. These approaches highlight the importance of tool grounding mechanisms [44] that translate linguistic plans into executable API sequences while maintaining personalization constraints.
+
+This synthesis highlights that modern agent systems achieve enhanced personalization through two primary strategies: 1) Memory-augmented architectures, which leverage editable memory graphs [137], reflection mechanisms [87], and hierarchical memory structures [133] to dynamically adapt to user preferences across various domains; and 2) Tool and API integration, which expand agent capabilities by balancing generalization with specialization. Future work may explore improving the contextual relevance and adaptability of memory systems while optimizing real-time tool interaction for seamless task execution.
+
+4.4.4 Personalized Generation. Based on the foundation of personalized planning and execution mechanisms, which enable agents to adapt strategies to user-specific contexts [44, 159], the next critical concern lies in personalized generation. This capability ensures that generated outputs not only align with factual correctness but also resonate with users' unique preferences, personality traits, and situational needs. Personalized generation bridges the gap between
+
+adaptive reasoning and human-aligned outcomes, allowing agents to produce contextually relevant and emotionally appropriate responses.
+
+(1). Alignment with User Fact. Alignment with User Fact emphasizes the accuracy, consistency, and factual grounding of personalized responses, ensuring they remain trustworthy across diverse user interactions. This is particularly challenging in personalized agents, where maintaining character authenticity while avoiding hallucinations requires balancing creativity with factual adherence. Recent advances address these challenges through improved training frameworks and evaluation metrics. For instance, Character-LLM [110] integrates memory-augmented architectures to reduce hallucinations while preserving character-specific traits. Wang et al. [135] investigate quantization effects on personality consistency in edge-deployed agents and stabilize outputs under computational constraints. Dai et al. [27] ensures multimodal consistency (text-image) in role-playing. These works highlight the importance of architectural innovations and rigorous evaluation in achieving reliability.
+(2). Alignment with User Preferences. Alignment with user preferences ensures that generated outputs reflect individualized personalities, values, and interaction styles. This requires agents to dynamically interpret implicit user cues and adapt responses accordingly. Wang et al. [139] benchmarks role-specific alignment. Ran et al. [96] improves personality fidelity via psychological scale datasets. Wang et al. [134] quantifies alignment via psychological interviews. Chen et al. [18] evaluates social adaptability in conversations.
+
+4.4.5 Discussion. The architectural evolution from RAG to personalized agents introduces significant advancements in human-AI interaction but also surfaces critical challenges that warrant further investigation.
+
+Personalized Understanding, while enabling interpretation of user intent and context, faces limitations in real-time adaptability and generalization. Current approaches like RoleLLM [139] and Character-LLM [110] demonstrate robust role-specific comprehension but struggle with dynamic user state tracking, particularly when handling evolving preferences or multi-session interactions. Furthermore, cultural specificity in benchmarks like CharacterEval [123] reveals gaps in global applicability, as agents trained on region-specific data often fail to generalize across diverse sociocultural contexts. Future work could explore hybrid architectures that combine continuous learning mechanisms with privacy-preserving federated learning to address these adaptability constraints while maintaining user trust.
+
+Personalized Planning and Execution, achieves remarkable task specialization through memory management and tool integration, yet suffers from scalability issues in complex environments. While frameworks like EMG-RAG [137] and VOYAGER [127] effectively manage user-specific constraints, their reliance on predefined API taxonomies limits emergent tool discovery in novel scenarios. The "cold-start" problem persists in domains requiring rapid skill acquisition, as seen in healthcare applications [2], where delayed API responses can compromise decision-making efficacy. A promising direction involves developing meta-reasoning architectures that dynamically prioritize memory recall versus tool invocation based on situational urgency and confidence thresholds.
+
+Personalized Generation balances factual accuracy with preference alignment but risks over-fitting, where excessive finetuning to user profiles may reinforce cognitive biases. Techniques address surface-level alignment but lack mechanisms for ethical boundary detection. For instance, agents might inadvertently propagate harmful stereotypes when mirroring user preferences without critical oversight. Future systems could integrate value-aligned reinforcement learning with human-in-the-loop validation to preserve authenticity while preventing detrimental customization.
+
+Table 2. Datasets and metrics for personalized RAG and Agent.
+
+| Field | Metrics Category | Metrics | Datasets |
| Pre-retrieval | Textual Quality | BLEU, ROUGE, EM | Avocado Research Email Collection [57, 85], Amazon review [57, 83], Reddit comments[57, 118], Amazon ESCI dataset[82, 97], PIP |
| Information Retrieval | MAP, MRR, NDCG, Precision, Recall, RBP | AOL[88, 174], WARRIORS[99], Personalized Results Re-Ranking benchmark [6], delicio.us [9, 15, 144, 172], Flickr [9, 108], CiteULike [10, 14], LRDP [12], Delicious [141], Bibsonomy [79], Wikipedia [8, 33] |
| Classification | Accuracy, Macro-F1 | SCAN [56, 173], AITA WORKSM[53, 80], Robust04 [61] |
| Others | XEntropy, PMS, Image-Align, PQEC, ProfOverlap | Amazon ESCI dataset[82, 97], PIP, Bibsonomy [79] |
| Retrieval | Textual Quality | BLEU, ROUGE, Dis, PPL | TOPDIAL [130], Pchatbot [93], DuLemon [150] |
| Information Retrieval | Recall, MRR, Precision, F1 | LiveChat [34], Pchatbot [93], DuLemon [150] |
| Classification | Accuracy, Succ | TOPDIAL [130], PersonalityEvid [119], DuLemon [150], PersonalityEdit [75] |
| Others | Fluency, Coherence, Plausibility, ES, DD, TPEI, PAE | PersonalityEvid [119], PersonalityEdit [75] |
| Generation | Textual Quality | BLEU, ROUGE, Dis, PPL, METEOR | LaMP [105], Long LaMP [55], Dulemon [150], PGraphRAG [5], AmazonQA/Products [29], Reddit [170], MedicalDialogue [162] |
| Classification | Accuracy, F1, Persona F1 | LaMP [105], Long LaMP [55], Dulemon [150], AmazonQA/Products [29], Reddit [170], MedicalDialogue [162] |
| Regression | MAE, RMSE | LaMP [105], Long LaMP [55], PGraphRAG [5] |
| Others | Fluency, Mean Success Rate, Median Relative Improvements | Personalized-Gen [3] |
| Agent | Textual Quality | BLEU, ROUGE, METEOR, CIDer, EM, Fluency, Coherence, Instruction Adherence, Consistency related metrics | RICO [126], RoleBench [139], Shao et al. [110], Socialbench [18], MMRole-Data [27], ROLEPERSONALITY [96], ChatHarui [134], Character-LLM-Data [153], Knowledge Behind Persona [41], Wang et al. [137], Wang et al. [135], Zheng et al. [169] |
| Information Retrieval | Recall, F1, Precision | Knowledge Behind Persona [41] |
| Classification | Accuracy, Failure Rate, Classification Accuracy, Preference Rate, Correctness | MIT-BIH Arrhythmia Database [148], VirtualHome [44], Socialbench [18], ARC [100], AGIEval [100], HellaSwag [100], MedMCQA [100], AQUA-RAT [100], LogiQA [100], LSAT-AR [100], LSAT-LR [100], LSAT-RC [100], SAT-English [100], SAT-Math [100], PersonalWAB [17], TravelPlanner+ [114] |
| Others | Pass@k, Executability, Productivity, Plausibility of the Story | Hong et al. [40], Zheng et al. [169] |
+
+# 5 EVALUATION AND DATASET
+
+In the evolving landscape of personalization, from RAG to advanced Agent-based systems, the evaluation of models relies heavily on diverse datasets and metrics tailored to specific tasks. This survey categorizes metrics into several key types: Textual Quality metrics (e.g., BLEU, ROUGE, METEOR) assess the fluency and coherence of generated outputs; Information Retrieval metrics (e.g., MAP, MRR, Recall) evaluate the accuracy and relevance of retrieved information; Classification metrics (e.g., Accuracy, F1) measure task-specific correctness; Regression metrics (e.g., MAE, RMSE) quantify prediction errors; and Other metrics (e.g., Fluency, Pass@k) address domain-specific or task-unique aspects like plausibility or executability. These metrics span pre-retrieval, retrieval, generation, and agent-based personalization approaches, reflecting their varied objectives. To provide a comprehensive overview, we compile an extensive list of datasets across these fields, as detailed in Table 2. These datasets, paired with their respective metrics, enable researchers to benchmark and refine personalized systems, from enhancing query rewriting to enabling autonomous agents in physical and virtual environments.
+
+# 6 CHALLENGES AND FUTURE DIRECTIONS
+
+Personalized RAG and agent-based systems still face several critical challenges that warrant further exploration. We list them as follows:
+
+- Balancing Personalization and Scalability: Integrating personalization data (such as preferences, history, and contextual signals) into RAG processes often increases computational complexity, making it difficult to maintain
+
+efficiency and scalability across large-scale systems. Future work could explore lightweight, adaptive embeddings and hybrid frameworks that seamlessly fuse user profiles with real-time contexts.
+
+- Evaluating Personalization Effectively: Current metrics like BLEU, ROUGE, and human evaluations fall short in capturing the nuanced alignment of outputs with dynamic user preferences, lacking tailored measures for personalization efficacy. Developing specialized benchmarks and metrics that assess long-term user satisfaction and adaptability is crucial for real-world applicability.
+- Preserving Privacy through Device-Cloud Collaboration: Personalized retrieval often involves processing sensitive user data, raising privacy concerns, especially with the increased global emphasis on data protection regulations, such as the European Union's General Data Protection Regulation (GDPR). Consequently, a promising approach is the collaborative integration of on-device small Language models which handle sensitive personal data locally, with cloud-based LLM, which provides broader contextual knowledge.
+- Personalized Agent Planning: Current research on agent planning remains mainly in its early stages, with much of the work focusing on building foundational frameworks such as GUI agents [81] and the application of agents across diverse domains [131]. Notably, the incorporation of personalized approaches has yet to be widely adopted. Exploring how to integrate personalized support into existing frameworks to enhance user experience represents a promising and valuable direction for future investigation.
+- Ensuring Ethical and Coherent Systems: Bias in data processing, privacy concerns in user profiling, and coherence across retrieval and generation stages remain unresolved. Future directions should prioritize ethical safeguards, privacy-preserving techniques, and cross-stage optimization to build trustworthy, unified personalized systems.
+
+# 7 CONCLUSION
+
+In this paper, we explore the landscape of personalization from Retrieval-Augmented Generation (RAG) to advanced LLM-based Agents, detailing adaptations across pre-retrieval, retrieval, and generation stages while extending into agentic capabilities. By reviewing recent literature, datasets, and metrics, we highlight the progress and diversity in enhancing user satisfaction through tailored AI systems. However, challenges such as scalability, effective evaluation, and ethical concerns underscore the need for innovative solutions. Future research should focus on lightweight frameworks, specialized benchmarks, and privacy-preserving techniques to advance personalized AI. Relevant papers and resources are also compiled online for ease of future research.
+
+# REFERENCES
+
+[1] 2021. BERT: a review of applications in natural language processing and understanding. arXiv preprint arXiv:2103.11943 (2021).
+[2] Mahyar Abbasian, Iman Azimi, Amir M Rahmani, and Ramesh Jain. 2023. Conversational health agents: A personalized llm-powered agent framework. arXiv preprint arXiv:2310.02374 (2023).
+[3] Bashar Alhafni, Vivek Kulkarni, Dhruv Kumar, and Vipul Raheja. 2024. Personalized Text Generation with Fine-Grained Linguistic Control. In Proceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024). 88–101.
+[4] Amazon. [n.d.]. Amazon Customer Review Dataset. Online dataset. https://nijianmo.github.io/amazon/
+[5] Steven Au, Cameron J Dimacali, Ojasmitha Pedirappagari, Namyong Park, Franck Dernoncourt, Yu Wang, Nikos Kanakaris, Hanieh Deilamsalehy, Ryan A Rossi, and Nesreen K Ahmed. 2025. Personalized Graph-Based Retrieval for Large Language Models. arXiv preprint arXiv:2501.02157 (2025).
+[6] Elias Bassani, Pranav Kasela, Alessandro Raganato, and Gabriella Pasi. 2022. A multi-domain benchmark for personalized search evaluation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3822-3827.
+[7] Elias Bassani, Nicola Tonellotto, and Gabriella Pasi. 2023. Personalized query expansion with contextual word embeddings. ACM Transactions on Information Systems 42, 2 (2023), 1-35.
+[8] Oliver Baumann and Mirco Schoenfeld. 2024. PSQE: Personalized Semantic Query Expansion for user-centric query disambiguation. (2024).
+
+[9] Matthias Bender, Tom Crecelius, Mouna Kacimi, Sebastian Michel, Thomas Neumann, Josiane Xavier Parreira, Ralf Schenkel, and Gerhard Weikum. 2008. Exploiting social relations for query expansion and result ranking. In 2008 IEEE 24th International Conference on Data Engineering Workshop. IEEE, 501-506.
+[10] Marin Bertier, Rachid Guerraoui, Vincent Leroy, and Anne-Marie Kermarrec. 2009. Toward personalized query expansion. In Proceedings of the Second ACM EuroSys Workshop on Social Network Systems. 7-12.
+[11] Keping Bi, Qingyao Ai, and W Bruce Croft. 2021. Learning a fine-grained review-based transformer model for personalized product search. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 123-132.
+[12] Claudio Biancalana and Alessandro Micarelli. 2009. Social tagging in query expansion: A new way for personalized web search. In 2009 International Conference on Computational Science and Engineering, Vol. 4. IEEE, 1060-1065.
+[13] Microsoft Bing. [n.d]. Bing Search Engine. https://www.bing.com
+[14] Mohamed Reda Bouadjenek, Hakim Hacid, and Mokrane Bouzeghoub. 2019. Personalized social query expansion using social annotations. Transactions on Large-Scale Data-and Knowledge-Centered Systems XL (2019), 1-25.
+[15] Mohamed Reda Bouadjenek, Hakim Hacid, Mokrane Bouzeghoub, and Johann Daigremont. 2011. Personalized social query expansion using social bookmarking systems. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval. 1113-1114.
+[16] Domenico Bulfamante. 2023. Generative enterprise search with extensible knowledge base using ai. Ph.D. Dissertation. Politecnico di Torino.
+[17] Hongru Cai, Yongqi Li, Wenjie Wang, ZHU Fengbin, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. [n. d]. Large Language Models Empowered Personalized Web Agents. In THE WEB CONFERENCE 2025.
+[18] Hongzhan Chen, Hehong Chen, Ming Yan, Wenshen Xu, Xing Gao, Weizhou Shen, Xiaojun Quan, Chenliang Li, Ji Zhang, Fei Huang, et al. 2024. Socialbench: Sociality evaluation of role-playing conversational agents. arXiv preprint arXiv:2403.13679 (2024).
+[19] Jiangjie Chen, Xintao Wang, Rui Xu, Siyu Yuan, Yikai Zhang, Wei Shi, Jian Xie, Shuang Li, Ruihan Yang, Tinghui Zhu, et al. 2024. From persona to personalization: A survey on role-playing language agents. arXiv preprint arXiv:2404.18231 (2024).
+[20] Ruizhe Chen, Xiaotian Zhang, Meng Luo, Wenhao Chai, and Zuozhu Liu. 2024. Pad: Personalized alignment of llms at decoding-time. arXiv preprint arXiv:2410.04070 (2024).
+[21] Weijie Chen, Ting Bai, Jinbo Su, Jian Luan, Wei Liu, and Chuan Shi. 2024. Kg-retriever: Efficient knowledge indexing for retrieval-augmented large language models. arXiv preprint arXiv:2412.05547 (2024).
+[22] Zheng Chen, Ziyan Jiang, Fan Yang, Eunah Cho, Xing Fan, Xiaojiang Huang, Yanbin Lu, and Aram Galstyan. 2023. Graph meets LLM: A novel approach to collaborative filtering for robust conversational understanding. arXiv preprint arXiv:2305.14449 (2023).
+[23] Hao Cheng, Shuo Wang, Wensheng Lu, Wei Zhang, Mingyang Zhou, Kezhong Lu, and Hao Liao. 2023. Explainable recommendation with personalized review retrieval and aspect learning. arXiv preprint arXiv:2306.12657 (2023).
+[24] Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. 2023. Adapting language models to compress contexts. arXiv preprint arXiv:2305.14788 (2023).
+[25] Eunah Cho, Ziyan Jiang, Jie Hao, Zheng Chen, Saurabh Gupta, Xing Fan, and Chenlei Guo. 2021. Personalized search-based query rewrite system for conversational ai. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI. 179-188.
+[26]Hyundong Cho, Karishma Sharma, Nicolaas Jedema, Leonardo FR Ribeiro, Alessandro Moschitti, Ravi Krishnan, and Jonathan May. 2025. TuningFree Personalized Alignment via Trial-Error-Explain In-Context Learning. arXiv preprint arXiv:2502.08972 (2025).
+[27] Yanqi Dai, Huanran Hu, Lei Wang, Shengjie Jin, Xu Chen, and Zhiwu Lu. 2024. Mmrole: A comprehensive framework for developing and evaluating multimodal role-playing agents. arXiv preprint arXiv:2408.04203 (2024).
+[28] Wenlong Deng, Christos Thrampoulidis, and Xiaoxiao Li. 2024. Unlocking the potential of prompt-tuning in bridging generalized and personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6087-6097.
+[29] Yang Deng, Yaliang Li, Wenxuan Zhang, Bolin Ding, and Wai Lam. 2022. Toward personalized answer generation in e-commerce via multiperspective preference modeling. ACM Transactions on Information Systems (TOIS) 40, 4 (2022), 1-28.
+[30] Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The Faiss library. (2024). arXiv:2401.08281 [cs.LG]
+[31] ESPN. [n.d.]. ESPN Sports Statistics Dataset. Online dataset.
+[32] 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.
+[33] Byron J Gao, David C Anastasiu, and Xing Jiang. 2010. Utilizing user-input contextual terms for query disambiguation. In Coling 2010: Posters. 329-337.
+[34] Jingsheng Gao, Yixin Lian, Ziyi Zhou, Yuzhuo Fu, and Baoyuan Wang. 2023. LiveChat: A large-scale personalized dialogue dataset automatically constructed from live streaming. arXiv preprint arXiv:2306.08401 (2023).
+[35] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. 2023. Retrievalaugmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 2 (2023).
+[36] Google. [n.d.]. Google Search. https://www.google.com
+
+[37] Jia-Chen Gu, Hui Liu, Zhen-Hua Ling, Quan Liu, Zhigang Chen, and Xiaodan Zhu. 2021. Partner matters! an empirical study on fusing personas for personalized response selection in retrieval-based chatbots. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 565-574.
+[38] Jie Hao, Yang Liu, Xing Fan, Saurabh Gupta, Saleh Soltan, Rakesh Chada, Pradeep Natarajan, Chenlei Guo, and Gokhan Tur. 2022. CGF: Constrained generation framework for query rewriting in conversational AI. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track. 475-483.
+[39] Nicola Henze, Peter Dolog, and Wolfgang Nejdl. 2004. Reasoning and ontologies for personalized e-learning in the semantic web. Journal of Educational Technology & Society 7, 4 (2004), 82-97.
+[40] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. 2023. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352 3, 4 (2023), 6.
+[41] WANG Hongru, Minda Hu, Yang Deng, Rui Wang, Fei Mi, Weichao Wang, Yasheng Wang, Wai-Chung Kwan, Irwin King, and Kam-Fai Wong. [n. d]. Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues. In The 2023 Conference on Empirical Methods in Natural Language Processing.
+[42] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR 1, 2 (2022), 3.
+[43] Qiushi Huang, Shuai Fu, Xubo Liu, Wenwu Wang, Tom Ko, Yu Zhang, and Lilian Tang. 2024. Learning retrieval augmentation for personalized dialogue generation. arXiv preprint arXiv:2406.18847 (2024).
+[44] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. 2022. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning. PMLR, 9118-9147.
+[45] Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of LLM agents: A survey. arXiv preprint arXiv:2402.02716 (2024).
+[46] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. 2023. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653 (2023).
+[47] Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, and Prithviraj Ammanabrolu. 2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv preprint arXiv:2310.11564 (2023).
+[48] Pengyue Jia, Yiding Liu, Xiangyu Zhao, Xiaopeng Li, Changying Hao, Shuaiqiang Wang, and Dawei Yin. 2024. MILL: Mutual Verification with Large Language Models for Zero-Shot Query Expansion. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2498-2518.
+[49] Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Zhang, and Yixin Zhu. 2023. Evaluating and inducing personality in pre-trained language models. Advances in Neural Information Processing Systems 36 (2023), 10622-10643.
+[50] Hideaki Joko, Shubham Chatterjee, Andrew Ramsay, Arjen P De Vries, Jeff Dalton, and Faegheh Hasibi. 2024. Doing personal laps: Llm-augmented dialogue construction for personalized multi-session conversational search. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 796-806.
+[51] Wang-Cheng Kang, Jianmo Ni, Nikhil Mehta, Maheswaran Sathiamoorthy, Lichan Hong, Ed Chi, and Derek Zhiyuan Cheng. 2023. Do llms understand user preferences? evaluating llms on user rating prediction. arXiv preprint arXiv:2305.06474 (2023).
+[52] Manojkumar Rangasamy Kannadasan and Grigor Aslanyan. 2019. Personalized query auto-completion through a lightweight representation of the user context. arXiv preprint arXiv:1905.01386 (2019).
+[53] Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Greg Corrado, Laszlo Lukacs, Marina Ganea, Peter Young, and Vivek Ramavajjala. 2016. Smart Reply: Automated Response Suggestion for Email. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (San Francisco, California, USA) (KDD '16). Association for Computing Machinery, New York, NY, USA, 955-964. https://doi.org/10.1145/2939672.2939801
+[54] Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. 2024. Reinforcement learning for optimizing rag for domain chatbots. arXiv preprint arXiv:2401.06800 (2024).
+[55] Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, et al. 2024. Longlamp: A benchmark for personalized long-form text generation. arXiv preprint arXiv:2407.11016 (2024).
+[56] Brenden Lake and Marco Baroni. 2018. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning. PMLR, 2873-2882.
+[57] Cheng Li, Mingyang Zhang, Qiao zhu Mei, Weize Kong, and Michael Bendersky. 2024. Learning to rewrite prompts for personalized text generation. In Proceedings of the ACM Web Conference 2024. 3367-3378.
+[58] Changhao Li, Yuchen Zhuang, Rushi Qiang, Haotian Sun, Hanjun Dai, Chao Zhang, and Bo Dai. 2024. Matryoshka: Learning to Drive Black-Box LLMs with LLMs. arXiv preprint arXiv:2410.20749 (2024).
+[59] Lei Li, Yongfeng Zhang, and Li Chen. 2023. Personalized prompt learning for explainable recommendation. ACM Transactions on Information Systems 41, 4 (2023), 1-26.
+[60] Sen Li, Fuyu Lv, Taiwei Jin, Guiyang Li, Yukun Zheng, Tao Zhuang, Qingwen Liu, Xiaoyi Zeng, James Kwok, and Qianli Ma. 2022. Query rewriting in taobao search. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3262-3271.
+
+[61] Xiaopeng Li, Lixin Su, Pengyue Jia, Xiangyu Zhao, Suqi Cheng, Junfeng Wang, and Dawei Yin. 2023. Agent4ranking: Semantic robust ranking via personalized query rewriting using multi-agent llm. arXiv preprint arXiv:2312.15450 (2023).
+[62] Xinyu Li, Ruiyang Zhou, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133 (2024).
+[63] Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, et al. 2024. Personal llm agents: Insights and survey about the capability, efficiency and security. arXiv preprint arXiv:2401.05459 (2024).
+[64] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281 (2023).
+[65] Ruixue Lian, Sixing Lu, Clint Solomon, Gustavo Aguilar, Pragaash Ponnusamy, Jialong Han, Chengyuan Ma, and Chenlei Guo. 2023. PersonalTM: Transformer memory for personalized retrieval. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2256-2260.
+[66] Shan-Mu Lin and Chuen-Min Huang. 2006. Personalized optimal search in local query expansion. In Proceedings of the 18th Conference on Computational Linguistics and Speech Processing. 221-236.
+[67] Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is chatgpt a good recommender? a preliminary study. arXiv preprint arXiv:2304.10149 (2023).
+[68] Jiahong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Jieming Zhu, Minda Hu, Menglin Yang, and Irwin King. 2025. A Survey of Personalized Large Language Models: Progress and Future Directions. arXiv preprint arXiv:2502.11528 (2025).
+[69] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics 12 (2024), 157-173.
+[70] Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024. Once: Boosting content-based recommendation with both open-and closed-source large language models. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining. 452-461.
+[71] Shuai Liu, Hyundong J Cho, Marjorie Freedman, Xuezhe Ma, and Jonathan May. 2023. RECAP: retrieval-enhanced context-aware prefix encoder for personalized dialogue response generation. arXiv preprint arXiv:2306.07206 (2023).
+[72] Tyler Lu and Craig Boutilier. 2011. Budgeted social choice: From consensus to personalized decision making. In *IJCAI*, Vol. 11, 280-286.
+[73] Zhengyi Ma, Zhicheng Dou, Yutao Zhu, Hanxun Zhong, and Ji-Rong Wen. 2021. One chatbot per person: Creating personalized chatbots based on implicit user profiles. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 555-564.
+[74] Aman Madaan, Niket Tandon, Peter Clark, and Yiming Yang. 2022. Memory-assisted prompt editing to improve GPT-3 after deployment. arXiv preprint arXiv:2201.06009 (2022).
+[75] Shengyu Mao, Xiaohan Wang, Mengru Wang, Yong Jiang, Pengjun Xie, Fei Huang, and Ningyu Zhang. 2024. Editing Personality for Large Language Models. In CCF International Conference on Natural Language Processing and Chinese Computing. Springer, 241-254.
+[76] Yuren Mao, Xuemei Dong, Wenyi Xu, Yunjun Gao, Bin Wei, and Ying Zhang. 2024. Fit-rag: black-box rag with factual information and token reduction. arXiv preprint arXiv:2403.14374 (2024).
+[77] Puneet Mathur, Zhe Liu, Ke Li, Yingyi Ma, Gil Keren, Zeeshan Ahmed, Dinesh Manocha, and Xuedong Zhang. 2023. Personal: Language model personalization via domain-distributed span aggregated k-nearest n-gram retrieval augmentation. In Findings of the Association for Computational Linguistics: EMNLP 2023. 11314-11328.
+[78] Fatemehsadat Mireshghallah, Vaishnavi Shrivastava, Milad Shokouhi, Taylor Berg-Kirkpatrick, Robert Sim, and Dimitrios Dimitriadis. 2021. Identifier: Implicit user representations for simple and effective personalized sentiment analysis. arXiv preprint arXiv:2110.00135 (2021).
+[79] Philippe Mulhem, Nawal Ould Amer, and Mathias Gery. 2016. Axiomatic term-based personalized query expansion using bookmarking system. In International Conference on Database and Expert Systems Applications. Springer, 235-243.
+[80] Sheshera Mysore, Zhuoran Lu, Mengting Wan, Longqi Yang, Steve Menezes, Tina Baghaee, Emmanuel Barajas Gonzalez, Jennifer Neville, and Tara Safavi. 2023. Pearl: Personalizing large language model writing assistants with generation-calibrated retrievers. arXiv preprint arXiv:2311.09180 (2023).
+[81] Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, et al. 2024. Gui agents: A survey. arXiv preprint arXiv:2412.13501 (2024).
+[82] Duy A Nguyen, Rishi Kesav Mohan, Van Yang, Pritom Saha Akash, and Kevin Chen-Chuan Chang. 2025. RL-based Query Rewriting with Distilled LLM for online E-Commerce Systems. arXiv preprint arXiv:2501.18056 (2025).
+[83] Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 188-197.
+[84] Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O'Banion, and Jun Xie. 2024. User-llm: Efficient llm contextualization with user embeddings. arXiv preprint arXiv:2402.13598 (2024).
+[85] Douglas Oard, William Webber, David Kirsch, and Sergey Golitsynski. 2015. Avocado research email collection. Philadelphia: Linguistic Data Consortium (2015).
+[86] U.S. National Library of Medicine. [n.d.]. PubMed: A Free Resource for Biomedical Literature. https://pubmed.ncbi.nlm.nih.gov/
+[87] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology. 1-22.
+
+[88] Greg Pass, Abdur Chowdhury, and Cayley Torgeson. 2006. A picture of search. In Proceedings of the 1st International Conference on Scalable Information Systems (Hong Kong) (InfoScale '06). Association for Computing Machinery, New York, NY, USA, 1-es. https://doi.org/10.1145/1146847.1146848
+[89] Vadim Igorevich Pavliukevich, Alina Khasanovna Zherdeva, Olesya Vladimirovna Makhnytkina, and Dmitriy Viktorovich Dyrmovskiy. [n. d.]. Improving RAG with LoRA finetuning for persona text generation. ([n. d.]).
+[90] Dan Peng, Zhihui Fu, and Jun Wang. 2024. Pocketllm: Enabling on-device fine-tuning for personalized llms. arXiv preprint arXiv:2407.01031 (2024).
+[91] Qiyao Peng, Hongtao Liu, Hongyan Xu, Qing Yang, Minglai Shao, and Wenjun Wang. 2024. Review-LLM: Harnessing Large Language Models for Personalized Review Generation. arXiv:2407.07487 [cs.CL] https://arxiv.org/abs/2407.07487
+[92] Hongjin Qian, Zhicheng Dou, Yutao Zhu, Yueyuan Ma, and Ji-Rong Wen. 2021. Learning implicit user profile for personalized retrieval-based chatbot. In proceedings of the 30th ACM international conference on Information & Knowledge Management. 1467-1477.
+[93] Hongjin Qian, Xiahe Li, Hanxun Zhong, Yu Guo, Yueyuan Ma, Yutao Zhu, Zhanliang Liu, Zhicheng Dou, and Ji-Rong Wen. 2021. Pchatbot: a large-scale dataset for personalized chatbot. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 2470-2477.
+[94] Xiaoru Qu, Yifan Wang, Zhao Li, and Jun Gao. 2024. Graph-enhanced prompt learning for personalized review generation. Data Science and Engineering 9, 3 (2024), 309-324.
+[95] A. Rajaraman and J.D. Ullman. 2011. Mining of Massive Datasets. Cambridge University Press. https://books.google.co.uk/books?id=OefRhZyYOb0C
+[96] Yiting Ran, Xintao Wang, Rui Xu, Xinfeng Yuan, Jiaqing Liang, Deqing Yang, and Yanghua Xiao. 2024. Capturing minds, not just words: Enhancing role-playing language models with personality-indicative data. arXiv preprint arXiv:2406.18921 (2024).
+[97] Chandan K. Reddy, Lluis Marquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopadhyay, Arnab Biswas, Anlu Xing, and Karthik Subbian. 2022. Shopping Queries Dataset: A Large-Scale ESCI Benchmark for Improving Product Search. (2022). arXiv:2206.06588
+[98] Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084 (2019).
+[99] Ruiyang Ren, Peng Qiu, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Hua Wu, Ji-Rong Wen, and Haifeng Wang. 2024. Bases: Large-scale web search user simulation with large language model based agents. arXiv preprint arXiv:2402.17505 (2024).
+[100] Matthew Renze and Erhan Guven. 2024. Self-reflection in llm agents: Effects on problem-solving performance. arXiv preprint arXiv:2405.06682 (2024).
+[101] 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).
+[102] Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval 3, 4 (2009), 333-389.
+[103] 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.
+[104] Alireza Salemi, Cheng Li, Mingyang Zhang, Qiao zhu Mei, Weize Kong, Tao Chen, Zhuowan Li, Michael Bendersky, and Hamed Zamani. 2025. Reasoning-Enhanced Self-Training for Long-Form Personalized Text Generation. arXiv preprint arXiv:2501.04167 (2025).
+[105] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 7370-7392.
+[106] Alireza Salemi and Hamed Zamani. 2024. Learning to Rank for Multiple Retrieval-Augmented Models through Iterative Utility Maximization. arXiv preprint arXiv:2410.09942 (2024).
+[107] Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect?. In International Conference on Machine Learning. PMLR, 29971-30004.
+[108] Rossano Schifanella, Alain Barrat, Ciro Cattuto, Benjamin Markines, and Filippo Menczer. 2010. Folks in folksonomies: social link prediction from shared metadata. In Proceedings of the third ACM international conference on Web search and data mining. 271-280.
+[109] Noor Shaker, Georgios Yannakakis, and Julian Togelius. 2010. Towards automatic personalized content generation for platform games. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, Vol. 6. 63-68.
+[110] Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2023. Character-llm: A trainable agent for role-playing. arXiv preprint arXiv:2310.10158 (2023).
+[111] Jocelyn Shen, Joel Mire, Hae Won Park, Cynthia Breazeal, and Maarten Sap. 2024. HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs. arXiv preprint arXiv:2405.17633 (2024).
+[112] Yunxiao Shi, Xing Zi, Zijing Shi, Haimin Zhang, Qiang Wu, and Min Xu. 2024. Eragent: Enhancing retrieval-augmented language models with improved accuracy, efficiency, and personalization. arXiv preprint arXiv:2405.06683 (2024).
+[113] Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. 2025. Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG. arXiv preprint arXiv:2501.09136 (2025).
+[114] Harmanpreet Singh, Nikhil Verma, Yixiao Wang, Manasa Bharadwaj, Homa Fashandi, Kevin Ferreira, and Chul Lee. 2024. Personal Large Language Model Agents: A Case Study on Tailored Travel Planning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 486-514.
+[115] Shamane Siriwardhana, Rivindu Weerasekera, Elliott Wen, Tharindu Kaluarachchi, Rajib Rana, and Suranga Nanayakkara. 2023. Improving the domain adaptation of retrieval augmented generation (RAG) models for open domain question answering. Transactions of the Association for Computational Linguistics 11 (2023), 1-17.
+[116] Mingyang Song and Mao Zheng. 2024. A Survey of Query Optimization in Large Language Models. arXiv preprint arXiv:2412.17558 (2024).
+
+[117] Spotify. 2023. Annoy: Approximate Nearest Neighbors in C++/Python. https://github.com/spotify/annoy
+[118] Stuck_In_the Matrix. 2015. Reddit Public Comments (2007-10 through 2015-05). (2015). https://www.reddit.com/r/datasets/comments/3bxlg7/i_have EVERY_publicly-available Reddit_COMMENT/
+[119] Lei Sun, Jinming Zhao, and Qin Jin. 2024. Revealing Personality Traits: A New Benchmark Dataset for Explanable Personality Recognition on Dialogues. arXiv preprint arXiv:2409.19723 (2024).
+[120] Zhaoxuan Tan, Zheyuan Liu, and Meng Jiang. 2024. Personalized pieces: Efficient personalized large language models through collaborative efforts. arXiv preprint arXiv:2406.10471 (2024).
+[121] 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 preprint arXiv:2402.04401 (2024).
+[122] Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2025. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. arXiv:2402.04401 [cs.CL] https://arxiv.org/abs/2402.04401
+[123] Quan Tu, Shilong Fan, Zihang Tian, and Rui Yan. 2024. Charactereval: A Chinese benchmark for role-playing conversational agent evaluation. arXiv preprint arXiv:2401.01275 (2024).
+[124] Cornell University. [n.d.]. arXiv: An Open Access Repository for Research. https://arxiv.org/
+[125] Hemanth Vemuri, Sheshansh Agrawal, Shivam Mittal, Deepak Saini, Akshay Soni, Abhinav V Sambasivan, Wenhao Lu, Yajun Wang, Mehul Parsana, Purushottam Kar, et al. 2023. Personalized retrieval over millions of items. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1014-1022.
+[126] Bryan Wang, Gang Li, and Yang Li. 2023. Enabling conversational interaction with mobile ui using large language models. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1-17.
+[127] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023).
+[128] Hongru Wang, Wenyu Huang, Yang Deng, Rui Wang, Zezhong Wang, Yufei Wang, Fei Mi, Jeff Z Pan, and Kam-Fai Wong. 2024. Unims-rag: A unified multi-source retrieval-augmented generation for personalized dialogue systems. arXiv preprint arXiv:2401.13256 (2024).
+[129] Hongru Wang, Rui Wang, Fei Mi, Yang Deng, Zezhong Wang, Bin Liang, Ruifeng Xu, and Kam-Fai Wong. 2023. Cue-CoT: Chain-of-thought prompting for responding to in-depth dialogue questions with LLMs. arXiv preprint arXiv:2305.11792 (2023).
+[130] Jian Wang, Yi Cheng, Dongding Lin, Chak Tou Leong, and Wenjie Li. 2023. Target-oriented proactive dialogue systems with personalization: Problem formulation and dataset curation. arXiv preprint arXiv:2310.07397 (2023).
+[131] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2024. A survey on large language model based autonomous agents. Frontiers of Computer Science 18, 6 (2024), 186345.
+[132] Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query expansion with large language models. arXiv preprint arXiv:2303.07678 (2023).
+[133] Lei Wang, Jingsen Zhang, Hao Yang, Zhiyuan Chen, Jiakai Tang, Zeyu Zhang, Xu Chen, Yankai Lin, Ruihua Song, Wayne Xin Zhao, et al. 2023. User behavior simulation with large language model based agents. arXiv preprint arXiv:2306.02552 (2023).
+[134] Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, et al. 2023. Incharacter: Evaluating personality fidelity in role-playing agents through psychological interviews. arXiv preprint arXiv:2310.17976 (2023).
+[135] Yixiao Wang, Homa Fashandi, and Kevin Ferreira. 2024. Investigating the Personality Consistency in Quantized Role-Playing Dialogue Agents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 239–255.
+[136] Yu Wang, Yifan Gao, Xiusi Chen, Haoming Jiang, Shiyang Li, Jingfeng Yang, Qingyu Yin, Zheng Li, Xian Li, Bing Yin, et al. [n.d.]. MEMORYLLM: Towards Self-Updatable Large Language Models. In Forty-first International Conference on Machine Learning.
+[137] Zheng Wang, Zhongyang Li, Zeren Jiang, Dandan Tu, and Wei Shi. 2024. Crafting Personalized Agents through Retrieval-Augmented Generation on Editable Memory Graphs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 4891-4906.
+[138] Zijie J Wang and Duen Horng Chau. 2024. MeMemo: On-device Retrieval Augmentation for Private and Personalized Text Generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2765-2770.
+[139] Zekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, et al. 2023. Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. arXiv preprint arXiv:2310.00746 (2023).
+[140] Tianxin Wei, Bowen Jin, Ruirui Li, Hansi Zeng, Zhengyang Wang, Jianhui Sun, Qingyu Yin, Hanqing Lu, Suhang Wang, Jingrui He, et al. 2024. Towards unified multi-modal personalization: Large vision-language models for generative recommendation and beyond. arXiv preprint arXiv:2403.10667 (2024).
+[141] Robert Wetzker, Carsten Zimmermann, and Christian Bauchage. 2008. Analyzing social bookmarking systems: A del. icio. us cookbook. In Proceedings of the ECAI 2008 Mining Social Data Workshop. 26-30.
+[142] Stanisław Wozniak, Bartlomiej Koptyra, Arkadiusz Janz, Przemysław Kazienko, and Jan Kocón. 2024. Personalized large language models. arXiv preprint arXiv:2402.09269 (2024).
+[143] Junde Wu, Jiayuan Zhu, Yunli Qi, Jingkun Chen, Min Xu, Filippo Menolascina, and Vicente Grau. 2024. Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187 (2024).
+[144] Xuan Wu, Dong Zhou, Yu Xu, and Seamus Lawless. 2017. Personalized query expansion utilizing multi-relational social data. In 2017 12th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP). IEEE, 65-70.
+
+[145] Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender Systems. 12-22.
+[146] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences 68, 2 (2025), 121101.
+[147] Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. 641-649.
+[148] Huatao Xu, Liying Han, Qirui Yang, Mo Li, and Mani Srivastava. 2024. Penetrative ai: Making llms comprehend the physical world. In Proceedings of the 25th International Workshop on Mobile Computing Systems and Applications. 1-7.
+[149] Hongyan Xu, Hongtao Liu, Pengfei Jiao, and Wenjun Wang. 2021. Transformer reasoning network for personalized review summarization. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1452-1461.
+[150] Xinchao Xu, Zhibin Gou, Wenquan Wu, Zheng-Yu Niu, Hua Wu, Haifeng Wang, and Shihang Wang. 2022. Long time no see! open-domain conversation with long-term persona memory. arXiv preprint arXiv:2203.05797 (2022).
+[151] Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, and Tat-Seng Chua. 2025. Personalized Generation In Large Model Era: A Survey. arXiv preprint arXiv:2503.02614 (2025).
+[152] Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. 2024. Personalized federated continual learning via multi-granularity prompt. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4023-4034.
+[153] Xiaoyan Yu, Tongxu Luo, Yifan Wei, Fangyu Lei, Yiming Huang, Hao Peng, and Liehuang Zhu. 2024. Neeko: Leveraging dynamic lora for efficient multi-character role-playing agent. arXiv preprint arXiv:2402.13717 (2024).
+[154] Xinfeng Yuan, Siyu Yuan, Yuhan Cui, Tianhe Lin, Xintao Wang, Rui Xu, Jiangjie Chen, and Deqing Yang. 2024. Evaluating character understanding of large language models via character profiling from fictional works. arXiv preprint arXiv:2404.12726 (2024).
+[155] Hansi Zeng, Surya Kallumadi, Zaid Alibadi, Rodrigo Nogueira, and Hamed Zamani. 2023. A personalized dense retrieval framework for unified information access. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 121-130.
+[156] Saber Zerhoudi and Michael Granitzer. 2024. PersonaRAG: Enhancing Retrieval-Augmented Generation Systems with User-Centric Agents. arXiv preprint arXiv:2407.09394 (2024).
+[157] Han Zhang, Songlin Wang, Kang Zhang, Zhiling Tang, Yunjiang Jiang, Yun Xiao, Weipeng Yan, and Wen-Yun Yang. 2020. Towards personalized and semantic retrieval: An end-to-end solution for e-commerce search via embedding learning. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2407-2416.
+[158] Jiarui Zhang. 2024. Guided profile generation improves personalization with llms. arXiv preprint arXiv:2409.13093 (2024).
+[159] Jesse Zhang, Jiahui Zhang, Karl Pertsch, Ziyi Liu, Xiang Ren, Minsuk Chang, Shao-Hua Sun, and Joseph J Lim. [n.d.]. Bootstrap Your Own Skills: Learning to Solve New Tasks with Large Language Model Guidance. In 7th Annual Conference on Robot Learning.
+[160] Kai Zhang, Yangyang Kang, Fubang Zhao, and Xiaozhong Liu. 2023. LLM-based medical assistant personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023).
+[161] Kaiyan Zhang, Jianyu Wang, Ermo Hua, Biqing Qi, Ning Ding, and Bowen Zhou. 2024. Cogenesis: A framework collaborating large and small language models for secure context-aware instruction following. arXiv preprint arXiv:2403.03129 (2024).
+[162] Kai Zhang, Fubang Zhao, Yangyang Kang, and Xiaozhong Liu. 2023. Memory-augmented llm personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023).
+[163] Wenlin Zhang, Chuhan Wu, Xiangyang Li, Yuhao Wang, Kuicai Dong, Yichao Wang, Xinyi Dai, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2025. LLMTreeRec: Unleashing the Power of Large Language Models for Cold-Start Recommendations. In Proceedings of the 31st International Conference on Computational Linguistics. 886-896.
+[164] Yanyue Zhang, Yulan He, and Deyu Zhou. 2025. Rehearse With User: Personalized Opinion Summarization via Role-Playing based on Large Language Models. arXiv preprint arXiv:2503.00449 (2025).
+[165] You Zhang, Jin Wang, Liang-Chih Yu, Dan Xu, and Xuejie Zhang. 2024. Personalized LoRA for human-centered text understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 1958-19596.
+[166] Yabin Zhang, Wenhui Yu, Erhan Zhang, Xu Chen, Lantao Hu, Peng Jiang, and Kun Gai. 2024. Recgpt: Generative personalized prompts for sequential recommendation via chatgpt training paradigm. arXiv preprint arXiv:2404.08675 (2024).
+[167] Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501 (2024).
+[168] Zhehao Zhang, Ryan A Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, et al. 2024. Personalization of large language models: A survey. arXiv preprint arXiv:2411.00027 (2024).
+[169] Yi Zheng, Chongyang Ma, Kanle Shi, and Haibin Huang. 2023. Agents meet okr: An object and key results driven agent system with hierarchical self-collaboration and self-evaluation. arXiv preprint arXiv:2311.16542 (2023).
+[170] Hanxun Zhong, Zhicheng Dou, Yutao Zhu, Hongjin Qian, and Ji-Rong Wen. 2022. Less is more: Learning to refine dialogue history for personalized dialogue generation. arXiv preprint arXiv:2204.08128 (2022).
+[171] Wanjun Zhong, Duyu Tang, Jiahai Wang, Jian Yin, and Nan Duan. 2021. UserAdapter: Few-shot user learning in sentiment analysis. In Findings of the Association for Computational Linguistics: ACL-JJCNLP 2021. 1484-1488.
+
+[172] Dong Zhou, Séamus Lawless, and Vincent Wade. 2012. Improving search via personalized query expansion using social media. Information retrieval 15 (2012), 218-242.
+[173] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. 2022. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022).
+[174] Yujia Zhou, Qiannan Zhu, Jiajie Jin, and Zhicheng Dou. 2024. Cognitive personalized search integrating large language models with an efficient memory mechanism. In Proceedings of the ACM Web Conference 2024. 1464-1473.
+[175] Yuchen Zhuang, Haotian Sun, Yue Yu, Rushi Qiang, Qifan Wang, Chao Zhang, and Bo Dai. [n.d.]. Hydra: Model factorization framework for black-box llm personalization, 2024. URL https://arxiv.org/abs/2406.02888 ([n.d.]).
+
+Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10147/images/3888cac0ece5531f078c8ec3406df643f228458552c866b9ad79273c9d66181b.jpg b/data/2025/2504_10xxx/2504.10147/images/3888cac0ece5531f078c8ec3406df643f228458552c866b9ad79273c9d66181b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c82b15a32294a0867b97456d7d2341cf69190887
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/3888cac0ece5531f078c8ec3406df643f228458552c866b9ad79273c9d66181b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18afa0a7fc35cc841edc4dcab73b26f362341d21f4a2d27bfa3dda9bb5a5f196
+size 170253
diff --git a/data/2025/2504_10xxx/2504.10147/images/431fadc2fb6b837e6448de02affe85f7a85af23cf964b4cebff641487aab910d.jpg b/data/2025/2504_10xxx/2504.10147/images/431fadc2fb6b837e6448de02affe85f7a85af23cf964b4cebff641487aab910d.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..378b451c9ddd4405eb4998fc8b727b3aa397e02e
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/431fadc2fb6b837e6448de02affe85f7a85af23cf964b4cebff641487aab910d.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d68c8eb13f17b17691c1e4287c45b8f5cc83c642645d64bbf67e7d91190cd0c
+size 39749
diff --git a/data/2025/2504_10xxx/2504.10147/images/684289db43c065be977860260a5ca5599fa5f676c9663e41befff2d9bcc9c089.jpg b/data/2025/2504_10xxx/2504.10147/images/684289db43c065be977860260a5ca5599fa5f676c9663e41befff2d9bcc9c089.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..20740e2df11badd80228961fe7d3413c8c8119ca
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/684289db43c065be977860260a5ca5599fa5f676c9663e41befff2d9bcc9c089.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5f7d16dbe27428ca20f0ca94e22d2fe117f8e77aedaba58a132be5cc110817b
+size 42125
diff --git a/data/2025/2504_10xxx/2504.10147/images/7122c71aefb1637d6b1b8ecf544934b92473bd4011b205e6f3793fb7593c1f2c.jpg b/data/2025/2504_10xxx/2504.10147/images/7122c71aefb1637d6b1b8ecf544934b92473bd4011b205e6f3793fb7593c1f2c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..7c8bd88c314ea62e3294a83c36a8c6f3da10aa2f
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/7122c71aefb1637d6b1b8ecf544934b92473bd4011b205e6f3793fb7593c1f2c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21aec649f1d7bf3b8aea1b7bb7f31f26dfdbccc3ef784946933441e6285a6df9
+size 3687
diff --git a/data/2025/2504_10xxx/2504.10147/images/8f9e5a0db9555c1a4ed96186bb98667b2c4fd2abae37751b9dee8a95b60f65c1.jpg b/data/2025/2504_10xxx/2504.10147/images/8f9e5a0db9555c1a4ed96186bb98667b2c4fd2abae37751b9dee8a95b60f65c1.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..eab7b3395f868fef3399cc2619b10c5510488ae1
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/8f9e5a0db9555c1a4ed96186bb98667b2c4fd2abae37751b9dee8a95b60f65c1.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:608f8605225d85a016df84ca5ab528b1fc2b810d6e253c0be178dfc2db74a2ba
+size 80585
diff --git a/data/2025/2504_10xxx/2504.10147/images/91c490fa5f8c0af657eecb79cedc896d6a77e119d6fb8493fe9338606a50ca55.jpg b/data/2025/2504_10xxx/2504.10147/images/91c490fa5f8c0af657eecb79cedc896d6a77e119d6fb8493fe9338606a50ca55.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f9d2fc7b1df39e38a1a0175541e8a06b36631f73
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/91c490fa5f8c0af657eecb79cedc896d6a77e119d6fb8493fe9338606a50ca55.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9555fe6ec8d37c2d9195083f4667eb4e1de0eb4d7e8fb959da1a52b0a0dce222
+size 3132
diff --git a/data/2025/2504_10xxx/2504.10147/images/c1f227eab52914058e52cd6428ef57f5faa6d829581489e901866e44230f1223.jpg b/data/2025/2504_10xxx/2504.10147/images/c1f227eab52914058e52cd6428ef57f5faa6d829581489e901866e44230f1223.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3243ccdeb19d540b617fb9c2d7716a841c316c58
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/c1f227eab52914058e52cd6428ef57f5faa6d829581489e901866e44230f1223.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff86b58b3122d2356ee77b8c781103684f8322145c96df115dd48a70e16db601
+size 150070
diff --git a/data/2025/2504_10xxx/2504.10147/images/c5b36e8fc5d85eef240bb82935b05407c1e84bf402d0060b0bb413dc04743f40.jpg b/data/2025/2504_10xxx/2504.10147/images/c5b36e8fc5d85eef240bb82935b05407c1e84bf402d0060b0bb413dc04743f40.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d0917dc122a54def6d80749c9dfda4893a8cdf27
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/c5b36e8fc5d85eef240bb82935b05407c1e84bf402d0060b0bb413dc04743f40.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ba59c5e64b32dd14b5c5b1dd8fd173693729868575000d7dc3c4f119916b527
+size 5289
diff --git a/data/2025/2504_10xxx/2504.10147/images/d86abcf5e5ec9f4d1a1e2cf178e6e2a2411500ea7111c2031609d3b0586d7569.jpg b/data/2025/2504_10xxx/2504.10147/images/d86abcf5e5ec9f4d1a1e2cf178e6e2a2411500ea7111c2031609d3b0586d7569.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f839dc773918e5e6d8e0f65338642f0d2d177192
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/d86abcf5e5ec9f4d1a1e2cf178e6e2a2411500ea7111c2031609d3b0586d7569.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:278441d29b632ade3748c4e2e61d1b08a3763781384b3964ffe2066b9d6e56ab
+size 164671
diff --git a/data/2025/2504_10xxx/2504.10147/images/eddf6ee96c6208637db90f8b20f4af142d11c804484f15c3dc8e58752b900628.jpg b/data/2025/2504_10xxx/2504.10147/images/eddf6ee96c6208637db90f8b20f4af142d11c804484f15c3dc8e58752b900628.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..212bf36fd92f654edeb42cd025df4a71e19ec8e3
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/eddf6ee96c6208637db90f8b20f4af142d11c804484f15c3dc8e58752b900628.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cb6f660e71e65b9bdca3ae6d5442539861ed7ea29196115966fdc5f695f5b37
+size 95542
diff --git a/data/2025/2504_10xxx/2504.10147/images/faaeeb7d36800aed003df61ec8252e783b23b8a2f81ce0769f1b50247566d8e0.jpg b/data/2025/2504_10xxx/2504.10147/images/faaeeb7d36800aed003df61ec8252e783b23b8a2f81ce0769f1b50247566d8e0.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f49e968a47848a9905a3507213391e2fe472fc00
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/images/faaeeb7d36800aed003df61ec8252e783b23b8a2f81ce0769f1b50247566d8e0.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02a0595cda3127c04f4f6ef4cca06d0e48da79f0ea55a7c405cd5450ca64c326
+size 2579
diff --git a/data/2025/2504_10xxx/2504.10147/layout.json b/data/2025/2504_10xxx/2504.10147/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..6aa343ae977e838d73ddae9fc0f0b0ba1e624243
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10147/layout.json
@@ -0,0 +1,14337 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 70,
+ 92,
+ 337,
+ 107
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 92,
+ 337,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 92,
+ 337,
+ 107
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 69,
+ 118,
+ 302,
+ 132
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 118,
+ 302,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 118,
+ 302,
+ 132
+ ],
+ "type": "text",
+ "content": "XIAOPENG LI*, City University of Hong Kong, Hong Kong"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 71,
+ 134,
+ 304,
+ 147
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 134,
+ 304,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 134,
+ 304,
+ 147
+ ],
+ "type": "text",
+ "content": "PENGYUE JIA*, City University of Hong Kong, Hong Kong"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 71,
+ 150,
+ 503,
+ 163
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 150,
+ 503,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 150,
+ 503,
+ 163
+ ],
+ "type": "text",
+ "content": "DERONG XU, City University of Hong Kong, Hong Kong and University of Science and Technology of China, China"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 71,
+ 167,
+ 272,
+ 179
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 167,
+ 272,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 167,
+ 272,
+ 179
+ ],
+ "type": "text",
+ "content": "YI WEN, City University of Hong Kong, Hong Kong"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 71,
+ 183,
+ 473,
+ 196
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 183,
+ 473,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 183,
+ 473,
+ 196
+ ],
+ "type": "text",
+ "content": "YINGYI ZHANG, City University of Hong Kong, Hong Kong and Dalian University of Technology, China"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 71,
+ 199,
+ 315,
+ 213
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 199,
+ 315,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 199,
+ 315,
+ 213
+ ],
+ "type": "text",
+ "content": "WENLIN ZHANG, City University of Hong Kong, Hong Kong"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 71,
+ 216,
+ 307,
+ 228
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 216,
+ 307,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 216,
+ 307,
+ 228
+ ],
+ "type": "text",
+ "content": "WANYU WANG, City University of Hong Kong, Hong Kong"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 71,
+ 232,
+ 268,
+ 243
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 232,
+ 268,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 232,
+ 268,
+ 243
+ ],
+ "type": "text",
+ "content": "YICHAO WANG, Noah's Ark Lab, Huawei, China"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 71,
+ 248,
+ 279,
+ 260
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 248,
+ 279,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 248,
+ 279,
+ 260
+ ],
+ "type": "text",
+ "content": "ZHAOCHENG DU, Noah's Ark Lab, Huawei, China"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 71,
+ 264,
+ 266,
+ 276
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 264,
+ 266,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 264,
+ 266,
+ 276
+ ],
+ "type": "text",
+ "content": "XIANGYANG LI, Noah's Ark Lab, Huawei, China"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 71,
+ 281,
+ 258,
+ 293
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 281,
+ 258,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 281,
+ 258,
+ 293
+ ],
+ "type": "text",
+ "content": "YONG LIU, Noah's Ark Lab, Huawei, Singapore"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 71,
+ 297,
+ 266,
+ 308
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 297,
+ 266,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 297,
+ 266,
+ 308
+ ],
+ "type": "text",
+ "content": "HUIFENG GUO, Noah's Ark Lab, Huawei, China"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 71,
+ 312,
+ 276,
+ 324
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 312,
+ 276,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 312,
+ 276,
+ 324
+ ],
+ "type": "text",
+ "content": "RUIMING TANG†, Noah's Ark Lab, Huawei, China"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 70,
+ 329,
+ 317,
+ 342
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 329,
+ 317,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 329,
+ 317,
+ 342
+ ],
+ "type": "text",
+ "content": "XIANGYU ZHAO†, City University of Hong Kong, Hong Kong"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 69,
+ 350,
+ 503,
+ 473
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 350,
+ 503,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 350,
+ 503,
+ 473
+ ],
+ "type": "text",
+ "content": "Personalization has become an essential capability in modern AI systems, enabling customized interactions that align with individual user preferences, contexts, and goals. Recent research has increasingly concentrated on Retrieval-Augmented Generation (RAG) frameworks and their evolution into more advanced agent-based architectures within personalized settings to enhance user satisfaction. Building on this foundation, this survey systematically examines personalization across the three core stages of RAG: pre-retrieval, retrieval, and generation. Beyond RAG, we further extend its capabilities into the realm of Personalized LLM-based Agents, which enhance traditional RAG systems with agentic functionalities, including user understanding, personalized planning and execution, and dynamic generation. For both personalization in RAG and agent-based personalization, we provide formal definitions, conduct a comprehensive review of recent literature, and summarize key datasets and evaluation metrics. Additionally, we discuss fundamental challenges, limitations, and promising research directions in this evolving field. Relevant papers and resources are continuously updated at the Github Repo| Field | Sub-field | Subsub-field | Papers |
| Pre-retrieval | Query Rewriting | Learning to Personalized Query Rewrite | CLE-QR [60], CGF [38], PEARL [80] |
| LLM to Personalized Query Rewrite | Least-to-Most Prompting [173], ERAGent [112], CoPS [174], Agent4Ranking [61], FIG [22], BASES [99] |
| Query Expansion | Tagging-based query expansion | Gossiple [10], Biancalana and Micarelli [12], SoQuES [15], Zhou et al. [172] |
| Else | Lin and Huang [66], Bender et al. [9], Axiomatic PQEC [79], WE-LM [144], PSQE [14], PQEWC [7] |
| Others | Bobo [33], Kannadasan and Aslanyan [52], PSQE [8] |
| Retrieval | Indexing | PEARL [80], KG-Retriever [21], EMG-RAG [137], PGraphRAG [5] |
| Retrieval | Dense Retrieval | MeMemo [138], RECAP [71], LAPDOG [43], Gu et al. [37], PersonalLM [77], UIA [155], XPERT [125], DPSR [157], RTM [11], Pearl [80], MemPrompt [74], EERRA [23], MALP [160], USER-LLM [84], PER-PCS [120] |
| Sparse Retrieval | OPPU [121], PAG [101], Au et al. [5], UniMS-RAG [128], Deng et al. [29], |
| Prompt-based Retrieval | LAPS [50], UniIMP [140], Shen et al. [111] |
| Others | Salem et al. [103], PersonalITM [65], Zhang et al. [165] |
| Post-retrieval | PersonaRAG [156], Pavliukevich et al. [89], UniMS-RAG [128], Salemi and Zamani [106], Zhang et al. [164], AutoCompressors [24], FIT-RAG [76] |
| Generation | Generation from Explicit Preferences | Direct Prompting | P² [49], Character Profiling [154] OpinionQA [107], Kang et al. [51], Liu et al. [67], Cue-CoT [129], TICL [26] |
| Profile-Augmented Prompting | GPG [158], Richardson et al. [101], ONCE [70], LLMTreeRec [163], KAR [145], Matryoshka [58] |
| Personalized-Prompt Prompting | Li et al. [57], RecGPT [166], PEPLER-D [59], GRAPA [94], SGPT [28], PFCL [152] |
| Generation from Implicit Preferences | Fine-tuning-Based Methods | PLoRA [165], LM-P [142], MiLP [165], OPPU [122], PER-PCS [120], Review-LLM [91], UserIdentifier [78], UserAdapter [171], HYDRA [175], PocketLLM [90], CoGenesis [161] |
| Reinforcement Learning-Based Methods | P-RLHF [62], P-SOUPS [47], PAD [20], REST-PG [104], Salemi et al. [103], RewrimerSIRI [57],Kulkarni et al. [54] |
| From RAG to Agent | Personalized Understanding | In user-profile understanding | Xu et al. [148], Abbasian et al. [2], |
| In agent's role understanding | RoleLLM [139], Character-LLM [110], Wang et al. [134], |
| In agent's user-role joint understanding | SocialBench [18], Dai et al. [27], Ran et al. [96], Wang et al. [126], Tu et al. [123], Necko [153] |
| Personalized Planning and Execution | Memory Management | EMG-RAG [137], Park et al. [87], Abbasian et al. [2], RecAgent [133], TravelPlanner+ [114], PersonalWAB [17], VOYAGER [127], MemoeryLLM [136] |
| Tool and API Calling | VOYAGER [127], Zhang et al. [159], PUMA [17], Wang et al. [126], PenetrativeAI [148], Huang et al. [44], [87], MetaGPT [40], OKR-Agent [169] |
| Personalized Generation | Alignment with User Fact | Character-LLM [110], Wang et al. [135], Dai et al. [27] |
| Alignment with User Preferences | Wang et al. [139], Ran et al. [96], Wang et al. [134], Chen et al. [18] |
",
+ "image_path": "3888cac0ece5531f078c8ec3406df643f228458552c866b9ad79273c9d66181b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 420,
+ 538,
+ 458
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 6,
+ "blocks": [
+ {
+ "bbox": [
+ 107,
+ 420,
+ 538,
+ 443
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 420,
+ 538,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 420,
+ 538,
+ 443
+ ],
+ "type": "text",
+ "content": "- User Historical Content: Implicit personalization derived from user-generated content, such as chat history, emails, reviews, and social media interactions."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 107,
+ 447,
+ 533,
+ 458
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 447,
+ 533,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 447,
+ 533,
+ 458
+ ],
+ "type": "text",
+ "content": "- Persona-Based User Simulation: The use of LLM-based agents to simulate and generate personalized interactions."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 107,
+ 465,
+ 538,
+ 491
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 465,
+ 538,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 465,
+ 538,
+ 491
+ ],
+ "type": "text",
+ "content": "Integrating this personalized information at various stages of the RAG and agent workflows enables dynamic alignment with human preferences, thereby making responses more user-centric and adaptive."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 107,
+ 505,
+ 282,
+ 514
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 505,
+ 282,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 505,
+ 282,
+ 514
+ ],
+ "type": "text",
+ "content": "3 HOW TO ADOPT PERSONALIZATION"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 522,
+ 435,
+ 533
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 522,
+ 435,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 522,
+ 435,
+ 533
+ ],
+ "type": "text",
+ "content": "We define the process of introducing personalization within the RAG pipeline as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 252,
+ 542,
+ 539,
+ 553
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 252,
+ 542,
+ 539,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 252,
+ 542,
+ 539,
+ 553
+ ],
+ "type": "interline_equation",
+ "content": "g = \\mathcal {G} (\\mathcal {R} (Q (q, p), C, p), \\text {p r o m p t}, p, \\theta) \\tag {1}",
+ "image_path": "c5b36e8fc5d85eef240bb82935b05407c1e84bf402d0060b0bb413dc04743f40.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "p"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": " denotes personalized information, and the process unfolds in three steps. In the pre-retrieval phase, query processing "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "(Q)"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": " refines the query "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "q"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": " using personalized information, such as through query rewriting or expansion. During the retrieval phase, the retriever "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "(\\mathcal{R})"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": " leverages "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "p"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": " to fetch relevant documents from the corpus "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "(C)"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": ". Finally, in the generation phase, the retrieved information, combined with "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "p"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": " and structured using the given prompt, id fed into the generator "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "(\\mathcal{G})"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": " with parameter "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "\\theta"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": " to produce the final response "
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "inline_equation",
+ "content": "g"
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": ". It is evident that personalized information directly influences multiple stages of the RAG pipeline. In this survey, we consider the agent system as a specialized application of the RAG framework, where personalization is incorporated in a manner similar to the RAG framework."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 321,
+ 658,
+ 326,
+ 665
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 321,
+ 658,
+ 326,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 321,
+ 658,
+ 326,
+ 665
+ ],
+ "type": "text",
+ "content": "4"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 70,
+ 99,
+ 500,
+ 304
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 70,
+ 99,
+ 500,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 99,
+ 500,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 99,
+ 500,
+ 304
+ ],
+ "type": "image",
+ "image_path": "8f9e5a0db9555c1a4ed96186bb98667b2c4fd2abae37751b9dee8a95b60f65c1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 193,
+ 310,
+ 379,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 193,
+ 310,
+ 379,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 193,
+ 310,
+ 379,
+ 320
+ ],
+ "type": "text",
+ "content": "Fig. 2. Overview of the personalized pre-retrieval stage."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 70,
+ 337,
+ 254,
+ 347
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 337,
+ 254,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 337,
+ 254,
+ 347
+ ],
+ "type": "text",
+ "content": "4 WHERE TO ADOPT PERSONALIZATION"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 70,
+ 354,
+ 146,
+ 364
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 354,
+ 146,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 354,
+ 146,
+ 364
+ ],
+ "type": "text",
+ "content": "4.1 Pre-retrieval"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 69,
+ 372,
+ 503,
+ 424
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 372,
+ 503,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 372,
+ 503,
+ 424
+ ],
+ "type": "text",
+ "content": "4.1.1 Definition. Pre-retrieval is a crucial step in information retrieval systems, where the original user query is enhanced or modified before the retrieval process to improve the relevance and quality of the search results, as shown in Figure 2. This process often incorporates additional contextual or personalized information to better align the query with the user's intent. The process can be formalized as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 261,
+ 432,
+ 503,
+ 444
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 261,
+ 432,
+ 503,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 261,
+ 432,
+ 503,
+ 444
+ ],
+ "type": "interline_equation",
+ "content": "q ^ {*} = Q (q, p) \\tag {2}",
+ "image_path": "faaeeb7d36800aed003df61ec8252e783b23b8a2f81ce0769f1b50247566d8e0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "p"
+ },
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "q"
+ },
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "type": "text",
+ "content": " denote the personalized information and original query, and "
+ },
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "q^{*}"
+ },
+ {
+ "bbox": [
+ 69,
+ 452,
+ 502,
+ 475
+ ],
+ "type": "text",
+ "content": " is the optimized query after query reformulation."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 69,
+ 487,
+ 503,
+ 552
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 487,
+ 503,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 487,
+ 503,
+ 552
+ ],
+ "type": "text",
+ "content": "4.1.2 Query Rewriting. Query rewriting in RAG at the pre-retrieval stage refers to the process of reformulating user queries to enhance retrieval effectiveness by improving relevance, disambiguating intent, or incorporating contextual information before retrieving documents from an external knowledge source. The literature on personalized query rewriting can be broadly classified into two primary categories: (1) Direct Personalized Query Rewriting and (2) Auxiliary Personalized Query Rewriting."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 69,
+ 562,
+ 503,
+ 655
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 562,
+ 503,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 562,
+ 503,
+ 655
+ ],
+ "type": "text",
+ "content": "(1). Direct Personalized Query Rewriting. The first category focuses on personalized query rewriting by using direct models. For example, Cho et al. [25] presents a personalized search-based query rewrite system for conversational AI that addresses user-specific semantic and phonetic errors. Nguyen et al. [82] apply reinforcement learning techniques to improve query rewriting in online e-commerce systems, leveraging distilled LLMs for personalized performance. CLE-QR [60] explores query rewriting in Taobao's search engine to enhance user satisfaction through customized query adaptation. CGF [38] introduces a constrained generation framework that allows for more flexible and personalized query rewriting in conversational AI. Li et al. [57] investigate learning methods to rewrite prompts for personalized"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 284,
+ 658,
+ 288,
+ 665
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 658,
+ 288,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 658,
+ 288,
+ 665
+ ],
+ "type": "text",
+ "content": "5"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 135
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 135
+ ],
+ "type": "text",
+ "content": "text generation, improving the relevance and engagement of AI-generated content. Additionally, PEARL [80] discusses personalizing large language model-based writing assistants through the integration of generation-calibrated retrievers, enhancing AI-generated content."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 147,
+ 541,
+ 296
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 147,
+ 541,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 147,
+ 541,
+ 296
+ ],
+ "type": "text",
+ "content": "(2). Auxiliary Personalized Query Rewriting. The second category emphasizes personalized query rewriting by using auxiliary mechanisms, such as retrieval, reasoning strategies, and external memory. Zhou et al. [173] propose a least-to-most prompting strategy that aids in complex reasoning within LLMs, which can be adapted for personalized text generation. ERAGent [112] enhance retrieval-augmented LLMs to improve personalization, efficiency, and accuracy, indirectly supporting personalized query rewriting for content generation. CoPS [174] integrate LLMs with memory mechanisms to create more personalized search experiences, which also influences content generation through better query understanding. Further, Agent4Ranking [61] employs multi-agent LLMs to perform semantic robust ranking, including personalized query rewriting to improve search rankings. FIG [22] combine graph-based methods with LLMs to query rewrite, improving personalized content generation and conversational interactions. Lastly, BASES [99] employ LLM-based agents to simulate large-scale web search user interactions, contributing to the development of personalized query rewriting strategies for content generation."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 308,
+ 541,
+ 389
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 308,
+ 541,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 308,
+ 541,
+ 389
+ ],
+ "type": "text",
+ "content": "4.1.3 Query Expansion. Query expansion enhances retrieval systems by expanding a user's original query with additional terms, synonyms, or refined structure to better capture intent. This improves the relevance and scope of retrieved documents. Recent advancements in LLMs have reinvigorated this field [46, 48, 132], leveraging their comprehension and generation abilities to expand queries using encoded knowledge or external retrieval, with notable success. Personalized query expansion, a subset, incorporates user-specific data to tailor results, boosting performance and customizing the search experience."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 107,
+ 400,
+ 541,
+ 655
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 7,
+ "blocks": [
+ {
+ "bbox": [
+ 107,
+ 400,
+ 541,
+ 521
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 400,
+ 541,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 400,
+ 541,
+ 521
+ ],
+ "type": "text",
+ "content": "(1). Tagging-based Query Expansion. By 2009, studies began incorporating tagging information to enhance personalized query expansion. For instance, Gossle [10] introduced the TagMap and TagRank algorithms, which dynamically selected tags from personalized networks constructed using the cosine similarity of user-item tag distances, improving recall performance. Similarly, Biancalana and Micarelli [12] recorded user queries and visited URLs, leveraging social bookmarking to extract relevant tags and build a personalized three-dimensional co-occurrence matrix. Based on this, multiple semantically categorized expanded queries were generated to better reflect user interests. Further advancements include SoQuES [15], which integrated tag semantic similarity with social proximity, and a graph-based approach [172] that utilized Tag-Topic models and pseudo-relevance feedback for term weighting, tailoring the expansion process to individual user preferences."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 533,
+ 539,
+ 655
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 533,
+ 539,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 533,
+ 539,
+ 655
+ ],
+ "type": "text",
+ "content": "(2). Else. Apart from tagging-based techniques, early research on Personalized Query Expansion primarily focused on modeling user personalization based on search history [66], social networks, or preferences derived from friendship networks [9]. The Axiomatic PQEC framework [79] formalized expansion rules using both local (user behavior-driven) and social (network-driven) strategies. In 2017, WE-LM [144] advanced this paradigm by modeling multi-relational networks with word embeddings across tag-word relationships, refining associations through affinity graphs. Later, PSQE [14] further improved tagging-based methods using utf-iuf user profiling, integrating a tag similarity graph with user profiles in the online phase to compute expansion terms relevant to user interests in real-time, achieving dynamic personalized expansion. In addition, PQEWC [7] leveraged clustering and contextual word embeddings to optimize query expansions dynamically."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "sub_type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 321,
+ 658,
+ 326,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 321,
+ 658,
+ 326,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 321,
+ 658,
+ 326,
+ 666
+ ],
+ "type": "text",
+ "content": "6"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 5
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 72,
+ 93,
+ 503,
+ 223
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 72,
+ 93,
+ 503,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 93,
+ 503,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 93,
+ 503,
+ 223
+ ],
+ "type": "image",
+ "image_path": "431fadc2fb6b837e6448de02affe85f7a85af23cf964b4cebff641487aab910d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 200,
+ 228,
+ 372,
+ 239
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 228,
+ 372,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 228,
+ 372,
+ 239
+ ],
+ "type": "text",
+ "content": "Fig. 3. Overview of the personalized retrieval stage."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 70,
+ 263,
+ 503,
+ 330
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 263,
+ 503,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 263,
+ 503,
+ 330
+ ],
+ "type": "text",
+ "content": "4.1.4 Others. Besides query rewriting and query expansion, other personalized query-related research focuses on areas like query disambiguation and query auto-completion [116]. Bobo [33] allows users to input contextual terms reflecting their domain knowledge. In 2019, a method [52] applied fastText embeddings from recent queries to rank candidates. In addition, PSQE [8] employed synthetic user profiles from Wikipedia and word2vec embeddings for query disambiguation."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 70,
+ 341,
+ 503,
+ 393
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 341,
+ 503,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 341,
+ 503,
+ 393
+ ],
+ "type": "text",
+ "content": "4.1.5 Discussion. While both query rewriting and query expansion aim to align user input with system understanding to enhance retrieval quality, their roles in personalization differ in fundamental ways. Understanding the distinct operational characteristics and application scenarios of each technique is essential for designing effective personalized retrieval systems. The key takeaways are listed as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 70,
+ 401,
+ 501,
+ 453
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 8,
+ "blocks": [
+ {
+ "bbox": [
+ 70,
+ 401,
+ 501,
+ 426
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 401,
+ 501,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 401,
+ 501,
+ 426
+ ],
+ "type": "text",
+ "content": "- Query rewriting is most beneficial when the original query is ambiguous, underspecified, or misaligned with retrieval intents, particularly in conversational or multi-turn settings."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 70,
+ 428,
+ 501,
+ 453
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 428,
+ 501,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 428,
+ 501,
+ 453
+ ],
+ "type": "text",
+ "content": "- Query expansion is most effective when the original query is relevant but incomplete - i.e., when it needs to be semantically broadened to cover additional relevant concepts."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 70,
+ 468,
+ 132,
+ 478
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 468,
+ 132,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 468,
+ 132,
+ 478
+ ],
+ "type": "text",
+ "content": "4.2 Retrieval"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "text",
+ "content": "4.2.1 Definition. The retrieval process involves finding the most relevant documents "
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "inline_equation",
+ "content": "D^{*}"
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "text",
+ "content": " from a corpus "
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "inline_equation",
+ "content": "C"
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "text",
+ "content": " based on a query "
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "inline_equation",
+ "content": "q^{*}"
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "text",
+ "content": ", as shown in Figure 3. To incorporate personalization, additional user-specific information "
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "inline_equation",
+ "content": "p"
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "text",
+ "content": " is integrated into the retrieval function "
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{R}"
+ },
+ {
+ "bbox": [
+ 70,
+ 485,
+ 501,
+ 538
+ ],
+ "type": "text",
+ "content": ". This allows the retrieval process to tailor the selected documents to align with individual user preferences or contexts, thereby enhancing the relevance and personalization of the generated outputs."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 253,
+ 545,
+ 503,
+ 559
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 253,
+ 545,
+ 503,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 253,
+ 545,
+ 503,
+ 559
+ ],
+ "type": "interline_equation",
+ "content": "D ^ {*} = \\mathcal {R} (q ^ {*}, C, p) \\tag {3}",
+ "image_path": "91c490fa5f8c0af657eecb79cedc896d6a77e119d6fb8493fe9338606a50ca55.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 70,
+ 566,
+ 501,
+ 605
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 566,
+ 501,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 566,
+ 501,
+ 605
+ ],
+ "type": "text",
+ "content": "In the retrieval process, personalization can primarily be introduced by focusing on three steps: indexing, retrieval, and post-retrieval. These steps ensure efficient and accurate retrieval of relevant documents or knowledge, while tailoring the process to individual user preferences. Below, we provide a detailed explanation of each step."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 70,
+ 616,
+ 503,
+ 654
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 616,
+ 503,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 616,
+ 503,
+ 654
+ ],
+ "type": "text",
+ "content": "4.2.2 Indexing. Indexing organizes knowledge base data into a structured format to facilitate efficient retrieval. Within the RAG pipeline, documents are either chunked or entirely encoded into representations before being integrated into searchable systems [30, 117]. Conventional encoding methods employ either sparse encoding techniques (e.g.,"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 289,
+ 665
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 289,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 289,
+ 665
+ ],
+ "type": "text",
+ "content": "7"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 121
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 121
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 121
+ ],
+ "type": "text",
+ "content": "TF-IDF [95], BM25 [102]) or dense encoding approaches leveraging pre-trained models, such as BERT [1], Siamese Encoders [98], or LLM-based encoders [64, 147]."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 124,
+ 541,
+ 217
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 124,
+ 541,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 124,
+ 541,
+ 217
+ ],
+ "type": "text",
+ "content": "To introduce personalization at the indexing stage, PEARL [80] generates user embeddings by encoding personal history data with models like DeBERTa. These embeddings are subsequently clustered to create personalized shared indices. Other approaches integrate knowledge graphs into indexing to enhance retrieval performance. For example, KG-R retriever [21] employs a Hierarchical Index Graph, consisting of a knowledge graph layer and a collaborative document layer, to improve RAG retrieval. EMG-RAG [137] incorporates personalized memory within an editable knowledge graph, enabling dynamic retrieval. Similarly, PGraphRAG [5] leverages user-centric knowledge graphs to enhance personalization in retrieval tasks."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 228,
+ 541,
+ 267
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 228,
+ 541,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 228,
+ 541,
+ 267
+ ],
+ "type": "text",
+ "content": "4.2.3 Retrieval. The Retrieval step matches a user query with the indexed knowledge base to fetch relevant candidates. It can be broadly categorized into four different types: (1) Dense Retrieval, (2) Sparse Retrieval, (3) Prompt-based Retrieval, and (4) Others."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 107,
+ 278,
+ 541,
+ 426
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 278,
+ 541,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 278,
+ 541,
+ 426
+ ],
+ "type": "text",
+ "content": "(1). Dense Retrieval. Dense retrieval methods often use vector embeddings and similarity metrics (e.g., cosine similarity) and achieve personalization by encoding user preferences, context, or interactions into query or document embeddings, enabling tailored results through similarity-based matching. For instance, MeMemo [138] retrieves personalized information by matching user-specific embeddings with document vectors, focusing on private, on-device text generation. Similarly, RECAP [71] and LAPDOG [43] enhance personalized dialogue generation by encoding queries and user profiles as dense vectors and retrieving top-N results, ensuring user-specific context drives the responses. In chatbots, Gu et al. [37] integrates conversational context and user profiles to align retrieved responses with user personas. PersonalM [77] employs group-wise contrastive learning, training its retrieval model to align user queries with domain-specific text fragments, thereby improving personalization. UIA [155] employs dual encoders to retrieve documents tailored to user preferences. XPERT [125] incorporates temporal events and user interactions into embeddings, enabling large-scale retrieval across millions of items."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 429,
+ 542,
+ 522
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 429,
+ 542,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 429,
+ 542,
+ 522
+ ],
+ "type": "text",
+ "content": "Dense retrieval also enhances specific applications like e-commerce, medical assistance, and language models. DPSR [157] and RTM [11] encode user queries and product information to personalize product searches dynamically. Pearl [80] and MemPrompt [74] retrieve personalized content by leveraging historical user data and memory-assisted mechanisms. EERRA [23] uses review embeddings as dense queries for recommendations. In medical assistance, MALP [160] and User-LLM [84] integrate short- and long-term user interactions into embeddings for contextualized, personalized responses. Finally, PER-PCS [120] retrieves relevant information using individual user histories, enhancing the personalization capabilities of large language models."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 533,
+ 542,
+ 655
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 533,
+ 542,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 533,
+ 542,
+ 655
+ ],
+ "type": "text",
+ "content": "(2). Sparse Retrieval. Sparse retrieval methods often rely on term-based matching (e.g., BM25) and apply personalization by assigning higher weights to terms or keywords that are more relevant to the user. OPPU [121] uses the BM25 algorithm to select the k most relevant records from the user's historical data for the current query. Similarly, PAG [101] incorporates user input and profiles to enhance summarization and retrieval, aligning sparse representations with personalization objectives for large language models. Au et al. [5] uses BM25 search algorithms to find entries related to the target user or neighboring users through the graph structure. UniMS-RAG [128] combines sparse and dense retrieval by leveraging multi-source knowledge, such as dialogue context and user images, to refine personalized responses in dialogue systems. Lastly, Deng et al. [29] apply sparse retrieval to support fact-based queries, considering user queries and preferences to enhance answer generation for e-commerce applications."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 321,
+ 658,
+ 327,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 321,
+ 658,
+ 327,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 321,
+ 658,
+ 327,
+ 666
+ ],
+ "type": "text",
+ "content": "8"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 7
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 70,
+ 96,
+ 504,
+ 362
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 4,
+ "blocks": [
+ {
+ "bbox": [
+ 70,
+ 96,
+ 503,
+ 244
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 96,
+ 503,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 96,
+ 503,
+ 244
+ ],
+ "type": "text",
+ "content": "(3). Prompt-based Retrieval. Prompt-based retrieval leverages prompts to guide retrieval from the model or external sources and introduces personalization by crafting user-specific prompts that guide the retrieval process. These prompts may include explicit user preferences, historical interactions, or detailed instructions that reflect the user's unique requirements. By embedding this personalized context directly into the prompt, the retrieval process can dynamically adjust to capture and return results that are most relevant to the user. LAPS [50] focuses on multi-session conversational search by storing user preferences and dialogue context, then using prompts to retrieve relevant information tailored to the user's biases and categories of interest. UniMP [140] employs user interaction histories as input to prompt-based retrieval, enabling personalized recommendations for multi-modal tasks, such as vision-language applications, by aligning prompts with user behavioral data. In contrast, Shen et al. [111] explores the use of LLMs to extract empathy and narrative styles from user-provided stories, but this work primarily focuses on style extraction and does not explicitly involve a retrieval component."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 70,
+ 255,
+ 504,
+ 362
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 255,
+ 504,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 255,
+ 504,
+ 362
+ ],
+ "type": "text",
+ "content": "(4). Others. Reinforcement learning-based retrieval personalizes the process by optimizing retrieval policies based on user feedback, learning user preferences over time to adjust strategies. Salemi et al. [103] combines models like BM25, RbR, and dense retrieval, refining them with reinforcement learning (RL) and knowledge distillation (KD) to adapt to user profiles for personalized outputs. Parameter-based retrieval leverages pre-trained model parameters to implicitly store and retrieve user-specific information, allowing direct retrieval from the model without traditional indices. PersonalTM [65] generates document identifiers (Document IDs) using a Transformer model, encoding query, history, and document relationships into its parameters for personalization. Similarly, Zhang et al. [165] uses parameterized representations to integrate user queries and histories, tailoring responses to individual preferences."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 70,
+ 373,
+ 502,
+ 411
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 373,
+ 502,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 373,
+ 502,
+ 411
+ ],
+ "type": "text",
+ "content": "4.2.4 Post-retrieval. Current Post-Retrieval methods primarily focus on refining retrieved documents or responses to improve relevance and coherence, current methodologies could be categorized into three parts (1) Re-ranking, (2) Summarization, and (3) Compression."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 70,
+ 422,
+ 504,
+ 654
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 9,
+ "blocks": [
+ {
+ "bbox": [
+ 70,
+ 422,
+ 504,
+ 529
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 422,
+ 504,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 422,
+ 504,
+ 529
+ ],
+ "type": "text",
+ "content": "(1). Re-ranking. Re-ranking enhances personalized content generation by prioritizing more relevant documents at the top. PersonaRAG [156] extends RAG by integrating user-centric agents, such as the Live Session Agent and the Document Ranking Agent, to refine document ranking and improve overall performance. Pavliukevich et al. [89] propose a cross-encoder BERT model for re-ranking external knowledge within a personalized context. UniMS-RAG [128] introduces a scoring mechanism that evaluates retrieved documents and outputs by optimizing the retriever. Besides, it includes an evidence attention mask, enabling re-ranking during inference and applying it to personalized datasets. Salemi and Zamani [106] present an iterative approach to optimizing ranking results based on the expectation-maximization algorithm, with performance validated in personalized scenarios."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 70,
+ 540,
+ 502,
+ 578
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 540,
+ 502,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 540,
+ 502,
+ 578
+ ],
+ "type": "text",
+ "content": "(2). Summarization. Summarization refers to the process of summarizing retrieved information to enhance performance. For instance, Zhang et al. [164] introduced a role-playing agent system to summarize retrieved history in order to improve the final Personalized Opinion Summarization process."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 70,
+ 589,
+ 503,
+ 654
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 589,
+ 503,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 589,
+ 503,
+ 654
+ ],
+ "type": "text",
+ "content": "(3). Compression. Compression involves condensing embeddings or retrieved content to enhance efficiency and effectiveness. Approaches like AutoCompressor [24] compress contextual embeddings into shorter semantic representations, and FIT-RAG [76] introduces a self-knowledge recognizer along with a sub-document-level token reduction mechanism to minimize tokens within RAG pipeline. However, few studies have specifically explored personalized fields, highlighting a promising direction for future research."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "sub_type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 284,
+ 658,
+ 289,
+ 665
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 284,
+ 658,
+ 289,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 284,
+ 658,
+ 289,
+ 665
+ ],
+ "type": "text",
+ "content": "9"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 8
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 541,
+ 342
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 541,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 541,
+ 342
+ ],
+ "type": "text",
+ "content": "4.2.5 Discussion. Indexing, retrieval, and post-retrieval methods each play a critical role in ensuring efficient and personalized information processing, with specific applications and trade-offs. Indexing focuses on organizing knowledge bases for efficient retrieval, using techniques such as sparse encoding methods like TF-IDF and BM25, which are efficient but limited in understanding semantics, and dense encoding methods like BERT and DeBERTa, which provide better semantic understanding but require significant computational resources. These methods are widely used in tasks like question answering and personalized recommendation systems. Retrieval involves matching user queries with relevant documents and can be categorized into dense retrieval, which provides high semantic understanding and personalization but is computationally expensive; sparse retrieval, which is efficient and interpretable but less capable of handling semantics; prompt-based retrieval, which is highly flexible and adaptable to user needs but requires careful engineering of prompts; and advanced methods like reinforcement learning-based approaches, which dynamically adapt to user feedback but are complex to implement. This step is essential in applications like personalized dialogue systems, search engines, and e-commerce. Post-retrieval methods refine retrieved results to enhance relevance and coherence through re-ranking, which improves personalization and prioritizes relevant content but increases computational overhead; summarization, which simplifies complex information for better user understanding but risks losing critical details; and compression, which reduces computational costs by condensing information but remains underexplored in personalized contexts. Together, these methods provide a comprehensive pipeline for delivering efficient, relevant, and personalized outputs, balancing their strengths in semantic understanding, relevance, and flexibility with challenges related to computational costs and implementation complexity."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 354,
+ 177,
+ 363
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 354,
+ 177,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 354,
+ 177,
+ 363
+ ],
+ "type": "text",
+ "content": "4.3 Generation"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "text",
+ "content": "4.3.1 Definition. Personalized generation incorporates user-specific retrieved documents "
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "inline_equation",
+ "content": "D^{*}"
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "text",
+ "content": ", task-specific prompt prompt, and user preference information "
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "inline_equation",
+ "content": "p"
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "text",
+ "content": " via the generator "
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{G}"
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "text",
+ "content": " parameterized by "
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "inline_equation",
+ "content": "\\theta"
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "text",
+ "content": " to produce tailored content "
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "inline_equation",
+ "content": "g^{*}"
+ },
+ {
+ "bbox": [
+ 107,
+ 370,
+ 541,
+ 410
+ ],
+ "type": "text",
+ "content": " aligned with individual preference, where the flow is shown in Figure 4. The generation process can be formulated as"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 274,
+ 417,
+ 539,
+ 430
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 274,
+ 417,
+ 539,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 274,
+ 417,
+ 539,
+ 430
+ ],
+ "type": "interline_equation",
+ "content": "g ^ {*} = \\mathcal {G} (D ^ {*}, \\text {p r o m p t}, p, \\theta). \\tag {4}",
+ "image_path": "7122c71aefb1637d6b1b8ecf544934b92473bd4011b205e6f3793fb7593c1f2c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "type": "text",
+ "content": "Personalized generation can be achieved by incorporating explicit and implicit preferences. Explicit preference-driven methodologies utilize direct input signals (e.g., "
+ },
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "type": "inline_equation",
+ "content": "D^{*}"
+ },
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "type": "text",
+ "content": ", prompt, and "
+ },
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "type": "inline_equation",
+ "content": "p"
+ },
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "type": "text",
+ "content": "), to tailor outputs to specific user preferences. Conversely, implicit preference-encoded approaches embed personalized information within the parameters "
+ },
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "type": "inline_equation",
+ "content": "\\theta"
+ },
+ {
+ "bbox": [
+ 107,
+ 437,
+ 541,
+ 503
+ ],
+ "type": "text",
+ "content": " of the generator model, during training, thereby facilitating preference alignment without the necessity for explicit runtime inputs."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 513,
+ 541,
+ 594
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 513,
+ 541,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 513,
+ 541,
+ 594
+ ],
+ "type": "text",
+ "content": "4.3.2 Generation from Explicit Preferences. Integrating explicit preferences into LLMs facilitates personalized content generation. Explicit preference information encompasses user demographic information (e.g., age, occupation, gender, location), user behavior sequences (reflecting historical behavioral patterns), and user historical output texts (capturing writing style and tone preferences). The injection of explicit preferences for personalized generation can be categorized into three types: (1) Direct-integrated Prompting, (2) Summary-augmented Prompting, and (3) Adaptive Prompting."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 107,
+ 603,
+ 541,
+ 656
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 603,
+ 541,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 603,
+ 541,
+ 656
+ ],
+ "type": "text",
+ "content": "(1). Direct-integrated Prompting. Integrating user explicit preferences into language models through prompting enables the prediction of users' intent and behavioral patterns, facilitating personalized content generation. For instance, "
+ },
+ {
+ "bbox": [
+ 107,
+ 603,
+ 541,
+ 656
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{P}^2"
+ },
+ {
+ "bbox": [
+ 107,
+ 603,
+ 541,
+ 656
+ ],
+ "type": "text",
+ "content": " [49], Character Profiling [154], and OpinionQA [107] integrate personalized data into LLMs through prompting for role-playing task, thereby aligning the model's responses with specified user profiles. Kang et al. [51] and Liu et al. [67]"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "text",
+ "content": "10"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 9
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 83,
+ 95,
+ 491,
+ 232
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 83,
+ 95,
+ 491,
+ 232
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 83,
+ 95,
+ 491,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 83,
+ 95,
+ 491,
+ 232
+ ],
+ "type": "image",
+ "image_path": "684289db43c065be977860260a5ca5599fa5f676c9663e41befff2d9bcc9c089.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 195,
+ 245,
+ 376,
+ 256
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 195,
+ 245,
+ 376,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 195,
+ 245,
+ 376,
+ 256
+ ],
+ "type": "text",
+ "content": "Fig. 4. Overview of the personalized generation stage."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 70,
+ 277,
+ 503,
+ 344
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 277,
+ 503,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 277,
+ 503,
+ 344
+ ],
+ "type": "text",
+ "content": "integrate interaction histories into LLMs via prompting to predict user rating for candidate items. Cue-CoT [129] employs chain-of-thought reasoning to infer user needs from contextual cues, enabling personalized responses to in-depth dialogue questions. Additionally, TICL [26] proposes a trial-and-error framework that critiques initial LLM-generated responses, derives explanations and integrates these negative examples into prompts to improve personalization alignment."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 69,
+ 354,
+ 504,
+ 606
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 7,
+ "blocks": [
+ {
+ "bbox": [
+ 69,
+ 354,
+ 504,
+ 475
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 354,
+ 504,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 354,
+ 504,
+ 475
+ ],
+ "type": "text",
+ "content": "(2). Summary-augmented Prompting. Direct integration of personalized information via prompting struggles with ambiguous intent signals: Lengthy interaction histories introduce noise that obscures critical behavioral patterns [69], while sparse behavioral data lacks sufficient context for LLMs to derive meaningful user preferences. To address these issues, recent approaches focus on summarizing user personalized intents and integrating them into prompts. For instance, GPG [158] extracts key user habits and preferences from personal contexts, enabling fine-grained personalization. Similarly, LLMs are employed to generate task-specific summaries of user preferences, enhancing retrieval-augmented personalized generation capabilities [101]. In recommendation systems, ONCE [70], LLMTreeRec [163], and KAR [145] leverage historical user-item interactions to summarize user preferences. Furthermore, Matryoshka [58] generates user preference summaries by dynamically retrieving and synthesizing historical data."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 69,
+ 485,
+ 504,
+ 606
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 485,
+ 504,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 485,
+ 504,
+ 606
+ ],
+ "type": "text",
+ "content": "(3). Adaptive Prompting. Manually designing personalized prompts demands both expert knowledge and significant labor, motivating the development of automated methods for personalized prompt generation. For example, Li et al. [57] trains a personalized prompt rewriter via supervised and reinforcement learning. RecGPT [166] and PEPLER-D [59] leverage prompt tuning to generate personalized prompts, enhancing sequential and explainable recommendations, respectively. GRAPA [94] integrates semantic and collaborative signals from user-item interaction graphs with graph neural networks to generate context-aware personalized prompts. SGPT [28] employs prompt tuning to jointly model common and group-specific patterns, bridging generalized and personalized federated learning paradigms. Furthermore, PFCL [152] achieves multi-granularity human preference modeling: coarse-grained prompts distill shared knowledge, while fine-grained prompts adapt to individual user characteristics."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 70,
+ 616,
+ 504,
+ 655
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 616,
+ 504,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 616,
+ 504,
+ 655
+ ],
+ "type": "text",
+ "content": "4.3.3 Generation from Implicit Preferences. Unlike explicit preference modeling, which captures user preferences through textual input, implicit preference-based methods incorporate personalization through internal parameters. This personalization is achieved either through Parameter-Efficient Fine-tuning (PEFT) techniques, such as LoRA [42],"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "text",
+ "content": "11"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 10
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 541,
+ 122
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 541,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 541,
+ 122
+ ],
+ "type": "text",
+ "content": "or reinforcement learning-based approaches for preference alignment [20, 57]. Based on these strategies, we classify existing methods into two categories: (1) Fine-tuning-Based Methods and (2) Reinforcement Learning-Based Methods."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 132,
+ 541,
+ 253
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 132,
+ 541,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 132,
+ 541,
+ 253
+ ],
+ "type": "text",
+ "content": "(1). Fine-tuning Based Methods. For fine-tuning methods, LoRA is the most widely adopted since it is resource-efficient and enables rapid adaptation without compromising model performance. PLoRA [165] introduces a personalized knowledge integration framework that combines task-specific LoRA with user-specific knowledge. Similarly, LM-P [142] personalizes information via LoRA by incorporating User ID as a personalization factor. MiLP [165] employs Bayesian optimization to determine the optimal personalization injection configuration, including LoRA settings, to effectively capture and utilize user-specific information. OPPU [122] and PER-PCS [120] follow a similar approach, leveraging user history data for fine-tuning LoRA-based personalization. However, PER-PCS differs by incorporating a gating module that selects the appropriate LoRA, enabling fine-grained personalization. Additionally, Review-LLM [91] integrates LoRA for supervised fine-tuning in the task of personalized review generation."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 255,
+ 541,
+ 335
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 255,
+ 541,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 255,
+ 541,
+ 335
+ ],
+ "type": "text",
+ "content": "Beyond LoRA-based approaches, alternative pipelines have been proposed for personalized generation. UserIdentifier [78] introduces a user-specific identifier, significantly reducing training costs while enhancing personalized demonstration. UserAdapter [171] proposes user-independent prefix embeddings, leveraging prefix tuning for personalization. Meanwhile, HYDRA [175] achieves implicit personalization by training user-specific headers. Recently, researchers have also explored fine-tuning personalized model on edge devices [90] and collaborative learning between small and large language models to enable more personalized generation [161]."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 107,
+ 346,
+ 541,
+ 550
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 346,
+ 541,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 346,
+ 541,
+ 550
+ ],
+ "type": "text",
+ "content": "(2). Reinforcement Learning Based Methods. Apart from fine-tuning based methods, recent research has explored reinforcement learning based techniques to personalize text generation by aligning outputs with user preferences. P-RLHF [62] has been proposed to jointly learn a user-specific and reward model to enable text generation that aligns with a user's styles or criteria. P-SOUPS [47] models multiple user preferences as a Multi-Objective Reinforcement Learning (MORL) problem, decomposing preferences into multiple dimensions, each trained independently. PAD [20] aligns text generation with human preferences during inference by utilizing token-level personalized rewards to guide the decoding process. REST-PG [104] introduces a framework that trains large language models to reason over personal data during response generation. This approach first generates reasoning paths to enhance the LLM's reasoning ability and then employs Expectation-Maximization Reinforced Self-Training to iteratively refine the model based on its high-reward outputs. Additionally, Salemi et al. [103] incorporate reinforcement learning into the RAG pipeline to improve retrieval accuracy, thereby enhancing the personalization of generated content. Other applications include RewriterSIRI [57], which has been introduced to generate text via RL-based personalized prompt rewriting using API-based LLMs, and Kulkarni et al. [54], who explore the use of reinforcement learning to optimize RAG for improving the relevance and coherence of chatbot responses in specialized domains, ultimately enhancing user satisfaction and engagement."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 561,
+ 541,
+ 654
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 561,
+ 541,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 561,
+ 541,
+ 654
+ ],
+ "type": "text",
+ "content": "4.3.4 Discussion. Personalized generation can be adopted via both explicit and implicit preference injection, yet they exhibit distinct characteristics that make them suitable for different scenarios. In explicit preference-based generation, personalization is clearly defined through user profile descriptions, contextual information, and similar inputs, which are incorporated into generators via prompts. A key advantage of this approach is explainability, as the personalized information is explicitly provided and easily traceable. Despite leveraging provided preferences and internal knowledge, explicit preference injection's personalization is constrained by model capabilities and irrelevant information interference. In contrast, implicit preference-based generation internalizes personalized information into"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 665
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 665
+ ],
+ "type": "text",
+ "content": "12"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 11
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 75,
+ 94,
+ 500,
+ 388
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 75,
+ 94,
+ 500,
+ 388
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 94,
+ 500,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 94,
+ 500,
+ 388
+ ],
+ "type": "image",
+ "image_path": "c1f227eab52914058e52cd6428ef57f5faa6d829581489e901866e44230f1223.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 160,
+ 397,
+ 414,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 160,
+ 397,
+ 414,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 397,
+ 414,
+ 408
+ ],
+ "type": "text",
+ "content": "Fig. 5. Overview of transition from personalized RAG to personalized agent."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 70,
+ 440,
+ 503,
+ 492
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 440,
+ 503,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 440,
+ 503,
+ 492
+ ],
+ "type": "text",
+ "content": "the generator's parameters through scene-specific personalized data, thereby adapting the model for more fine-grained personalization. However, these methods typically incur substantial training and computational costs, as they require fine-tuning the generator's internal parameters. Therefore, selecting between these approaches should be guided by the specific application scenario and resource constraints."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 70,
+ 509,
+ 174,
+ 520
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 509,
+ 174,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 509,
+ 174,
+ 520
+ ],
+ "type": "text",
+ "content": "4.4 From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 69,
+ 525,
+ 503,
+ 593
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 525,
+ 503,
+ 593
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 525,
+ 503,
+ 593
+ ],
+ "type": "text",
+ "content": "4.4.1 Definition. A personalized LLM-based agent is a system designed to dynamically incorporate user context, memory, and external tools or APIs to support highly personalized and goal-oriented interactions [19, 45, 146], and solve problems in a goal-oriented manner [63, 113]. From the previously introduced stages of RAG, we observe that the evolution of personalized RAG reveals a structural convergence with agent architectures. We analyze them from three key perspectives:"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 70,
+ 601,
+ 503,
+ 654
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 601,
+ 503,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 601,
+ 503,
+ 654
+ ],
+ "type": "text",
+ "content": "- Personalized Understanding: This phase within the agent parallels the query understanding and rewriting process of RAG as outlined in Section 4.1. However, it extends beyond static semantic parsing by incorporating dynamic user profiling [139] and role modeling [110]. This integration enables the agent to dynamically align interactions with implicit user preferences, facilitating personalized responses and task-specific adaptations [96]."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "text",
+ "content": "13"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 12
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 232
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 4,
+ "blocks": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 538,
+ 162
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 538,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 538,
+ 162
+ ],
+ "type": "text",
+ "content": "- Personalized Planning and Execution: This phase in agents mirrors RAG's retrieval process in Section 4.2 yet it advances beyond static document retrieval by incorporating real-time memory management [87] and sophisticated tool and API calling [127]. This approach ensures the dynamic alignment of external knowledge with personalized constraints, such as integrating medical history in healthcare agents [2], to deliver context-aware and user-specific outcomes."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 165,
+ 539,
+ 232
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 165,
+ 539,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 165,
+ 539,
+ 232
+ ],
+ "type": "text",
+ "content": "- Personalized Generation: This phase in agents mirrors RAG's generative process in Section 4.3 but transcends static template-based generation by integrating user preference and fact alignment. Agents dynamically enforce user preferences and ensure fact consistency through role-specific mechanisms (e.g., social adaptability in conversational agents [2]), enabling outputs to evolve in harmony with personalized and situational constraints rather than relying solely on predefined generative frameworks."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 107,
+ 238,
+ 539,
+ 304
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 238,
+ 539,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 238,
+ 539,
+ 304
+ ],
+ "type": "text",
+ "content": "In general we frame agent architectures as \"personalized RAG++\", where persistent memory [137] replaces static indexes, and tool APIs [17] serve as dynamic knowledge connectors, enabling complicated, human-aligned interactions beyond one-shot retrieval, as shown in Figure 5. This progression highlights that as RAG systems incorporate deeper personalization—requiring user-state tracking, adaptive tool usage, and context-aware generation, they inherently adopt agent-like capabilities."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 316,
+ 539,
+ 381
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 316,
+ 539,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 316,
+ 539,
+ 381
+ ],
+ "type": "text",
+ "content": "4.4.2 Personalized Understanding. Personalized understanding refers to an agent's ability to accurately interpret user inputs by integrating user intent recognition and contextual analysis. This process ensures interactions that are both meaningful and contextually appropriate. The rationale behind this classification lies in its capacity to address three core aspects of understanding: recognizing user intent, analyzing context, and leveraging user profiles. Each of these aspects plays a distinct role in improving the agent's performance."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 392,
+ 539,
+ 654
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 10,
+ "blocks": [
+ {
+ "bbox": [
+ 107,
+ 392,
+ 538,
+ 486
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 392,
+ 538,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 392,
+ 538,
+ 486
+ ],
+ "type": "text",
+ "content": "(1). User-profile Understanding. In user-profile understanding, an agent's personalized ability primarily depends on its capacity to accurately model and understand the user's preferences, context, and intentions. Xu et al. [148] proposes a framework in which LLMs are designed to understand the physical world, thereby facilitating a deeper connection between the agent and its environment, which is essential for accurate task execution. Abbasian et al. [2] further expands this understanding by emphasizing the importance of personalization in health agents, where the user's profile directly influences the behavior and decisions of the agent. This user understanding is foundational to ensuring that the AI agent performs tasks in a way that aligns with individual user needs."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 107,
+ 498,
+ 538,
+ 605
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 498,
+ 538,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 498,
+ 538,
+ 605
+ ],
+ "type": "text",
+ "content": "(2). Role Understanding. In agent's role understanding, the role of the agent within these environments is also crucial. Recent studies focus on enhancing role-playing capabilities within LLMs. Wang et al. [139] introduce RoleLLM, a benchmark that aims to elicit and refine the role-playing abilities of LLMs, demonstrating how role understanding influences agent performance in conversational tasks. Similarly, Shao et al. [110] present Character-LLM, a trainable agent framework for role-playing, which tailors its responses based on predefined roles. Wang et al. [134] introduce a method for evaluating personality fidelity in role-playing agents through psychological interviews, aiming to enhance the realism and consistency of AI-driven characters. This role understanding allows for more contextually appropriate interactions, increasing the relevance and utility of AI agents across various applications."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 616,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 616,
+ 539,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 616,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": "(3). User-role Joint Understanding. In agent's user-role joint understanding, the intersection of user and role understanding is explored through frameworks that evaluate and enhance the social and personality aspects of LLMs. SocialBench Chen et al. [18] provides a sociality evaluation framework for role-playing agents. Dai et al. [27], and"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "sub_type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "text",
+ "content": "14"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 13
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 70,
+ 96,
+ 503,
+ 162
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 96,
+ 503,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 96,
+ 503,
+ 162
+ ],
+ "type": "text",
+ "content": "[96] extend this by incorporating multi-modal data and personality-indicative information, respectively, which allows agents to better adapt to both user and role understanding in dynamic environments. Furthermore, Wang et al. [126] offers a perspective on how role and environment understanding can improve user experience. Tu et al. [123] contribute by providing a benchmark specifically for evaluating role-playing agents in the Chinese context, adding a cultural dimension to role understanding. Finally, Neeko [153] further advances role-based interactions."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 70,
+ 175,
+ 504,
+ 242
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 175,
+ 504,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 175,
+ 504,
+ 242
+ ],
+ "type": "text",
+ "content": "4.4.3 Personalized Planning and Execution. Personalized planning and execution refer to the process of designing and implementing strategies or actions that are specifically tailored to an individual's unique context, and goals [44, 87, 114, 159]. It requires agents to dynamically integrate long-term memory, real-time reasoning, and external tool utilization [40, 41, 169], as demonstrated in healthcare decision support [2] and travel planning scenarios [17]. We analyze two fundamental components that enable this personalization in the following."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 70,
+ 253,
+ 504,
+ 508
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 6,
+ "blocks": [
+ {
+ "bbox": [
+ 70,
+ 253,
+ 504,
+ 374
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 253,
+ 504,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 253,
+ 504,
+ 374
+ ],
+ "type": "text",
+ "content": "(1). Memory Management. Effective memory systems allow agents to integrate users' historical preferences, behavioral patterns, and contextual habits, enhancing their ability to make planning and tailor interactions to user-specific needs [17, 127, 136]. The EMG-RAG framework [137] combines editable memory graphs with retrieval-augmented generation to maintain dynamic user profiles, while Park et al. [87] implements memory streams and periodic reflection mechanisms to simulate human-like behavior. In healthcare applications, Abbasian et al. [2] integrates multimodal user data through specialized memory modules to optimize treatment recommendations. For recommendation systems, RecAgent [133] employs hierarchical memory structures to model user interaction patterns across multiple domains. Recent advances like TravelPlanner+ [114] demonstrate how memory-augmented LLMs achieve higher relevance in personalized itinerary generation compared to generic planners."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 70,
+ 386,
+ 504,
+ 508
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 386,
+ 504,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 386,
+ 504,
+ 508
+ ],
+ "type": "text",
+ "content": "(2). Tool and API Calling. The integration of external tools expands agents' capabilities beyond pure linguistic reasoning, enabling agents to interact with users and perform personalized tasks [17, 126, 127, 148, 159]. For instance, VOYAGER [127] establishes a paradigm for lifelong skill acquisition through automatic API curriculum learning and skill library construction. In robotics, Zhang et al. [159] develops a bootstrapping framework where LLMs guide robots in tool-mediated skill discovery, enabling a high success rate in novel object manipulation tasks. The PUMA framework [17] demonstrates how personalized web agents can achieve performance gains in e-commerce tasks through adaptive API orchestration. For mobile interaction, Wang et al. [126] implements few-shot tool learning to handle diverse UI operations with minimal training data. These approaches highlight the importance of tool grounding mechanisms [44] that translate linguistic plans into executable API sequences while maintaining personalization constraints."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 70,
+ 510,
+ 504,
+ 589
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 510,
+ 504,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 510,
+ 504,
+ 589
+ ],
+ "type": "text",
+ "content": "This synthesis highlights that modern agent systems achieve enhanced personalization through two primary strategies: 1) Memory-augmented architectures, which leverage editable memory graphs [137], reflection mechanisms [87], and hierarchical memory structures [133] to dynamically adapt to user preferences across various domains; and 2) Tool and API integration, which expand agent capabilities by balancing generalization with specialization. Future work may explore improving the contextual relevance and adaptability of memory systems while optimizing real-time tool interaction for seamless task execution."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 70,
+ 601,
+ 504,
+ 654
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 601,
+ 504,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 601,
+ 504,
+ 654
+ ],
+ "type": "text",
+ "content": "4.4.4 Personalized Generation. Based on the foundation of personalized planning and execution mechanisms, which enable agents to adapt strategies to user-specific contexts [44, 159], the next critical concern lies in personalized generation. This capability ensures that generated outputs not only align with factual correctness but also resonate with users' unique preferences, personality traits, and situational needs. Personalized generation bridges the gap between"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "text",
+ "content": "15"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 14
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 122
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 122
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 96,
+ 539,
+ 122
+ ],
+ "type": "text",
+ "content": "adaptive reasoning and human-aligned outcomes, allowing agents to produce contextually relevant and emotionally appropriate responses."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 140,
+ 542,
+ 348
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 5,
+ "blocks": [
+ {
+ "bbox": [
+ 107,
+ 140,
+ 541,
+ 262
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 140,
+ 541,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 140,
+ 541,
+ 262
+ ],
+ "type": "text",
+ "content": "(1). Alignment with User Fact. Alignment with User Fact emphasizes the accuracy, consistency, and factual grounding of personalized responses, ensuring they remain trustworthy across diverse user interactions. This is particularly challenging in personalized agents, where maintaining character authenticity while avoiding hallucinations requires balancing creativity with factual adherence. Recent advances address these challenges through improved training frameworks and evaluation metrics. For instance, Character-LLM [110] integrates memory-augmented architectures to reduce hallucinations while preserving character-specific traits. Wang et al. [135] investigate quantization effects on personality consistency in edge-deployed agents and stabilize outputs under computational constraints. Dai et al. [27] ensures multimodal consistency (text-image) in role-playing. These works highlight the importance of architectural innovations and rigorous evaluation in achieving reliability."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 280,
+ 542,
+ 348
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 280,
+ 542,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 280,
+ 542,
+ 348
+ ],
+ "type": "text",
+ "content": "(2). Alignment with User Preferences. Alignment with user preferences ensures that generated outputs reflect individualized personalities, values, and interaction styles. This requires agents to dynamically interpret implicit user cues and adapt responses accordingly. Wang et al. [139] benchmarks role-specific alignment. Ran et al. [96] improves personality fidelity via psychological scale datasets. Wang et al. [134] quantifies alignment via psychological interviews. Chen et al. [18] evaluates social adaptability in conversations."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 107,
+ 365,
+ 539,
+ 390
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 365,
+ 539,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 365,
+ 539,
+ 390
+ ],
+ "type": "text",
+ "content": "4.4.5 Discussion. The architectural evolution from RAG to personalized agents introduces significant advancements in human-AI interaction but also surfaces critical challenges that warrant further investigation."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 392,
+ 541,
+ 486
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 392,
+ 541,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 392,
+ 541,
+ 486
+ ],
+ "type": "text",
+ "content": "Personalized Understanding, while enabling interpretation of user intent and context, faces limitations in real-time adaptability and generalization. Current approaches like RoleLLM [139] and Character-LLM [110] demonstrate robust role-specific comprehension but struggle with dynamic user state tracking, particularly when handling evolving preferences or multi-session interactions. Furthermore, cultural specificity in benchmarks like CharacterEval [123] reveals gaps in global applicability, as agents trained on region-specific data often fail to generalize across diverse sociocultural contexts. Future work could explore hybrid architectures that combine continuous learning mechanisms with privacy-preserving federated learning to address these adaptability constraints while maintaining user trust."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 107,
+ 489,
+ 541,
+ 583
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 489,
+ 541,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 489,
+ 541,
+ 583
+ ],
+ "type": "text",
+ "content": "Personalized Planning and Execution, achieves remarkable task specialization through memory management and tool integration, yet suffers from scalability issues in complex environments. While frameworks like EMG-RAG [137] and VOYAGER [127] effectively manage user-specific constraints, their reliance on predefined API taxonomies limits emergent tool discovery in novel scenarios. The \"cold-start\" problem persists in domains requiring rapid skill acquisition, as seen in healthcare applications [2], where delayed API responses can compromise decision-making efficacy. A promising direction involves developing meta-reasoning architectures that dynamically prioritize memory recall versus tool invocation based on situational urgency and confidence thresholds."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 585,
+ 541,
+ 652
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 585,
+ 541,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 585,
+ 541,
+ 652
+ ],
+ "type": "text",
+ "content": "Personalized Generation balances factual accuracy with preference alignment but risks over-fitting, where excessive finetuning to user profiles may reinforce cognitive biases. Techniques address surface-level alignment but lack mechanisms for ethical boundary detection. For instance, agents might inadvertently propagate harmful stereotypes when mirroring user preferences without critical oversight. Future systems could integrate value-aligned reinforcement learning with human-in-the-loop validation to preserve authenticity while preventing detrimental customization."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "text",
+ "content": "16"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 15
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 73,
+ 110,
+ 504,
+ 387
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 181,
+ 94,
+ 391,
+ 104
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 94,
+ 391,
+ 104
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 94,
+ 391,
+ 104
+ ],
+ "type": "text",
+ "content": "Table 2. Datasets and metrics for personalized RAG and Agent."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 73,
+ 110,
+ 504,
+ 387
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 73,
+ 110,
+ 504,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 73,
+ 110,
+ 504,
+ 387
+ ],
+ "type": "table",
+ "html": "| Field | Metrics Category | Metrics | Datasets |
| Pre-retrieval | Textual Quality | BLEU, ROUGE, EM | Avocado Research Email Collection [57, 85], Amazon review [57, 83], Reddit comments[57, 118], Amazon ESCI dataset[82, 97], PIP |
| Information Retrieval | MAP, MRR, NDCG, Precision, Recall, RBP | AOL[88, 174], WARRIORS[99], Personalized Results Re-Ranking benchmark [6], delicio.us [9, 15, 144, 172], Flickr [9, 108], CiteULike [10, 14], LRDP [12], Delicious [141], Bibsonomy [79], Wikipedia [8, 33] |
| Classification | Accuracy, Macro-F1 | SCAN [56, 173], AITA WORKSM[53, 80], Robust04 [61] |
| Others | XEntropy, PMS, Image-Align, PQEC, ProfOverlap | Amazon ESCI dataset[82, 97], PIP, Bibsonomy [79] |
| Retrieval | Textual Quality | BLEU, ROUGE, Dis, PPL | TOPDIAL [130], Pchatbot [93], DuLemon [150] |
| Information Retrieval | Recall, MRR, Precision, F1 | LiveChat [34], Pchatbot [93], DuLemon [150] |
| Classification | Accuracy, Succ | TOPDIAL [130], PersonalityEvid [119], DuLemon [150], PersonalityEdit [75] |
| Others | Fluency, Coherence, Plausibility, ES, DD, TPEI, PAE | PersonalityEvid [119], PersonalityEdit [75] |
| Generation | Textual Quality | BLEU, ROUGE, Dis, PPL, METEOR | LaMP [105], Long LaMP [55], Dulemon [150], PGraphRAG [5], AmazonQA/Products [29], Reddit [170], MedicalDialogue [162] |
| Classification | Accuracy, F1, Persona F1 | LaMP [105], Long LaMP [55], Dulemon [150], AmazonQA/Products [29], Reddit [170], MedicalDialogue [162] |
| Regression | MAE, RMSE | LaMP [105], Long LaMP [55], PGraphRAG [5] |
| Others | Fluency, Mean Success Rate, Median Relative Improvements | Personalized-Gen [3] |
| Agent | Textual Quality | BLEU, ROUGE, METEOR, CIDer, EM, Fluency, Coherence, Instruction Adherence, Consistency related metrics | RICO [126], RoleBench [139], Shao et al. [110], Socialbench [18], MMRole-Data [27], ROLEPERSONALITY [96], ChatHarui [134], Character-LLM-Data [153], Knowledge Behind Persona [41], Wang et al. [137], Wang et al. [135], Zheng et al. [169] |
| Information Retrieval | Recall, F1, Precision | Knowledge Behind Persona [41] |
| Classification | Accuracy, Failure Rate, Classification Accuracy, Preference Rate, Correctness | MIT-BIH Arrhythmia Database [148], VirtualHome [44], Socialbench [18], ARC [100], AGIEval [100], HellaSwag [100], MedMCQA [100], AQUA-RAT [100], LogiQA [100], LSAT-AR [100], LSAT-LR [100], LSAT-RC [100], SAT-English [100], SAT-Math [100], PersonalWAB [17], TravelPlanner+ [114] |
| Others | Pass@k, Executability, Productivity, Plausibility of the Story | Hong et al. [40], Zheng et al. [169] |
",
+ "image_path": "d86abcf5e5ec9f4d1a1e2cf178e6e2a2411500ea7111c2031609d3b0586d7569.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 70,
+ 404,
+ 207,
+ 414
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 404,
+ 207,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 404,
+ 207,
+ 414
+ ],
+ "type": "text",
+ "content": "5 EVALUATION AND DATASET"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 70,
+ 421,
+ 503,
+ 570
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 421,
+ 503,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 421,
+ 503,
+ 570
+ ],
+ "type": "text",
+ "content": "In the evolving landscape of personalization, from RAG to advanced Agent-based systems, the evaluation of models relies heavily on diverse datasets and metrics tailored to specific tasks. This survey categorizes metrics into several key types: Textual Quality metrics (e.g., BLEU, ROUGE, METEOR) assess the fluency and coherence of generated outputs; Information Retrieval metrics (e.g., MAP, MRR, Recall) evaluate the accuracy and relevance of retrieved information; Classification metrics (e.g., Accuracy, F1) measure task-specific correctness; Regression metrics (e.g., MAE, RMSE) quantify prediction errors; and Other metrics (e.g., Fluency, Pass@k) address domain-specific or task-unique aspects like plausibility or executability. These metrics span pre-retrieval, retrieval, generation, and agent-based personalization approaches, reflecting their varied objectives. To provide a comprehensive overview, we compile an extensive list of datasets across these fields, as detailed in Table 2. These datasets, paired with their respective metrics, enable researchers to benchmark and refine personalized systems, from enhancing query rewriting to enabling autonomous agents in physical and virtual environments."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 70,
+ 582,
+ 263,
+ 593
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 582,
+ 263,
+ 593
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 582,
+ 263,
+ 593
+ ],
+ "type": "text",
+ "content": "6 CHALLENGES AND FUTURE DIRECTIONS"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 70,
+ 599,
+ 501,
+ 624
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 599,
+ 501,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 599,
+ 501,
+ 624
+ ],
+ "type": "text",
+ "content": "Personalized RAG and agent-based systems still face several critical challenges that warrant further exploration. We list them as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 70,
+ 630,
+ 501,
+ 654
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 630,
+ 501,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 630,
+ 501,
+ 654
+ ],
+ "type": "text",
+ "content": "- Balancing Personalization and Scalability: Integrating personalization data (such as preferences, history, and contextual signals) into RAG processes often increases computational complexity, making it difficult to maintain"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 501,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 501,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 501,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "text",
+ "content": "17"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 16
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 114,
+ 96,
+ 539,
+ 121
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 96,
+ 539,
+ 121
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 96,
+ 539,
+ 121
+ ],
+ "type": "text",
+ "content": "efficiency and scalability across large-scale systems. Future work could explore lightweight, adaptive embeddings and hybrid frameworks that seamlessly fuse user profiles with real-time contexts."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 124,
+ 541,
+ 355
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 7,
+ "blocks": [
+ {
+ "bbox": [
+ 107,
+ 124,
+ 541,
+ 177
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 124,
+ 541,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 124,
+ 541,
+ 177
+ ],
+ "type": "text",
+ "content": "- Evaluating Personalization Effectively: Current metrics like BLEU, ROUGE, and human evaluations fall short in capturing the nuanced alignment of outputs with dynamic user preferences, lacking tailored measures for personalization efficacy. Developing specialized benchmarks and metrics that assess long-term user satisfaction and adaptability is crucial for real-world applicability."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 178,
+ 541,
+ 245
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 178,
+ 541,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 178,
+ 541,
+ 245
+ ],
+ "type": "text",
+ "content": "- Preserving Privacy through Device-Cloud Collaboration: Personalized retrieval often involves processing sensitive user data, raising privacy concerns, especially with the increased global emphasis on data protection regulations, such as the European Union's General Data Protection Regulation (GDPR). Consequently, a promising approach is the collaborative integration of on-device small Language models which handle sensitive personal data locally, with cloud-based LLM, which provides broader contextual knowledge."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 107,
+ 247,
+ 541,
+ 312
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 247,
+ 541,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 247,
+ 541,
+ 312
+ ],
+ "type": "text",
+ "content": "- Personalized Agent Planning: Current research on agent planning remains mainly in its early stages, with much of the work focusing on building foundational frameworks such as GUI agents [81] and the application of agents across diverse domains [131]. Notably, the incorporation of personalized approaches has yet to be widely adopted. Exploring how to integrate personalized support into existing frameworks to enhance user experience represents a promising and valuable direction for future investigation."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 316,
+ 541,
+ 355
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 316,
+ 541,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 316,
+ 541,
+ 355
+ ],
+ "type": "text",
+ "content": "- Ensuring Ethical and Coherent Systems: Bias in data processing, privacy concerns in user profiling, and coherence across retrieval and generation stages remain unresolved. Future directions should prioritize ethical safeguards, privacy-preserving techniques, and cross-stage optimization to build trustworthy, unified personalized systems."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 108,
+ 371,
+ 185,
+ 381
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 371,
+ 185,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 371,
+ 185,
+ 381
+ ],
+ "type": "text",
+ "content": "7 CONCLUSION"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 388,
+ 541,
+ 482
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 388,
+ 541,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 388,
+ 541,
+ 482
+ ],
+ "type": "text",
+ "content": "In this paper, we explore the landscape of personalization from Retrieval-Augmented Generation (RAG) to advanced LLM-based Agents, detailing adaptations across pre-retrieval, retrieval, and generation stages while extending into agentic capabilities. By reviewing recent literature, datasets, and metrics, we highlight the progress and diversity in enhancing user satisfaction through tailored AI systems. However, challenges such as scalability, effective evaluation, and ethical concerns underscore the need for innovative solutions. Future research should focus on lightweight frameworks, specialized benchmarks, and privacy-preserving techniques to advance personalized AI. Relevant papers and resources are also compiled online for ease of future research."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 108,
+ 499,
+ 167,
+ 509
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 499,
+ 167,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 499,
+ 167,
+ 509
+ ],
+ "type": "text",
+ "content": "REFERENCES"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 114,
+ 514,
+ 540,
+ 653
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 19,
+ "blocks": [
+ {
+ "bbox": [
+ 114,
+ 514,
+ 492,
+ 523
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 514,
+ 492,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 514,
+ 492,
+ 523
+ ],
+ "type": "text",
+ "content": "[1] 2021. BERT: a review of applications in natural language processing and understanding. arXiv preprint arXiv:2103.11943 (2021)."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 114,
+ 524,
+ 539,
+ 542
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 524,
+ 539,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 524,
+ 539,
+ 542
+ ],
+ "type": "text",
+ "content": "[2] Mahyar Abbasian, Iman Azimi, Amir M Rahmani, and Ramesh Jain. 2023. Conversational health agents: A personalized llm-powered agent framework. arXiv preprint arXiv:2310.02374 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 114,
+ 544,
+ 539,
+ 563
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 544,
+ 539,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 544,
+ 539,
+ 563
+ ],
+ "type": "text",
+ "content": "[3] Bashar Alhafni, Vivek Kulkarni, Dhruv Kumar, and Vipul Raheja. 2024. Personalized Text Generation with Fine-Grained Linguistic Control. In Proceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024). 88–101."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 563,
+ 427,
+ 573
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 563,
+ 427,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 563,
+ 427,
+ 573
+ ],
+ "type": "text",
+ "content": "[4] Amazon. [n.d.]. Amazon Customer Review Dataset. Online dataset. https://nijianmo.github.io/amazon/"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 115,
+ 574,
+ 540,
+ 601
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 574,
+ 540,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 574,
+ 540,
+ 601
+ ],
+ "type": "text",
+ "content": "[5] Steven Au, Cameron J Dimacali, Ojasmitha Pedirappagari, Namyong Park, Franck Dernoncourt, Yu Wang, Nikos Kanakaris, Hanieh Deilamsalehy, Ryan A Rossi, and Nesreen K Ahmed. 2025. Personalized Graph-Based Retrieval for Large Language Models. arXiv preprint arXiv:2501.02157 (2025)."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 603,
+ 539,
+ 622
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 603,
+ 539,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 603,
+ 539,
+ 622
+ ],
+ "type": "text",
+ "content": "[6] Elias Bassani, Pranav Kasela, Alessandro Raganato, and Gabriella Pasi. 2022. A multi-domain benchmark for personalized search evaluation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3822-3827."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 624,
+ 539,
+ 642
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 624,
+ 539,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 624,
+ 539,
+ 642
+ ],
+ "type": "text",
+ "content": "[7] Elias Bassani, Nicola Tonellotto, and Gabriella Pasi. 2023. Personalized query expansion with contextual word embeddings. ACM Transactions on Information Systems 42, 2 (2023), 1-35."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 114,
+ 643,
+ 527,
+ 653
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 643,
+ 527,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 643,
+ 527,
+ 653
+ ],
+ "type": "text",
+ "content": "[8] Oliver Baumann and Mirco Schoenfeld. 2024. PSQE: Personalized Semantic Query Expansion for user-centric query disambiguation. (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 538,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "text",
+ "content": "18"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 17
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 75,
+ 98,
+ 503,
+ 645
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 30,
+ "blocks": [
+ {
+ "bbox": [
+ 78,
+ 98,
+ 503,
+ 126
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 98,
+ 503,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 98,
+ 503,
+ 126
+ ],
+ "type": "text",
+ "content": "[9] Matthias Bender, Tom Crecelius, Mouna Kacimi, Sebastian Michel, Thomas Neumann, Josiane Xavier Parreira, Ralf Schenkel, and Gerhard Weikum. 2008. Exploiting social relations for query expansion and result ranking. In 2008 IEEE 24th International Conference on Data Engineering Workshop. IEEE, 501-506."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 75,
+ 127,
+ 503,
+ 147
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 127,
+ 503,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 127,
+ 503,
+ 147
+ ],
+ "type": "text",
+ "content": "[10] Marin Bertier, Rachid Guerraoui, Vincent Leroy, and Anne-Marie Kermarrec. 2009. Toward personalized query expansion. In Proceedings of the Second ACM EuroSys Workshop on Social Network Systems. 7-12."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 75,
+ 148,
+ 503,
+ 167
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 148,
+ 503,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 148,
+ 503,
+ 167
+ ],
+ "type": "text",
+ "content": "[11] Keping Bi, Qingyao Ai, and W Bruce Croft. 2021. Learning a fine-grained review-based transformer model for personalized product search. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 123-132."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 75,
+ 167,
+ 503,
+ 186
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 167,
+ 503,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 167,
+ 503,
+ 186
+ ],
+ "type": "text",
+ "content": "[12] Claudio Biancalana and Alessandro Micarelli. 2009. Social tagging in query expansion: A new way for personalized web search. In 2009 International Conference on Computational Science and Engineering, Vol. 4. IEEE, 1060-1065."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 75,
+ 187,
+ 282,
+ 197
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 187,
+ 282,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 187,
+ 282,
+ 197
+ ],
+ "type": "text",
+ "content": "[13] Microsoft Bing. [n.d]. Bing Search Engine. https://www.bing.com"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 75,
+ 198,
+ 503,
+ 217
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 198,
+ 503,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 198,
+ 503,
+ 217
+ ],
+ "type": "text",
+ "content": "[14] Mohamed Reda Bouadjenek, Hakim Hacid, and Mokrane Bouzeghoub. 2019. Personalized social query expansion using social annotations. Transactions on Large-Scale Data-and Knowledge-Centered Systems XL (2019), 1-25."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 75,
+ 217,
+ 503,
+ 246
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 217,
+ 503,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 217,
+ 503,
+ 246
+ ],
+ "type": "text",
+ "content": "[15] Mohamed Reda Bouadjenek, Hakim Hacid, Mokrane Bouzeghoub, and Johann Daigremont. 2011. Personalized social query expansion using social bookmarking systems. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval. 1113-1114."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 75,
+ 247,
+ 486,
+ 256
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 247,
+ 486,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 247,
+ 486,
+ 256
+ ],
+ "type": "text",
+ "content": "[16] Domenico Bulfamante. 2023. Generative enterprise search with extensible knowledge base using ai. Ph.D. Dissertation. Politecnico di Torino."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 75,
+ 258,
+ 503,
+ 276
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 258,
+ 503,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 258,
+ 503,
+ 276
+ ],
+ "type": "text",
+ "content": "[17] Hongru Cai, Yongqi Li, Wenjie Wang, ZHU Fengbin, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. [n. d]. Large Language Models Empowered Personalized Web Agents. In THE WEB CONFERENCE 2025."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 75,
+ 277,
+ 503,
+ 296
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 277,
+ 503,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 277,
+ 503,
+ 296
+ ],
+ "type": "text",
+ "content": "[18] Hongzhan Chen, Hehong Chen, Ming Yan, Wenshen Xu, Xing Gao, Weizhou Shen, Xiaojun Quan, Chenliang Li, Ji Zhang, Fei Huang, et al. 2024. Socialbench: Sociality evaluation of role-playing conversational agents. arXiv preprint arXiv:2403.13679 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 75,
+ 297,
+ 503,
+ 316
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 297,
+ 503,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 297,
+ 503,
+ 316
+ ],
+ "type": "text",
+ "content": "[19] Jiangjie Chen, Xintao Wang, Rui Xu, Siyu Yuan, Yikai Zhang, Wei Shi, Jian Xie, Shuang Li, Ruihan Yang, Tinghui Zhu, et al. 2024. From persona to personalization: A survey on role-playing language agents. arXiv preprint arXiv:2404.18231 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 75,
+ 317,
+ 503,
+ 336
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 317,
+ 503,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 317,
+ 503,
+ 336
+ ],
+ "type": "text",
+ "content": "[20] Ruizhe Chen, Xiaotian Zhang, Meng Luo, Wenhao Chai, and Zuozhu Liu. 2024. Pad: Personalized alignment of llms at decoding-time. arXiv preprint arXiv:2410.04070 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 75,
+ 337,
+ 503,
+ 356
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 337,
+ 503,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 337,
+ 503,
+ 356
+ ],
+ "type": "text",
+ "content": "[21] Weijie Chen, Ting Bai, Jinbo Su, Jian Luan, Wei Liu, and Chuan Shi. 2024. Kg-retriever: Efficient knowledge indexing for retrieval-augmented large language models. arXiv preprint arXiv:2412.05547 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 75,
+ 357,
+ 503,
+ 376
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 357,
+ 503,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 357,
+ 503,
+ 376
+ ],
+ "type": "text",
+ "content": "[22] Zheng Chen, Ziyan Jiang, Fan Yang, Eunah Cho, Xing Fan, Xiaojiang Huang, Yanbin Lu, and Aram Galstyan. 2023. Graph meets LLM: A novel approach to collaborative filtering for robust conversational understanding. arXiv preprint arXiv:2305.14449 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 75,
+ 376,
+ 503,
+ 396
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 376,
+ 503,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 376,
+ 503,
+ 396
+ ],
+ "type": "text",
+ "content": "[23] Hao Cheng, Shuo Wang, Wensheng Lu, Wei Zhang, Mingyang Zhou, Kezhong Lu, and Hao Liao. 2023. Explainable recommendation with personalized review retrieval and aspect learning. arXiv preprint arXiv:2306.12657 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 75,
+ 396,
+ 503,
+ 415
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 396,
+ 503,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 396,
+ 503,
+ 415
+ ],
+ "type": "text",
+ "content": "[24] Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. 2023. Adapting language models to compress contexts. arXiv preprint arXiv:2305.14788 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 75,
+ 417,
+ 503,
+ 436
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 417,
+ 503,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 417,
+ 503,
+ 436
+ ],
+ "type": "text",
+ "content": "[25] Eunah Cho, Ziyan Jiang, Jie Hao, Zheng Chen, Saurabh Gupta, Xing Fan, and Chenlei Guo. 2021. Personalized search-based query rewrite system for conversational ai. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI. 179-188."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 75,
+ 437,
+ 503,
+ 456
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 437,
+ 503,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 437,
+ 503,
+ 456
+ ],
+ "type": "text",
+ "content": "[26]Hyundong Cho, Karishma Sharma, Nicolaas Jedema, Leonardo FR Ribeiro, Alessandro Moschitti, Ravi Krishnan, and Jonathan May. 2025. TuningFree Personalized Alignment via Trial-Error-Explain In-Context Learning. arXiv preprint arXiv:2502.08972 (2025)."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 75,
+ 456,
+ 503,
+ 475
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 456,
+ 503,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 456,
+ 503,
+ 475
+ ],
+ "type": "text",
+ "content": "[27] Yanqi Dai, Huanran Hu, Lei Wang, Shengjie Jin, Xu Chen, and Zhiwu Lu. 2024. Mmrole: A comprehensive framework for developing and evaluating multimodal role-playing agents. arXiv preprint arXiv:2408.04203 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 75,
+ 476,
+ 503,
+ 495
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 476,
+ 503,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 476,
+ 503,
+ 495
+ ],
+ "type": "text",
+ "content": "[28] Wenlong Deng, Christos Thrampoulidis, and Xiaoxiao Li. 2024. Unlocking the potential of prompt-tuning in bridging generalized and personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6087-6097."
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 75,
+ 496,
+ 503,
+ 515
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 496,
+ 503,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 496,
+ 503,
+ 515
+ ],
+ "type": "text",
+ "content": "[29] Yang Deng, Yaliang Li, Wenxuan Zhang, Bolin Ding, and Wai Lam. 2022. Toward personalized answer generation in e-commerce via multiperspective preference modeling. ACM Transactions on Information Systems (TOIS) 40, 4 (2022), 1-28."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 75,
+ 516,
+ 503,
+ 535
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 516,
+ 503,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 516,
+ 503,
+ 535
+ ],
+ "type": "text",
+ "content": "[30] Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The Faiss library. (2024). arXiv:2401.08281 [cs.LG]"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 75,
+ 536,
+ 264,
+ 545
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 536,
+ 264,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 536,
+ 264,
+ 545
+ ],
+ "type": "text",
+ "content": "[31] ESPN. [n.d.]. ESPN Sports Statistics Dataset. Online dataset."
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 75,
+ 546,
+ 503,
+ 574
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 546,
+ 503,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 546,
+ 503,
+ 574
+ ],
+ "type": "text",
+ "content": "[32] 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": 25
+ },
+ {
+ "bbox": [
+ 75,
+ 576,
+ 503,
+ 594
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 576,
+ 503,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 576,
+ 503,
+ 594
+ ],
+ "type": "text",
+ "content": "[33] Byron J Gao, David C Anastasiu, and Xing Jiang. 2010. Utilizing user-input contextual terms for query disambiguation. In Coling 2010: Posters. 329-337."
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 75,
+ 596,
+ 503,
+ 615
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 596,
+ 503,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 596,
+ 503,
+ 615
+ ],
+ "type": "text",
+ "content": "[34] Jingsheng Gao, Yixin Lian, Ziyi Zhou, Yuzhuo Fu, and Baoyuan Wang. 2023. LiveChat: A large-scale personalized dialogue dataset automatically constructed from live streaming. arXiv preprint arXiv:2306.08401 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 75,
+ 616,
+ 503,
+ 635
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 616,
+ 503,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 616,
+ 503,
+ 635
+ ],
+ "type": "text",
+ "content": "[35] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. 2023. Retrievalaugmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 2 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 75,
+ 635,
+ 251,
+ 645
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 635,
+ 251,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 635,
+ 251,
+ 645
+ ],
+ "type": "text",
+ "content": "[36] Google. [n.d.]. Google Search. https://www.google.com"
+ }
+ ]
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 291,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 291,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 291,
+ 666
+ ],
+ "type": "text",
+ "content": "19"
+ }
+ ]
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 18
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 111,
+ 98,
+ 539,
+ 655
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 26,
+ "blocks": [
+ {
+ "bbox": [
+ 111,
+ 98,
+ 539,
+ 126
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 98,
+ 539,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 98,
+ 539,
+ 126
+ ],
+ "type": "text",
+ "content": "[37] Jia-Chen Gu, Hui Liu, Zhen-Hua Ling, Quan Liu, Zhigang Chen, and Xiaodan Zhu. 2021. Partner matters! an empirical study on fusing personas for personalized response selection in retrieval-based chatbots. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 565-574."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 111,
+ 127,
+ 539,
+ 156
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 127,
+ 539,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 127,
+ 539,
+ 156
+ ],
+ "type": "text",
+ "content": "[38] Jie Hao, Yang Liu, Xing Fan, Saurabh Gupta, Saleh Soltan, Rakesh Chada, Pradeep Natarajan, Chenlei Guo, and Gokhan Tur. 2022. CGF: Constrained generation framework for query rewriting in conversational AI. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track. 475-483."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 111,
+ 157,
+ 539,
+ 176
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 157,
+ 539,
+ 176
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 157,
+ 539,
+ 176
+ ],
+ "type": "text",
+ "content": "[39] Nicola Henze, Peter Dolog, and Wolfgang Nejdl. 2004. Reasoning and ontologies for personalized e-learning in the semantic web. Journal of Educational Technology & Society 7, 4 (2004), 82-97."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 111,
+ 177,
+ 539,
+ 197
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 177,
+ 539,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 177,
+ 539,
+ 197
+ ],
+ "type": "text",
+ "content": "[40] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. 2023. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352 3, 4 (2023), 6."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 111,
+ 198,
+ 539,
+ 226
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 198,
+ 539,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 198,
+ 539,
+ 226
+ ],
+ "type": "text",
+ "content": "[41] WANG Hongru, Minda Hu, Yang Deng, Rui Wang, Fei Mi, Weichao Wang, Yasheng Wang, Wai-Chung Kwan, Irwin King, and Kam-Fai Wong. [n. d]. Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues. In The 2023 Conference on Empirical Methods in Natural Language Processing."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 111,
+ 227,
+ 539,
+ 247
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 227,
+ 539,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 227,
+ 539,
+ 247
+ ],
+ "type": "text",
+ "content": "[42] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR 1, 2 (2022), 3."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 111,
+ 247,
+ 539,
+ 266
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 247,
+ 539,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 247,
+ 539,
+ 266
+ ],
+ "type": "text",
+ "content": "[43] Qiushi Huang, Shuai Fu, Xubo Liu, Wenwu Wang, Tom Ko, Yu Zhang, and Lilian Tang. 2024. Learning retrieval augmentation for personalized dialogue generation. arXiv preprint arXiv:2406.18847 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 111,
+ 267,
+ 539,
+ 286
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 267,
+ 539,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 267,
+ 539,
+ 286
+ ],
+ "type": "text",
+ "content": "[44] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. 2022. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning. PMLR, 9118-9147."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 111,
+ 287,
+ 539,
+ 306
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 287,
+ 539,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 287,
+ 539,
+ 306
+ ],
+ "type": "text",
+ "content": "[45] Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of LLM agents: A survey. arXiv preprint arXiv:2402.02716 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 111,
+ 307,
+ 539,
+ 326
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 307,
+ 539,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 307,
+ 539,
+ 326
+ ],
+ "type": "text",
+ "content": "[46] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. 2023. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 111,
+ 327,
+ 539,
+ 355
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 327,
+ 539,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 327,
+ 539,
+ 355
+ ],
+ "type": "text",
+ "content": "[47] Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, and Prithviraj Ammanabrolu. 2023. Personalized soups: Personalized large language model alignment via post-hoc parameter merging. arXiv preprint arXiv:2310.11564 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 111,
+ 356,
+ 539,
+ 386
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 356,
+ 539,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 356,
+ 539,
+ 386
+ ],
+ "type": "text",
+ "content": "[48] Pengyue Jia, Yiding Liu, Xiangyu Zhao, Xiaopeng Li, Changying Hao, Shuaiqiang Wang, and Dawei Yin. 2024. MILL: Mutual Verification with Large Language Models for Zero-Shot Query Expansion. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2498-2518."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 111,
+ 387,
+ 539,
+ 406
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 387,
+ 539,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 387,
+ 539,
+ 406
+ ],
+ "type": "text",
+ "content": "[49] Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Zhang, and Yixin Zhu. 2023. Evaluating and inducing personality in pre-trained language models. Advances in Neural Information Processing Systems 36 (2023), 10622-10643."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 111,
+ 407,
+ 539,
+ 435
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 407,
+ 539,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 407,
+ 539,
+ 435
+ ],
+ "type": "text",
+ "content": "[50] Hideaki Joko, Shubham Chatterjee, Andrew Ramsay, Arjen P De Vries, Jeff Dalton, and Faegheh Hasibi. 2024. Doing personal laps: Llm-augmented dialogue construction for personalized multi-session conversational search. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 796-806."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 111,
+ 436,
+ 539,
+ 456
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 436,
+ 539,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 436,
+ 539,
+ 456
+ ],
+ "type": "text",
+ "content": "[51] Wang-Cheng Kang, Jianmo Ni, Nikhil Mehta, Maheswaran Sathiamoorthy, Lichan Hong, Ed Chi, and Derek Zhiyuan Cheng. 2023. Do llms understand user preferences? evaluating llms on user rating prediction. arXiv preprint arXiv:2305.06474 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 111,
+ 456,
+ 539,
+ 475
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 456,
+ 539,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 456,
+ 539,
+ 475
+ ],
+ "type": "text",
+ "content": "[52] Manojkumar Rangasamy Kannadasan and Grigor Aslanyan. 2019. Personalized query auto-completion through a lightweight representation of the user context. arXiv preprint arXiv:1905.01386 (2019)."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 111,
+ 476,
+ 539,
+ 515
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 476,
+ 539,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 476,
+ 539,
+ 515
+ ],
+ "type": "text",
+ "content": "[53] Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Greg Corrado, Laszlo Lukacs, Marina Ganea, Peter Young, and Vivek Ramavajjala. 2016. Smart Reply: Automated Response Suggestion for Email. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (San Francisco, California, USA) (KDD '16). Association for Computing Machinery, New York, NY, USA, 955-964. https://doi.org/10.1145/2939672.2939801"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 111,
+ 516,
+ 539,
+ 535
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 516,
+ 539,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 516,
+ 539,
+ 535
+ ],
+ "type": "text",
+ "content": "[54] Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. 2024. Reinforcement learning for optimizing rag for domain chatbots. arXiv preprint arXiv:2401.06800 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 111,
+ 536,
+ 539,
+ 555
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 536,
+ 539,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 536,
+ 539,
+ 555
+ ],
+ "type": "text",
+ "content": "[55] Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, et al. 2024. Longlamp: A benchmark for personalized long-form text generation. arXiv preprint arXiv:2407.11016 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 111,
+ 556,
+ 539,
+ 575
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 556,
+ 539,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 556,
+ 539,
+ 575
+ ],
+ "type": "text",
+ "content": "[56] Brenden Lake and Marco Baroni. 2018. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning. PMLR, 2873-2882."
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 111,
+ 576,
+ 539,
+ 595
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 576,
+ 539,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 576,
+ 539,
+ 595
+ ],
+ "type": "text",
+ "content": "[57] Cheng Li, Mingyang Zhang, Qiao zhu Mei, Weize Kong, and Michael Bendersky. 2024. Learning to rewrite prompts for personalized text generation. In Proceedings of the ACM Web Conference 2024. 3367-3378."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 111,
+ 596,
+ 539,
+ 615
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 596,
+ 539,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 596,
+ 539,
+ 615
+ ],
+ "type": "text",
+ "content": "[58] Changhao Li, Yuchen Zhuang, Rushi Qiang, Haotian Sun, Hanjun Dai, Chao Zhang, and Bo Dai. 2024. Matryoshka: Learning to Drive Black-Box LLMs with LLMs. arXiv preprint arXiv:2410.20749 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 111,
+ 616,
+ 539,
+ 635
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 616,
+ 539,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 616,
+ 539,
+ 635
+ ],
+ "type": "text",
+ "content": "[59] Lei Li, Yongfeng Zhang, and Li Chen. 2023. Personalized prompt learning for explainable recommendation. ACM Transactions on Information Systems 41, 4 (2023), 1-26."
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 111,
+ 635,
+ 539,
+ 655
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 635,
+ 539,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 635,
+ 539,
+ 655
+ ],
+ "type": "text",
+ "content": "[60] Sen Li, Fuyu Lv, Taiwei Jin, Guiyang Li, Yukun Zheng, Tao Zhuang, Qingwen Liu, Xiaoyi Zeng, James Kwok, and Qianli Ma. 2022. Query rewriting in taobao search. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3262-3271."
+ }
+ ]
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 320,
+ 658,
+ 327,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 327,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 327,
+ 666
+ ],
+ "type": "text",
+ "content": "20"
+ }
+ ]
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 19
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 75,
+ 98,
+ 503,
+ 655
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 29,
+ "blocks": [
+ {
+ "bbox": [
+ 75,
+ 98,
+ 503,
+ 117
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 98,
+ 503,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 98,
+ 503,
+ 117
+ ],
+ "type": "text",
+ "content": "[61] Xiaopeng Li, Lixin Su, Pengyue Jia, Xiangyu Zhao, Suqi Cheng, Junfeng Wang, and Dawei Yin. 2023. Agent4ranking: Semantic robust ranking via personalized query rewriting using multi-agent llm. arXiv preprint arXiv:2312.15450 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 75,
+ 118,
+ 503,
+ 137
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 118,
+ 503,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 118,
+ 503,
+ 137
+ ],
+ "type": "text",
+ "content": "[62] Xinyu Li, Ruiyang Zhou, Zachary C Lipton, and Liu Leqi. 2024. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 75,
+ 138,
+ 502,
+ 157
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 138,
+ 502,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 138,
+ 502,
+ 157
+ ],
+ "type": "text",
+ "content": "[63] Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, et al. 2024. Personal llm agents: Insights and survey about the capability, efficiency and security. arXiv preprint arXiv:2401.05459 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 75,
+ 158,
+ 502,
+ 177
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 158,
+ 502,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 158,
+ 502,
+ 177
+ ],
+ "type": "text",
+ "content": "[64] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 75,
+ 178,
+ 503,
+ 205
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 178,
+ 503,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 178,
+ 503,
+ 205
+ ],
+ "type": "text",
+ "content": "[65] Ruixue Lian, Sixing Lu, Clint Solomon, Gustavo Aguilar, Pragaash Ponnusamy, Jialong Han, Chengyuan Ma, and Chenlei Guo. 2023. PersonalTM: Transformer memory for personalized retrieval. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2256-2260."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 75,
+ 207,
+ 503,
+ 226
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 207,
+ 503,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 207,
+ 503,
+ 226
+ ],
+ "type": "text",
+ "content": "[66] Shan-Mu Lin and Chuen-Min Huang. 2006. Personalized optimal search in local query expansion. In Proceedings of the 18th Conference on Computational Linguistics and Speech Processing. 221-236."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 75,
+ 228,
+ 503,
+ 246
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 228,
+ 503,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 228,
+ 503,
+ 246
+ ],
+ "type": "text",
+ "content": "[67] Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is chatgpt a good recommender? a preliminary study. arXiv preprint arXiv:2304.10149 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 75,
+ 247,
+ 503,
+ 266
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 247,
+ 503,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 247,
+ 503,
+ 266
+ ],
+ "type": "text",
+ "content": "[68] Jiahong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Jieming Zhu, Minda Hu, Menglin Yang, and Irwin King. 2025. A Survey of Personalized Large Language Models: Progress and Future Directions. arXiv preprint arXiv:2502.11528 (2025)."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 75,
+ 267,
+ 503,
+ 286
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 267,
+ 503,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 267,
+ 503,
+ 286
+ ],
+ "type": "text",
+ "content": "[69] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics 12 (2024), 157-173."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 75,
+ 287,
+ 503,
+ 306
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 287,
+ 503,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 287,
+ 503,
+ 306
+ ],
+ "type": "text",
+ "content": "[70] Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024. Once: Boosting content-based recommendation with both open-and closed-source large language models. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining. 452-461."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 75,
+ 307,
+ 503,
+ 326
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 307,
+ 503,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 307,
+ 503,
+ 326
+ ],
+ "type": "text",
+ "content": "[71] Shuai Liu, Hyundong J Cho, Marjorie Freedman, Xuezhe Ma, and Jonathan May. 2023. RECAP: retrieval-enhanced context-aware prefix encoder for personalized dialogue response generation. arXiv preprint arXiv:2306.07206 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 75,
+ 327,
+ 479,
+ 336
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 327,
+ 479,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 327,
+ 479,
+ 336
+ ],
+ "type": "text",
+ "content": "[72] Tyler Lu and Craig Boutilier. 2011. Budgeted social choice: From consensus to personalized decision making. In *IJCAI*, Vol. 11, 280-286."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 75,
+ 338,
+ 503,
+ 356
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 338,
+ 503,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 338,
+ 503,
+ 356
+ ],
+ "type": "text",
+ "content": "[73] Zhengyi Ma, Zhicheng Dou, Yutao Zhu, Hanxun Zhong, and Ji-Rong Wen. 2021. One chatbot per person: Creating personalized chatbots based on implicit user profiles. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 555-564."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 75,
+ 357,
+ 503,
+ 376
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 357,
+ 503,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 357,
+ 503,
+ 376
+ ],
+ "type": "text",
+ "content": "[74] Aman Madaan, Niket Tandon, Peter Clark, and Yiming Yang. 2022. Memory-assisted prompt editing to improve GPT-3 after deployment. arXiv preprint arXiv:2201.06009 (2022)."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 75,
+ 376,
+ 503,
+ 396
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 376,
+ 503,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 376,
+ 503,
+ 396
+ ],
+ "type": "text",
+ "content": "[75] Shengyu Mao, Xiaohan Wang, Mengru Wang, Yong Jiang, Pengjun Xie, Fei Huang, and Ningyu Zhang. 2024. Editing Personality for Large Language Models. In CCF International Conference on Natural Language Processing and Chinese Computing. Springer, 241-254."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 75,
+ 397,
+ 503,
+ 415
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 397,
+ 503,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 397,
+ 503,
+ 415
+ ],
+ "type": "text",
+ "content": "[76] Yuren Mao, Xuemei Dong, Wenyi Xu, Yunjun Gao, Bin Wei, and Ying Zhang. 2024. Fit-rag: black-box rag with factual information and token reduction. arXiv preprint arXiv:2403.14374 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 75,
+ 417,
+ 503,
+ 445
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 417,
+ 503,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 417,
+ 503,
+ 445
+ ],
+ "type": "text",
+ "content": "[77] Puneet Mathur, Zhe Liu, Ke Li, Yingyi Ma, Gil Keren, Zeeshan Ahmed, Dinesh Manocha, and Xuedong Zhang. 2023. Personal: Language model personalization via domain-distributed span aggregated k-nearest n-gram retrieval augmentation. In Findings of the Association for Computational Linguistics: EMNLP 2023. 11314-11328."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 75,
+ 447,
+ 503,
+ 465
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 447,
+ 503,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 447,
+ 503,
+ 465
+ ],
+ "type": "text",
+ "content": "[78] Fatemehsadat Mireshghallah, Vaishnavi Shrivastava, Milad Shokouhi, Taylor Berg-Kirkpatrick, Robert Sim, and Dimitrios Dimitriadis. 2021. Identifier: Implicit user representations for simple and effective personalized sentiment analysis. arXiv preprint arXiv:2110.00135 (2021)."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 75,
+ 467,
+ 503,
+ 485
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 467,
+ 503,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 467,
+ 503,
+ 485
+ ],
+ "type": "text",
+ "content": "[79] Philippe Mulhem, Nawal Ould Amer, and Mathias Gery. 2016. Axiomatic term-based personalized query expansion using bookmarking system. In International Conference on Database and Expert Systems Applications. Springer, 235-243."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 75,
+ 487,
+ 503,
+ 514
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 487,
+ 503,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 487,
+ 503,
+ 514
+ ],
+ "type": "text",
+ "content": "[80] Sheshera Mysore, Zhuoran Lu, Mengting Wan, Longqi Yang, Steve Menezes, Tina Baghaee, Emmanuel Barajas Gonzalez, Jennifer Neville, and Tara Safavi. 2023. Pearl: Personalizing large language model writing assistants with generation-calibrated retrievers. arXiv preprint arXiv:2311.09180 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 75,
+ 517,
+ 503,
+ 535
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 517,
+ 503,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 517,
+ 503,
+ 535
+ ],
+ "type": "text",
+ "content": "[81] Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, et al. 2024. Gui agents: A survey. arXiv preprint arXiv:2412.13501 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 75,
+ 536,
+ 503,
+ 555
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 536,
+ 503,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 536,
+ 503,
+ 555
+ ],
+ "type": "text",
+ "content": "[82] Duy A Nguyen, Rishi Kesav Mohan, Van Yang, Pritom Saha Akash, and Kevin Chen-Chuan Chang. 2025. RL-based Query Rewriting with Distilled LLM for online E-Commerce Systems. arXiv preprint arXiv:2501.18056 (2025)."
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 75,
+ 556,
+ 503,
+ 585
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 556,
+ 503,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 556,
+ 503,
+ 585
+ ],
+ "type": "text",
+ "content": "[83] Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 188-197."
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 75,
+ 586,
+ 503,
+ 605
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 586,
+ 503,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 586,
+ 503,
+ 605
+ ],
+ "type": "text",
+ "content": "[84] Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O'Banion, and Jun Xie. 2024. User-llm: Efficient llm contextualization with user embeddings. arXiv preprint arXiv:2402.13598 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 75,
+ 606,
+ 503,
+ 624
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 606,
+ 503,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 606,
+ 503,
+ 624
+ ],
+ "type": "text",
+ "content": "[85] Douglas Oard, William Webber, David Kirsch, and Sergey Golitsynski. 2015. Avocado research email collection. Philadelphia: Linguistic Data Consortium (2015)."
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 75,
+ 625,
+ 452,
+ 635
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 625,
+ 452,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 625,
+ 452,
+ 635
+ ],
+ "type": "text",
+ "content": "[86] U.S. National Library of Medicine. [n.d.]. PubMed: A Free Resource for Biomedical Literature. https://pubmed.ncbi.nlm.nih.gov/"
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 75,
+ 636,
+ 503,
+ 655
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 75,
+ 636,
+ 503,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 75,
+ 636,
+ 503,
+ 655
+ ],
+ "type": "text",
+ "content": "[87] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology. 1-22."
+ }
+ ]
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "text",
+ "content": "21"
+ }
+ ]
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 20
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 110,
+ 98,
+ 539,
+ 655
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 31,
+ "blocks": [
+ {
+ "bbox": [
+ 112,
+ 98,
+ 539,
+ 117
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 98,
+ 539,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 98,
+ 539,
+ 117
+ ],
+ "type": "text",
+ "content": "[88] Greg Pass, Abdur Chowdhury, and Cayley Torgeson. 2006. A picture of search. In Proceedings of the 1st International Conference on Scalable Information Systems (Hong Kong) (InfoScale '06). Association for Computing Machinery, New York, NY, USA, 1-es. https://doi.org/10.1145/1146847.1146848"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 118,
+ 539,
+ 137
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 118,
+ 539,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 118,
+ 539,
+ 137
+ ],
+ "type": "text",
+ "content": "[89] Vadim Igorevich Pavliukevich, Alina Khasanovna Zherdeva, Olesya Vladimirovna Makhnytkina, and Dmitriy Viktorovich Dyrmovskiy. [n. d.]. Improving RAG with LoRA finetuning for persona text generation. ([n. d.])."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 112,
+ 137,
+ 537,
+ 147
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 137,
+ 537,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 137,
+ 537,
+ 147
+ ],
+ "type": "text",
+ "content": "[90] Dan Peng, Zhihui Fu, and Jun Wang. 2024. Pocketllm: Enabling on-device fine-tuning for personalized llms. arXiv preprint arXiv:2407.01031 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 111,
+ 148,
+ 539,
+ 167
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 148,
+ 539,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 148,
+ 539,
+ 167
+ ],
+ "type": "text",
+ "content": "[91] Qiyao Peng, Hongtao Liu, Hongyan Xu, Qing Yang, Minglai Shao, and Wenjun Wang. 2024. Review-LLM: Harnessing Large Language Models for Personalized Review Generation. arXiv:2407.07487 [cs.CL] https://arxiv.org/abs/2407.07487"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 111,
+ 167,
+ 539,
+ 187
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 167,
+ 539,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 167,
+ 539,
+ 187
+ ],
+ "type": "text",
+ "content": "[92] Hongjin Qian, Zhicheng Dou, Yutao Zhu, Yueyuan Ma, and Ji-Rong Wen. 2021. Learning implicit user profile for personalized retrieval-based chatbot. In proceedings of the 30th ACM international conference on Information & Knowledge Management. 1467-1477."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 111,
+ 188,
+ 539,
+ 216
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 188,
+ 539,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 188,
+ 539,
+ 216
+ ],
+ "type": "text",
+ "content": "[93] Hongjin Qian, Xiahe Li, Hanxun Zhong, Yu Guo, Yueyuan Ma, Yutao Zhu, Zhanliang Liu, Zhicheng Dou, and Ji-Rong Wen. 2021. Pchatbot: a large-scale dataset for personalized chatbot. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 2470-2477."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 111,
+ 217,
+ 539,
+ 236
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 217,
+ 539,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 217,
+ 539,
+ 236
+ ],
+ "type": "text",
+ "content": "[94] Xiaoru Qu, Yifan Wang, Zhao Li, and Jun Gao. 2024. Graph-enhanced prompt learning for personalized review generation. Data Science and Engineering 9, 3 (2024), 309-324."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 111,
+ 237,
+ 539,
+ 247
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 237,
+ 539,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 237,
+ 539,
+ 247
+ ],
+ "type": "text",
+ "content": "[95] A. Rajaraman and J.D. Ullman. 2011. Mining of Massive Datasets. Cambridge University Press. https://books.google.co.uk/books?id=OefRhZyYOb0C"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 111,
+ 247,
+ 539,
+ 266
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 247,
+ 539,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 247,
+ 539,
+ 266
+ ],
+ "type": "text",
+ "content": "[96] Yiting Ran, Xintao Wang, Rui Xu, Xinfeng Yuan, Jiaqing Liang, Deqing Yang, and Yanghua Xiao. 2024. Capturing minds, not just words: Enhancing role-playing language models with personality-indicative data. arXiv preprint arXiv:2406.18921 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 111,
+ 267,
+ 539,
+ 286
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 267,
+ 539,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 267,
+ 539,
+ 286
+ ],
+ "type": "text",
+ "content": "[97] Chandan K. Reddy, Lluis Marquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopadhyay, Arnab Biswas, Anlu Xing, and Karthik Subbian. 2022. Shopping Queries Dataset: A Large-Scale ESCI Benchmark for Improving Product Search. (2022). arXiv:2206.06588"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 111,
+ 287,
+ 537,
+ 297
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 287,
+ 537,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 287,
+ 537,
+ 297
+ ],
+ "type": "text",
+ "content": "[98] Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084 (2019)."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 111,
+ 297,
+ 539,
+ 316
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 297,
+ 539,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 297,
+ 539,
+ 316
+ ],
+ "type": "text",
+ "content": "[99] Ruiyang Ren, Peng Qiu, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Hua Wu, Ji-Rong Wen, and Haifeng Wang. 2024. Bases: Large-scale web search user simulation with large language model based agents. arXiv preprint arXiv:2402.17505 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 110,
+ 317,
+ 539,
+ 335
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 317,
+ 539,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 317,
+ 539,
+ 335
+ ],
+ "type": "text",
+ "content": "[100] Matthew Renze and Erhan Guven. 2024. Self-reflection in llm agents: Effects on problem-solving performance. arXiv preprint arXiv:2405.06682 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 110,
+ 337,
+ 539,
+ 356
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 337,
+ 539,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 337,
+ 539,
+ 356
+ ],
+ "type": "text",
+ "content": "[101] 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": 15
+ },
+ {
+ "bbox": [
+ 110,
+ 357,
+ 539,
+ 376
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 357,
+ 539,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 357,
+ 539,
+ 376
+ ],
+ "type": "text",
+ "content": "[102] Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval 3, 4 (2009), 333-389."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 110,
+ 376,
+ 539,
+ 396
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 376,
+ 539,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 376,
+ 539,
+ 396
+ ],
+ "type": "text",
+ "content": "[103] 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": 17
+ },
+ {
+ "bbox": [
+ 110,
+ 396,
+ 539,
+ 416
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 396,
+ 539,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 396,
+ 539,
+ 416
+ ],
+ "type": "text",
+ "content": "[104] Alireza Salemi, Cheng Li, Mingyang Zhang, Qiao zhu Mei, Weize Kong, Tao Chen, Zhuowan Li, Michael Bendersky, and Hamed Zamani. 2025. Reasoning-Enhanced Self-Training for Long-Form Personalized Text Generation. arXiv preprint arXiv:2501.04167 (2025)."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 110,
+ 417,
+ 539,
+ 436
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 417,
+ 539,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 417,
+ 539,
+ 436
+ ],
+ "type": "text",
+ "content": "[105] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. LaMP: When Large Language Models Meet Personalization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 7370-7392."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 110,
+ 437,
+ 539,
+ 455
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 437,
+ 539,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 437,
+ 539,
+ 455
+ ],
+ "type": "text",
+ "content": "[106] Alireza Salemi and Hamed Zamani. 2024. Learning to Rank for Multiple Retrieval-Augmented Models through Iterative Utility Maximization. arXiv preprint arXiv:2410.09942 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 110,
+ 456,
+ 539,
+ 475
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 456,
+ 539,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 456,
+ 539,
+ 475
+ ],
+ "type": "text",
+ "content": "[107] Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect?. In International Conference on Machine Learning. PMLR, 29971-30004."
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 110,
+ 476,
+ 539,
+ 495
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 476,
+ 539,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 476,
+ 539,
+ 495
+ ],
+ "type": "text",
+ "content": "[108] Rossano Schifanella, Alain Barrat, Ciro Cattuto, Benjamin Markines, and Filippo Menczer. 2010. Folks in folksonomies: social link prediction from shared metadata. In Proceedings of the third ACM international conference on Web search and data mining. 271-280."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 110,
+ 496,
+ 539,
+ 515
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 496,
+ 539,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 496,
+ 539,
+ 515
+ ],
+ "type": "text",
+ "content": "[109] Noor Shaker, Georgios Yannakakis, and Julian Togelius. 2010. Towards automatic personalized content generation for platform games. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, Vol. 6. 63-68."
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 110,
+ 516,
+ 537,
+ 525
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 516,
+ 537,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 516,
+ 537,
+ 525
+ ],
+ "type": "text",
+ "content": "[110] Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2023. Character-llm: A trainable agent for role-playing. arXiv preprint arXiv:2310.10158 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 110,
+ 526,
+ 539,
+ 545
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 526,
+ 539,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 526,
+ 539,
+ 545
+ ],
+ "type": "text",
+ "content": "[111] Jocelyn Shen, Joel Mire, Hae Won Park, Cynthia Breazeal, and Maarten Sap. 2024. HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs. arXiv preprint arXiv:2405.17633 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 110,
+ 546,
+ 539,
+ 565
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 546,
+ 539,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 546,
+ 539,
+ 565
+ ],
+ "type": "text",
+ "content": "[112] Yunxiao Shi, Xing Zi, Zijing Shi, Haimin Zhang, Qiang Wu, and Min Xu. 2024. Eragent: Enhancing retrieval-augmented language models with improved accuracy, efficiency, and personalization. arXiv preprint arXiv:2405.06683 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 110,
+ 566,
+ 539,
+ 585
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 566,
+ 539,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 566,
+ 539,
+ 585
+ ],
+ "type": "text",
+ "content": "[113] Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. 2025. Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG. arXiv preprint arXiv:2501.09136 (2025)."
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 110,
+ 586,
+ 539,
+ 614
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 586,
+ 539,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 586,
+ 539,
+ 614
+ ],
+ "type": "text",
+ "content": "[114] Harmanpreet Singh, Nikhil Verma, Yixiao Wang, Manasa Bharadwaj, Homa Fashandi, Kevin Ferreira, and Chul Lee. 2024. Personal Large Language Model Agents: A Case Study on Tailored Travel Planning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 486-514."
+ }
+ ]
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 110,
+ 616,
+ 539,
+ 644
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 616,
+ 539,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 616,
+ 539,
+ 644
+ ],
+ "type": "text",
+ "content": "[115] Shamane Siriwardhana, Rivindu Weerasekera, Elliott Wen, Tharindu Kaluarachchi, Rajib Rana, and Suranga Nanayakkara. 2023. Improving the domain adaptation of retrieval augmented generation (RAG) models for open domain question answering. Transactions of the Association for Computational Linguistics 11 (2023), 1-17."
+ }
+ ]
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 110,
+ 646,
+ 522,
+ 655
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 646,
+ 522,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 646,
+ 522,
+ 655
+ ],
+ "type": "text",
+ "content": "[116] Mingyang Song and Mao Zheng. 2024. A Survey of Query Optimization in Large Language Models. arXiv preprint arXiv:2412.17558 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 665
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 665
+ ],
+ "type": "text",
+ "content": "22"
+ }
+ ]
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 21
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 72,
+ 98,
+ 502,
+ 654
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 30,
+ "blocks": [
+ {
+ "bbox": [
+ 72,
+ 98,
+ 394,
+ 107
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 98,
+ 394,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 98,
+ 394,
+ 107
+ ],
+ "type": "text",
+ "content": "[117] Spotify. 2023. Annoy: Approximate Nearest Neighbors in C++/Python. https://github.com/spotify/annoy"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 72,
+ 108,
+ 502,
+ 126
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 108,
+ 502,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 108,
+ 502,
+ 126
+ ],
+ "type": "text",
+ "content": "[118] Stuck_In_the Matrix. 2015. Reddit Public Comments (2007-10 through 2015-05). (2015). https://www.reddit.com/r/datasets/comments/3bxlg7/i_have EVERY_publicly-available Reddit_COMMENT/"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 72,
+ 128,
+ 502,
+ 146
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 128,
+ 502,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 128,
+ 502,
+ 146
+ ],
+ "type": "text",
+ "content": "[119] Lei Sun, Jinming Zhao, and Qin Jin. 2024. Revealing Personality Traits: A New Benchmark Dataset for Explanable Personality Recognition on Dialogues. arXiv preprint arXiv:2409.19723 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 72,
+ 148,
+ 502,
+ 167
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 148,
+ 502,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 148,
+ 502,
+ 167
+ ],
+ "type": "text",
+ "content": "[120] 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": 5
+ },
+ {
+ "bbox": [
+ 72,
+ 168,
+ 502,
+ 186
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 168,
+ 502,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 168,
+ 502,
+ 186
+ ],
+ "type": "text",
+ "content": "[121] 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 preprint arXiv:2402.04401 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 72,
+ 187,
+ 502,
+ 206
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 187,
+ 502,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 187,
+ 502,
+ 206
+ ],
+ "type": "text",
+ "content": "[122] Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2025. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. arXiv:2402.04401 [cs.CL] https://arxiv.org/abs/2402.04401"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 72,
+ 207,
+ 502,
+ 226
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 207,
+ 502,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 207,
+ 502,
+ 226
+ ],
+ "type": "text",
+ "content": "[123] Quan Tu, Shilong Fan, Zihang Tian, and Rui Yan. 2024. Charactereval: A Chinese benchmark for role-playing conversational agent evaluation. arXiv preprint arXiv:2401.01275 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 72,
+ 227,
+ 351,
+ 236
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 227,
+ 351,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 227,
+ 351,
+ 236
+ ],
+ "type": "text",
+ "content": "[124] Cornell University. [n.d.]. arXiv: An Open Access Repository for Research. https://arxiv.org/"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 72,
+ 237,
+ 502,
+ 266
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 237,
+ 502,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 237,
+ 502,
+ 266
+ ],
+ "type": "text",
+ "content": "[125] Hemanth Vemuri, Sheshansh Agrawal, Shivam Mittal, Deepak Saini, Akshay Soni, Abhinav V Sambasivan, Wenhao Lu, Yajun Wang, Mehul Parsana, Purushottam Kar, et al. 2023. Personalized retrieval over millions of items. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1014-1022."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 72,
+ 266,
+ 502,
+ 286
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 266,
+ 502,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 266,
+ 502,
+ 286
+ ],
+ "type": "text",
+ "content": "[126] Bryan Wang, Gang Li, and Yang Li. 2023. Enabling conversational interaction with mobile ui using large language models. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1-17."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 72,
+ 287,
+ 502,
+ 306
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 287,
+ 502,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 287,
+ 502,
+ 306
+ ],
+ "type": "text",
+ "content": "[127] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 72,
+ 307,
+ 502,
+ 326
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 307,
+ 502,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 307,
+ 502,
+ 326
+ ],
+ "type": "text",
+ "content": "[128] Hongru Wang, Wenyu Huang, Yang Deng, Rui Wang, Zezhong Wang, Yufei Wang, Fei Mi, Jeff Z Pan, and Kam-Fai Wong. 2024. Unims-rag: A unified multi-source retrieval-augmented generation for personalized dialogue systems. arXiv preprint arXiv:2401.13256 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 72,
+ 327,
+ 502,
+ 346
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 327,
+ 502,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 327,
+ 502,
+ 346
+ ],
+ "type": "text",
+ "content": "[129] Hongru Wang, Rui Wang, Fei Mi, Yang Deng, Zezhong Wang, Bin Liang, Ruifeng Xu, and Kam-Fai Wong. 2023. Cue-CoT: Chain-of-thought prompting for responding to in-depth dialogue questions with LLMs. arXiv preprint arXiv:2305.11792 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 72,
+ 346,
+ 502,
+ 365
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 346,
+ 502,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 346,
+ 502,
+ 365
+ ],
+ "type": "text",
+ "content": "[130] Jian Wang, Yi Cheng, Dongding Lin, Chak Tou Leong, and Wenjie Li. 2023. Target-oriented proactive dialogue systems with personalization: Problem formulation and dataset curation. arXiv preprint arXiv:2310.07397 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 72,
+ 366,
+ 502,
+ 385
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 366,
+ 502,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 366,
+ 502,
+ 385
+ ],
+ "type": "text",
+ "content": "[131] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2024. A survey on large language model based autonomous agents. Frontiers of Computer Science 18, 6 (2024), 186345."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 72,
+ 386,
+ 496,
+ 396
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 386,
+ 496,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 386,
+ 496,
+ 396
+ ],
+ "type": "text",
+ "content": "[132] Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query expansion with large language models. arXiv preprint arXiv:2303.07678 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 72,
+ 396,
+ 502,
+ 415
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 396,
+ 502,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 396,
+ 502,
+ 415
+ ],
+ "type": "text",
+ "content": "[133] Lei Wang, Jingsen Zhang, Hao Yang, Zhiyuan Chen, Jiakai Tang, Zeyu Zhang, Xu Chen, Yankai Lin, Ruihua Song, Wayne Xin Zhao, et al. 2023. User behavior simulation with large language model based agents. arXiv preprint arXiv:2306.02552 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 72,
+ 417,
+ 502,
+ 436
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 417,
+ 502,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 417,
+ 502,
+ 436
+ ],
+ "type": "text",
+ "content": "[134] Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, et al. 2023. Incharacter: Evaluating personality fidelity in role-playing agents through psychological interviews. arXiv preprint arXiv:2310.17976 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 72,
+ 437,
+ 502,
+ 456
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 437,
+ 502,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 437,
+ 502,
+ 456
+ ],
+ "type": "text",
+ "content": "[135] Yixiao Wang, Homa Fashandi, and Kevin Ferreira. 2024. Investigating the Personality Consistency in Quantized Role-Playing Dialogue Agents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 239–255."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 72,
+ 456,
+ 502,
+ 475
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 456,
+ 502,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 456,
+ 502,
+ 475
+ ],
+ "type": "text",
+ "content": "[136] Yu Wang, Yifan Gao, Xiusi Chen, Haoming Jiang, Shiyang Li, Jingfeng Yang, Qingyu Yin, Zheng Li, Xian Li, Bing Yin, et al. [n.d.]. MEMORYLLM: Towards Self-Updatable Large Language Models. In Forty-first International Conference on Machine Learning."
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 72,
+ 476,
+ 502,
+ 495
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 476,
+ 502,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 476,
+ 502,
+ 495
+ ],
+ "type": "text",
+ "content": "[137] Zheng Wang, Zhongyang Li, Zeren Jiang, Dandan Tu, and Wei Shi. 2024. Crafting Personalized Agents through Retrieval-Augmented Generation on Editable Memory Graphs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 4891-4906."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 72,
+ 496,
+ 502,
+ 515
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 496,
+ 502,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 496,
+ 502,
+ 515
+ ],
+ "type": "text",
+ "content": "[138] Zijie J Wang and Duen Horng Chau. 2024. MeMemo: On-device Retrieval Augmentation for Private and Personalized Text Generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2765-2770."
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 72,
+ 516,
+ 502,
+ 544
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 516,
+ 502,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 516,
+ 502,
+ 544
+ ],
+ "type": "text",
+ "content": "[139] Zekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, et al. 2023. Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. arXiv preprint arXiv:2310.00746 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 72,
+ 546,
+ 502,
+ 574
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 546,
+ 502,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 546,
+ 502,
+ 574
+ ],
+ "type": "text",
+ "content": "[140] Tianxin Wei, Bowen Jin, Ruirui Li, Hansi Zeng, Zhengyang Wang, Jianhui Sun, Qingyu Yin, Hanqing Lu, Suhang Wang, Jingrui He, et al. 2024. Towards unified multi-modal personalization: Large vision-language models for generative recommendation and beyond. arXiv preprint arXiv:2403.10667 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 72,
+ 576,
+ 502,
+ 595
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 576,
+ 502,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 576,
+ 502,
+ 595
+ ],
+ "type": "text",
+ "content": "[141] Robert Wetzker, Carsten Zimmermann, and Christian Bauchage. 2008. Analyzing social bookmarking systems: A del. icio. us cookbook. In Proceedings of the ECAI 2008 Mining Social Data Workshop. 26-30."
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 72,
+ 596,
+ 502,
+ 615
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 596,
+ 502,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 596,
+ 502,
+ 615
+ ],
+ "type": "text",
+ "content": "[142] Stanisław Wozniak, Bartlomiej Koptyra, Arkadiusz Janz, Przemysław Kazienko, and Jan Kocón. 2024. Personalized large language models. arXiv preprint arXiv:2402.09269 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 72,
+ 616,
+ 502,
+ 635
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 616,
+ 502,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 616,
+ 502,
+ 635
+ ],
+ "type": "text",
+ "content": "[143] Junde Wu, Jiayuan Zhu, Yunli Qi, Jingkun Chen, Min Xu, Filippo Menolascina, and Vicente Grau. 2024. Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 72,
+ 635,
+ 502,
+ 654
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 635,
+ 502,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 635,
+ 502,
+ 654
+ ],
+ "type": "text",
+ "content": "[144] Xuan Wu, Dong Zhou, Yu Xu, and Seamus Lawless. 2017. Personalized query expansion utilizing multi-relational social data. In 2017 12th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP). IEEE, 65-70."
+ }
+ ]
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 213,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 213,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 70,
+ 72,
+ 213,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "text",
+ "content": "23"
+ }
+ ]
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 22
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 108,
+ 98,
+ 539,
+ 654
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 29,
+ "blocks": [
+ {
+ "bbox": [
+ 108,
+ 98,
+ 539,
+ 126
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 98,
+ 539,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 98,
+ 539,
+ 126
+ ],
+ "type": "text",
+ "content": "[145] Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender Systems. 12-22."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 108,
+ 127,
+ 539,
+ 147
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 127,
+ 539,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 127,
+ 539,
+ 147
+ ],
+ "type": "text",
+ "content": "[146] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences 68, 2 (2025), 121101."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 108,
+ 148,
+ 539,
+ 167
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 148,
+ 539,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 148,
+ 539,
+ 167
+ ],
+ "type": "text",
+ "content": "[147] Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. 641-649."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 108,
+ 167,
+ 539,
+ 187
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 167,
+ 539,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 167,
+ 539,
+ 187
+ ],
+ "type": "text",
+ "content": "[148] Huatao Xu, Liying Han, Qirui Yang, Mo Li, and Mani Srivastava. 2024. Penetrative ai: Making llms comprehend the physical world. In Proceedings of the 25th International Workshop on Mobile Computing Systems and Applications. 1-7."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 108,
+ 188,
+ 539,
+ 206
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 188,
+ 539,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 188,
+ 539,
+ 206
+ ],
+ "type": "text",
+ "content": "[149] Hongyan Xu, Hongtao Liu, Pengfei Jiao, and Wenjun Wang. 2021. Transformer reasoning network for personalized review summarization. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1452-1461."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 108,
+ 207,
+ 539,
+ 226
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 207,
+ 539,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 207,
+ 539,
+ 226
+ ],
+ "type": "text",
+ "content": "[150] Xinchao Xu, Zhibin Gou, Wenquan Wu, Zheng-Yu Niu, Hua Wu, Haifeng Wang, and Shihang Wang. 2022. Long time no see! open-domain conversation with long-term persona memory. arXiv preprint arXiv:2203.05797 (2022)."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 108,
+ 227,
+ 539,
+ 247
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 227,
+ 539,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 227,
+ 539,
+ 247
+ ],
+ "type": "text",
+ "content": "[151] Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, and Tat-Seng Chua. 2025. Personalized Generation In Large Model Era: A Survey. arXiv preprint arXiv:2503.02614 (2025)."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 108,
+ 247,
+ 539,
+ 266
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 247,
+ 539,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 247,
+ 539,
+ 266
+ ],
+ "type": "text",
+ "content": "[152] Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. 2024. Personalized federated continual learning via multi-granularity prompt. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4023-4034."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 108,
+ 267,
+ 539,
+ 286
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 267,
+ 539,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 267,
+ 539,
+ 286
+ ],
+ "type": "text",
+ "content": "[153] Xiaoyan Yu, Tongxu Luo, Yifan Wei, Fangyu Lei, Yiming Huang, Hao Peng, and Liehuang Zhu. 2024. Neeko: Leveraging dynamic lora for efficient multi-character role-playing agent. arXiv preprint arXiv:2402.13717 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 108,
+ 287,
+ 539,
+ 306
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 287,
+ 539,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 287,
+ 539,
+ 306
+ ],
+ "type": "text",
+ "content": "[154] Xinfeng Yuan, Siyu Yuan, Yuhan Cui, Tianhe Lin, Xintao Wang, Rui Xu, Jiangjie Chen, and Deqing Yang. 2024. Evaluating character understanding of large language models via character profiling from fictional works. arXiv preprint arXiv:2404.12726 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 108,
+ 307,
+ 539,
+ 326
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 307,
+ 539,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 307,
+ 539,
+ 326
+ ],
+ "type": "text",
+ "content": "[155] Hansi Zeng, Surya Kallumadi, Zaid Alibadi, Rodrigo Nogueira, and Hamed Zamani. 2023. A personalized dense retrieval framework for unified information access. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 121-130."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 108,
+ 327,
+ 539,
+ 346
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 327,
+ 539,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 327,
+ 539,
+ 346
+ ],
+ "type": "text",
+ "content": "[156] Saber Zerhoudi and Michael Granitzer. 2024. PersonaRAG: Enhancing Retrieval-Augmented Generation Systems with User-Centric Agents. arXiv preprint arXiv:2407.09394 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 108,
+ 346,
+ 539,
+ 376
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 346,
+ 539,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 346,
+ 539,
+ 376
+ ],
+ "type": "text",
+ "content": "[157] Han Zhang, Songlin Wang, Kang Zhang, Zhiling Tang, Yunjiang Jiang, Yun Xiao, Weipeng Yan, and Wen-Yun Yang. 2020. Towards personalized and semantic retrieval: An end-to-end solution for e-commerce search via embedding learning. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2407-2416."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 108,
+ 376,
+ 473,
+ 386
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 376,
+ 473,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 376,
+ 473,
+ 386
+ ],
+ "type": "text",
+ "content": "[158] Jiarui Zhang. 2024. Guided profile generation improves personalization with llms. arXiv preprint arXiv:2409.13093 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 108,
+ 387,
+ 539,
+ 406
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 387,
+ 539,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 387,
+ 539,
+ 406
+ ],
+ "type": "text",
+ "content": "[159] Jesse Zhang, Jiahui Zhang, Karl Pertsch, Ziyi Liu, Xiang Ren, Minsuk Chang, Shao-Hua Sun, and Joseph J Lim. [n.d.]. Bootstrap Your Own Skills: Learning to Solve New Tasks with Large Language Model Guidance. In 7th Annual Conference on Robot Learning."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 108,
+ 407,
+ 539,
+ 426
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 407,
+ 539,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 407,
+ 539,
+ 426
+ ],
+ "type": "text",
+ "content": "[160] Kai Zhang, Yangyang Kang, Fubang Zhao, and Xiaozhong Liu. 2023. LLM-based medical assistant personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 108,
+ 426,
+ 539,
+ 445
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 426,
+ 539,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 426,
+ 539,
+ 445
+ ],
+ "type": "text",
+ "content": "[161] Kaiyan Zhang, Jianyu Wang, Ermo Hua, Biqing Qi, Ning Ding, and Bowen Zhou. 2024. Cogenesis: A framework collaborating large and small language models for secure context-aware instruction following. arXiv preprint arXiv:2403.03129 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 108,
+ 447,
+ 539,
+ 465
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 447,
+ 539,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 447,
+ 539,
+ 465
+ ],
+ "type": "text",
+ "content": "[162] Kai Zhang, Fubang Zhao, Yangyang Kang, and Xiaozhong Liu. 2023. Memory-augmented llm personalization with short-and long-term memory coordination. arXiv preprint arXiv:2309.11696 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 108,
+ 466,
+ 539,
+ 495
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 466,
+ 539,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 466,
+ 539,
+ 495
+ ],
+ "type": "text",
+ "content": "[163] Wenlin Zhang, Chuhan Wu, Xiangyang Li, Yuhao Wang, Kuicai Dong, Yichao Wang, Xinyi Dai, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2025. LLMTreeRec: Unleashing the Power of Large Language Models for Cold-Start Recommendations. In Proceedings of the 31st International Conference on Computational Linguistics. 886-896."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 108,
+ 496,
+ 539,
+ 515
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 496,
+ 539,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 496,
+ 539,
+ 515
+ ],
+ "type": "text",
+ "content": "[164] Yanyue Zhang, Yulan He, and Deyu Zhou. 2025. Rehearse With User: Personalized Opinion Summarization via Role-Playing based on Large Language Models. arXiv preprint arXiv:2503.00449 (2025)."
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 108,
+ 516,
+ 539,
+ 535
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 516,
+ 539,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 516,
+ 539,
+ 535
+ ],
+ "type": "text",
+ "content": "[165] You Zhang, Jin Wang, Liang-Chih Yu, Dan Xu, and Xuejie Zhang. 2024. Personalized LoRA for human-centered text understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 1958-19596."
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 108,
+ 536,
+ 539,
+ 555
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 536,
+ 539,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 536,
+ 539,
+ 555
+ ],
+ "type": "text",
+ "content": "[166] Yabin Zhang, Wenhui Yu, Erhan Zhang, Xu Chen, Lantao Hu, Peng Jiang, and Kun Gai. 2024. Recgpt: Generative personalized prompts for sequential recommendation via chatgpt training paradigm. arXiv preprint arXiv:2404.08675 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 108,
+ 556,
+ 539,
+ 575
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 556,
+ 539,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 556,
+ 539,
+ 575
+ ],
+ "type": "text",
+ "content": "[167] Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 108,
+ 576,
+ 539,
+ 595
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 576,
+ 539,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 576,
+ 539,
+ 595
+ ],
+ "type": "text",
+ "content": "[168] Zhehao Zhang, Ryan A Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, et al. 2024. Personalization of large language models: A survey. arXiv preprint arXiv:2411.00027 (2024)."
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 108,
+ 596,
+ 539,
+ 615
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 596,
+ 539,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 596,
+ 539,
+ 615
+ ],
+ "type": "text",
+ "content": "[169] Yi Zheng, Chongyang Ma, Kanle Shi, and Haibin Huang. 2023. Agents meet okr: An object and key results driven agent system with hierarchical self-collaboration and self-evaluation. arXiv preprint arXiv:2311.16542 (2023)."
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 108,
+ 616,
+ 539,
+ 635
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 616,
+ 539,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 616,
+ 539,
+ 635
+ ],
+ "type": "text",
+ "content": "[170] Hanxun Zhong, Zhicheng Dou, Yutao Zhu, Hongjin Qian, and Ji-Rong Wen. 2022. Less is more: Learning to refine dialogue history for personalized dialogue generation. arXiv preprint arXiv:2204.08128 (2022)."
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 108,
+ 635,
+ 539,
+ 654
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 635,
+ 539,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 635,
+ 539,
+ 654
+ ],
+ "type": "text",
+ "content": "[171] Wanjun Zhong, Duyu Tang, Jiahai Wang, Jian Yin, and Nan Duan. 2021. UserAdapter: Few-shot user learning in sentiment analysis. In Findings of the Association for Computational Linguistics: ACL-JJCNLP 2021. 1484-1488."
+ }
+ ]
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 72,
+ 281,
+ 81
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 72,
+ 539,
+ 80
+ ],
+ "type": "text",
+ "content": "X. Li and P. Jia, et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 658,
+ 328,
+ 666
+ ],
+ "type": "text",
+ "content": "24"
+ }
+ ]
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 23
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 72,
+ 98,
+ 503,
+ 177
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 6,
+ "blocks": [
+ {
+ "bbox": [
+ 72,
+ 98,
+ 503,
+ 116
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 98,
+ 503,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 98,
+ 503,
+ 116
+ ],
+ "type": "text",
+ "content": "[172] Dong Zhou, Séamus Lawless, and Vincent Wade. 2012. Improving search via personalized query expansion using social media. Information retrieval 15 (2012), 218-242."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 72,
+ 118,
+ 503,
+ 137
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 118,
+ 503,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 118,
+ 503,
+ 137
+ ],
+ "type": "text",
+ "content": "[173] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. 2022. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022)."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 72,
+ 138,
+ 502,
+ 156
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 138,
+ 502,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 138,
+ 502,
+ 156
+ ],
+ "type": "text",
+ "content": "[174] Yujia Zhou, Qiannan Zhu, Jiajie Jin, and Zhicheng Dou. 2024. Cognitive personalized search integrating large language models with an efficient memory mechanism. In Proceedings of the ACM Web Conference 2024. 1464-1473."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 72,
+ 158,
+ 502,
+ 177
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 72,
+ 158,
+ 502,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 72,
+ 158,
+ 502,
+ 177
+ ],
+ "type": "text",
+ "content": "[175] Yuchen Zhuang, Haotian Sun, Yue Yu, Rushi Qiang, Qifan Wang, Chao Zhang, and Bo Dai. [n.d.]. Hydra: Model factorization framework for black-box llm personalization, 2024. URL https://arxiv.org/abs/2406.02888 ([n.d.])."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 71,
+ 189,
+ 308,
+ 198
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 189,
+ 308,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 189,
+ 308,
+ 198
+ ],
+ "type": "text",
+ "content": "Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 71,
+ 72,
+ 212,
+ 81
+ ],
+ "type": "text",
+ "content": "A Survey of Personalization: From RAG to Agent"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 72,
+ 503,
+ 80
+ ],
+ "type": "text",
+ "content": "Conference acronym 'XX, June 03-05, 2018, Woodstock, NY"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 283,
+ 658,
+ 290,
+ 666
+ ],
+ "type": "text",
+ "content": "25"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 24
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10157/9641b7f1-5ad9-4487-b9c3-7734cccafedc_content_list.json b/data/2025/2504_10xxx/2504.10157/9641b7f1-5ad9-4487-b9c3-7734cccafedc_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..e1bd8f3b0da8f30090a2d3f893d11159c5b392c3
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/9641b7f1-5ad9-4487-b9c3-7734cccafedc_content_list.json
@@ -0,0 +1,2572 @@
+[
+ {
+ "type": "text",
+ "text": "SocioVerse: A World Model for Social Simulation Powered by LLM Agents and A Pool of 10 Million Real-World Users",
+ "text_level": 1,
+ "bbox": [
+ 196,
+ 122,
+ 802,
+ 195
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Xinnong Zhang $^{1,2\\dagger}$ , Jiayu Lin $^{1,2\\dagger}$ , Xinyi Mou $^{2\\dagger}$ , Shiyue Yang $^{2}$ , Xiawei Liu $^{2}$ , Libo Sun $^{2}$ , Hanjia Lyu $^{3}$ , Yihang Yang $^{2}$ , Weihong Qi $^{4}$ , Yue Chen $^{2}$ , Guanying Li $^{2}$ , Ling Yan $^{5}$ , Yao Hu $^{5}$ , Siming Chen $^{2}$ , Yu Wang $^{2}$ , Xuanjing Huang $^{2}$ , Jiebo Luo $^{3}$ , Shiping Tang $^{2}$ , Libo Wu $^{1,2}$ , Baohua Zhou $^{2}$ , Zhongyu Wei $^{1,2}$",
+ "bbox": [
+ 186,
+ 244,
+ 815,
+ 306
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "$^{1}$ Shanghai Innovation Institute, $^{2}$ Fudan University, $^{3}$ University of Rochester, $^{4}$ Indiana University, $^{5}$ Xiaohongshu Inc. zywei@fudan.edu.cn \nSocioVerse: https://github.com/FudanDISC/SocioVerse",
+ "bbox": [
+ 282,
+ 305,
+ 712,
+ 362
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/ee37a034cc7bfd54903875bca0d8ebdb66151851f1d8c1aae3fe11feeefe474d.jpg",
+ "image_caption": [
+ "Figure 1: An illustration of the SocioVerse in the case of Ukraine issue. The alignment challenges are well handled regarding environment, user, scenario, and behavior."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 183,
+ 378,
+ 802,
+ 627
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 459,
+ 674,
+ 537,
+ 690
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Social simulation is transforming traditional social science research by modeling human behavior through interactions between virtual individuals and their environments. With recent advances in large language models (LLMs), this approach has shown growing potential in capturing individual differences and predicting group behaviors. However, existing methods face alignment challenges related to the environment, target users, interaction mechanisms, and behavioral patterns. To this end, we introduce SocioVerse, an LLM-agent-driven world model for social simulation. Our framework features four powerful alignment components and a user pool of 10 million real individuals. To validate its effectiveness, we conducted large-scale simulation experiments across three distinct domains: politics, news, and economics. Results demonstrate that SocioVerse can reflect large-scale population dynamics while ensuring diversity, credibility, and representativeness through standardized procedures and minimal manual adjustments.",
+ "bbox": [
+ 228,
+ 704,
+ 767,
+ 886
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "aside_text",
+ "text": "arXiv:2504.10157v3 [cs.CL] 15 Jul 2025",
+ "bbox": [
+ 22,
+ 282,
+ 57,
+ 715
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_footnote",
+ "text": "These authors contribute equally to this work.",
+ "bbox": [
+ 192,
+ 896,
+ 470,
+ 911
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 Introduction",
+ "text_level": 1,
+ "bbox": [
+ 173,
+ 89,
+ 312,
+ 104
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The study of human behavior aims to understand how individuals and groups act in various social contexts and serves as a cornerstone of social science research. Traditionally, this has been accomplished using methods such as surveys, interviews, and observations [10, 18, 44]. However, these approaches often encounter challenges, including high costs, limited sample sizes, and ethical concerns. As a result, researchers have resorted to alternative methods for studying human behavior.",
+ "bbox": [
+ 169,
+ 132,
+ 826,
+ 202
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Social simulation has emerged as an effective method for addressing this issue, where researchers use agents to model human behavior, observe their reactions, and translate these findings into insights about human behavior [48, 50]. By assigning behavioral rules to autonomous agents, researchers can explore how micro-level decisions lead to emergent macro-level patterns through the agent-based models [11, 21]. This approach enables capturing specific groups' preferences on particular topics and forecasting potential social dynamics. Furthermore, recent advancements in large language models (LLMs) have significantly enhanced agents' reasoning and decision-making capabilities, enabling them to operate and interact within increasingly realistic and complex environments [3, 35, 37].",
+ "bbox": [
+ 169,
+ 207,
+ 823,
+ 319
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Recent studies have explored social simulation across various levels and scenarios, from mimicking well-known individuals and mirroring specific situations to modeling large-scale social dynamics [4, 29, 34, 36, 49, 60]. However, they share a common challenge: alignment between the simulated environment and the real world, which manifests across multiple dimensions and raises several key questions that remain to be addressed, as shown in Figure 1.",
+ "bbox": [
+ 169,
+ 325,
+ 826,
+ 395
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Q1. How to align the simulated environment with the real world?",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 401,
+ 629,
+ 414
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In the real world, new events occur every day and new content is continuously generated. The behavior of real users is rooted in these ever-evolving social contexts and policy agendas. However, the static knowledge of LLMs prevents them from aligning with the dynamic nature of the real-world social environment [2, 15]. There is a gap between the simulated context and the real world, which results in discrepancies between the simulation process and outcomes compared to those in reality. Therefore, it is necessary to establish an update mechanism to keep the simulated environment synchronized with the real world.",
+ "bbox": [
+ 169,
+ 415,
+ 826,
+ 512
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Q2. How to align simulated agents with target users precisely?",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 518,
+ 609,
+ 532
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The composition of users in the real world is both complex and diverse, making it impractical to enumerate all users in every scenario. Therefore, it is essential to identify target users whose distribution aligns with that of the users in the corresponding scenario, thereby accurately reflecting the real-world composition and relationships [17, 45]. Based on this, precise target user simulation also requires providing agents with a detailed and comprehensive description of the corresponding users, often involving the integration of high-fidelity demographic, contextual, and behavioral data.",
+ "bbox": [
+ 169,
+ 532,
+ 823,
+ 614
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Q3. How to align the interaction mechanism with the real world among different scenarios?",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 621,
+ 810,
+ 635
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The diversity of social interactions presents challenges in social simulation design, requiring deliberate choices regarding the number of individuals, social structures, interaction patterns, and message dissemination mechanisms, to align with the real world. This often results in independently constructed task-specific simulation pipelines performing repetitive work, which reduces their generalizability and scalability [26, 58]. Therefore, there is a need for unified simulation frameworks based on systematic categorization to standardize simulation components and facilitate extensibility across different social scenarios.",
+ "bbox": [
+ 169,
+ 635,
+ 826,
+ 732
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Q4. How to align the behavioral pattern with the real-world groups?",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 738,
+ 651,
+ 753
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "When the environment perceived by agents, the user composition, and the interaction mechanisms are aligned with the real world, agents are expected to exhibit responses consistent with those of the corresponding real users. However, current LLMs exhibit inherent bias and limitations in such reasoning, failing to infer different types of user behaviors [16, 60]. Therefore, it is necessary to systematically collect behavior-driving factors across different user characteristics and adopt appropriate modeling approaches to effectively capture diverse behavior patterns.",
+ "bbox": [
+ 169,
+ 752,
+ 825,
+ 835
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In this paper, we propose SocioVerse, a world model for social simulation driven by LLM-based agents based on a large-scale real-world user pool. As shown in Figure 2, we design modular components to address the above questions. The Social Environment injects up-to-date and external real-world information into the simulation. The User Engine and Scenario Engine respectively reconstruct realistic user context and orchestrate the simulation process to align the simulation with",
+ "bbox": [
+ 169,
+ 842,
+ 825,
+ 912
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "2",
+ "bbox": [
+ 493,
+ 935,
+ 504,
+ 946
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/af88ad03ac5ce1c47c53dd172206932141236f13f7154e620648f8b4af53663e.jpg",
+ "image_caption": [
+ "Figure 2: An illustration of SocioVerse framework involving 4 powerful parts. The social environment provides an updated context for the simulation. During the simulation, the behavior engine takes the simulation setting, user profiles, and social information from the scenario engine, user engine, and social environment, respectively, and generates the results according to the query."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 183,
+ 97,
+ 808,
+ 398
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "the real world. Given this rich contextual setup, the Behavior Engine then drives agents to reproduce human behaviors accordingly.",
+ "bbox": [
+ 169,
+ 508,
+ 823,
+ 539
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "To support the framework, we construct a user pool of 10 million individuals by collecting real-world social media data to power the user engine. Comparable in scale to the entire populations of Hungary or Greece, this extensive pool enables diverse and large-scale social simulations. For any customized simulation task, various sampling strategies can be applied to extract target user groups from the pool to support the simulation process.",
+ "bbox": [
+ 169,
+ 542,
+ 823,
+ 614
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "We conduct three simulations using the SocioVerse framework, each differing in research domain, user composition, and social environment: (a) presidential election prediction, (b) breaking news feedback, and (c) national economic survey. For each task, we compare the simulation results with real-world situations. Extensive and comprehensive experiments demonstrate that our framework serves as a robust foundation for building standardized and accurate large-scale social simulations. In summary, our key contributions are as follows:",
+ "bbox": [
+ 169,
+ 619,
+ 826,
+ 703
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- SocioVerse: We propose a world model for social simulation comprising four powerful alignment modules, enabling diverse and trustworthy social simulations (as illustrated in Figure 2).",
+ "- 10M User Pool: A user pool of 10 million individuals, constructed from real-world behavioral data, enables large-scale and diverse social simulations, ranging from small interest groups to large citizen communities.",
+ "- Three Illustrative Simulations: We demonstrate the framework's capabilities through three distinct scenarios: presidential election prediction, breaking news feedback, and a national economic survey, providing a foundation for future research."
+ ],
+ "bbox": [
+ 215,
+ 727,
+ 826,
+ 891
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "3",
+ "bbox": [
+ 493,
+ 935,
+ 503,
+ 946
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2 Methods",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 89,
+ 279,
+ 104
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Overall Framework",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 121,
+ 316,
+ 135
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "The SocioVerse framework follows a structured pipeline to achieve realistic social simulation results, as shown in Figure 2: (1) Social Environment collects updated information and contextual knowledge. Within the simulation environment, (2) User Engine aligns the simulated agents with target users, (3) Scenario Engine aligns the interaction structure with diverse scenarios, and (4) Behavior Engine aligns the behavioral pattern with real-world target groups.",
+ "bbox": [
+ 169,
+ 148,
+ 826,
+ 219
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.1 Social Environment",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 237,
+ 351,
+ 250
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Function The social environment provides event-related context to align the simulation environment with real-world conditions. By integrating up-to-date events, social statistics, and preference content into LLM-based agents, it enhances the realism of the simulation and improve agent decision-making.",
+ "bbox": [
+ 169,
+ 263,
+ 826,
+ 306
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Components The social environment should encompass as much real-world social, cultural, and technological context as possible. It can be broadly categorized into three types: social structural information, social dynamic information, and personalized context.",
+ "bbox": [
+ 169,
+ 321,
+ 823,
+ 364
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Social Structure: Social structural information provides agents with a rich knowledge base encompassing demographic distributions, cultural norms, urban infrastructures, and collective behavior patterns [57]. This data allows agents to behave in a way that aligns with the typical characteristics of their assigned demographic or geographic profile. For example, by incorporating regional dialect preferences, work-life habits, and common social values, the simulation can more accurately reflect public discourse trends, mobility behaviors, and economic interactions.",
+ "bbox": [
+ 169,
+ 369,
+ 825,
+ 455
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Social Dynamics: Social dynamics encompass time-sensitive content continuously generated in the real world, such as news events and policy changes. Typically, this engine maintains an up-to-date event base to continuously collect real-world event news from mainstream news, and all the news articles contain time stamps and event-related tags so that LLM-based agents can comb through the timeline of the events and react accordingly [37].",
+ "bbox": [
+ 169,
+ 460,
+ 823,
+ 531
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Personalized Context: In addition to the macro social environment, individuals also receive different personalized information feeds. Previous studies have explored that the recommendation system can enhance the behavior diversity of the agent [31, 56, 60]. Consequently, the preference content component constructs relevant posts and pushes them to agents according to their social interaction network and interesting topics.",
+ "bbox": [
+ 169,
+ 536,
+ 825,
+ 606
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.2 User Engine",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 625,
+ 299,
+ 638
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Function The user engine aligns simulated agents with a rich set of real-world user samples, enabling the creation of complex target users within the simulation.",
+ "bbox": [
+ 169,
+ 650,
+ 826,
+ 680
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Components To support diverse user composition and effective user retrieval and description, the user engine incorporates a large user pool and a wide range of user labels.",
+ "bbox": [
+ 169,
+ 696,
+ 823,
+ 724
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "User Pools: The user pool is designed to collect extensive digital footprints of individuals across social media platforms, enabling a more comprehensive characterization of real-world behavioral patterns and expression tendencies. To this end, we constructed a user pool covering a variety of social media platforms, including $\\mathbf{X}^1$ and Rednote| Q01 | Voting Behavior |
| Question | ORDER OF MAJOR PARTY CANDIDATE NAMES |
| Value Labels | 1. Democrat first / Republican second2. Republican first / Democrat second |
| Q02 | Social Security |
| Question | Next I am going to read you a list of federal programs. For each one, I would like you to tell me whether you would like to see spending increased, decreased, or kept the same.What about Social Security? Should federal spending on Social Security be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q03 | Education |
| Question | What about public schools? Should federal spending on public schools be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q04 | Immigration |
| Question | What about tightening border security to prevent illegal immigration? Should federal spending on tightening border security to prevent illegal immigration be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q05 | Criminal Justice |
| Question | What about dealing with crime? Should federal spending on dealing with crime be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q06 | Social Welfare |
| Question | What about welfare programs? Should federal spending on welfare programs be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q07 | Infrastructure |
| Question | What about building and repairing highways? Should federal spending on building and repairing highways be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF\n1. Increased\n2. Decreased\n3. Kept the same |
| Q08 | Aid to Poor |
| Question | What about aid to the poor? Should federal spending on aid to the poor be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF\n1. Increased\n2. Decreased\n3. Kept the same |
| Q09 | Environment |
| Question | What about protecting the environment? Should federal spending on protecting the environment be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF\n1. Increased\n2. Decreased\n3. Kept the same |
| Q10 | Government |
| Question | How much do you feel that having elections makes the government pay attention to what the people think? |
| Value Labels | -2. DK/RF\n1. A good deal\n2. Some\n3. Not much |
| Q11 | Economy |
| Question | Which party do you think would do a better job of handling the nation's economy? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q12 | Health Care |
| Question | Which party do you think would do a better job of handling health care? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q13 | Immigration |
| Question | Which party do you think would do a better job of handling immigration? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q14 | Taxes |
| Question | Which party do you think would do a better job of handling taxes? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q15 | Environment |
| Question | Which party do you think would do a better job of handling the environment? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q16 | Education |
| Question | Some people think the government should provide fewer services even in areas such as health and education in order to reduce spending.\nOther people feel it is important for the government to provide many more services even if it means an increase in spending.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Government should provide fewer services\n2. Neutral\n3. Government should provide more services |
| Q17 | Defense |
| Question | Some people believe that we should spend less money for defense.\nOthers feel that defense spending should be increased.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Decrease defense spending\n2. Neutral\n3. Increase defense spending |
| Q18 | Health Care |
| Question | There is much concern about the rapid rise in medical and hospital costs.\nSome people feel there should be a government insurance plan which would cover all medical and hospital expenses for everyone.\nOthers feel that all medical expenses should be paid by individuals through private insurance plans like Blue Cross or other company paid plans.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Government insurance plan\n2. Neutral\n3. Private insurance plan |
| Q19 | Social Welfare |
| Question | Some people feel the government in Washington should see to it that every person has a job and a good standard of living.\nOthers think the government should just let each person get ahead on their own.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Government should see to jobs and standard of living\n2. Neutral\n3. Government should let each person get ahead on own |
| Q20 | Aid to Blacks |
| Question | Some people feel that the government in Washington should make every effort to improve the social and economic position of blacks.\nOthers feel that the government should not make any special effort to help blacks because they should help themselves.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Government should help blacks\n2. Neutral\n3. Blacks should help themselves |
| Q21 | Environment |
| Question | Some people think we need much tougher government regulations on business in order to protect the environment.\nOthers think that current regulations to protect the environment are already too much of a burden on business.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Tougher regulations on business needed to protect environment\n2. Neutral\n3. Regulations to protect environment already too much a burden on business |
| Q22 | Abortion |
| Question | Would you be pleased, upset, or neither pleased nor upset if the Supreme Court reduced abortion rights? |
| Value Labels | -2. DK/RF\n1. Pleased\n2. Upset\n3. Neither pleased nor upset |
| Q23 | Criminal Justice |
| Question | Do you favor or oppose the death penalty for persons convicted of murder? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose |
| Q24 | US Position in World |
| Question | Do you agree or disagree with this statement: ‘This country would be better off if we just stayed home and did not concern ourselves with problems in other parts of the world.’ |
| Value Labels | -2. DK/RF\n1. Agree\n2. Disagree |
| Q25 | US Position in World |
| Question | How willing should the United States be to use military force to solve international problems? |
| Value Labels | -2. DK/RF\n1. Willing\n2. Moderately willing\n3. Not willing |
| Q26 | Inequality |
| Question | Do you think the difference in incomes between rich people and poor people in the United States today is larger, smaller, or about the same as it was 20 years ago? |
| Value Labels | -2. DK/RF\n1. Larger\n2. Smaller\n3. About the same |
| Q27 | Environment |
| Question | Do you think the federal government should be doing more about rising temperatures, should be doing less, or is it currently doing the right amount? |
| Value Labels | -2. DK/RF\n1. Should be doing more\n2. Should be doing less\n3. Is currently doing the right amount |
| Q28 | Parental Leave |
| Question | Do you favor, oppose, or neither favor nor oppose requiring employers to offer paid leave to parents of new children? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q29 | LGBTQ+ Rights |
| Question | Do you think business owners who provide wedding-related services should be allowed to refuse services to same-sex couples if same-sex marriage violates their religious beliefs, or do you think business owners should be required to provide services regardless of a couple's sexual orientation? |
| Value Labels | -2. DK/RF\n1. Should be allowed to refuse\n2. Should be required to provide services |
| Q30 | LGBTQ+ Rights |
| Question | Should transgender people - that is, people who identify themselves as the sex or gender different from the one they were born as - have to use the bathrooms of the gender they were born as, or should they be allowed to use the bathrooms of their identified gender? |
| Value Labels | -2. DK/RF\n1. Have to use the bathrooms of the gender they were born as\n2. Be allowed to use the bathrooms of their identified gender |
| Q31 | LGBTQ+ Rights |
| Question | Do you favor or oppose laws to protect gays and lesbians against job discrimination? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose |
| Q32 | LGBTQ+ Rights |
| Question | Do you think gay or lesbian couples should be legally permitted to adopt children? |
| Value Labels | -2. DK/RF\n1. Yes\n2. No |
| Q33 | LGBTQ+ Rights |
| Question | Which comes closest to your view? You can just tell me the number of your choice. |
| Value Labels | -2. DK/RF 1. Gay and lesbian couples should be allowed to legally marry\n2. Gay and lesbian couples should be allowed to form civil unions but not legally marry\n3. There should be no legal recognition of gay or lesbian couples' relationship |
| Q34 | Immigration |
| Question | Some people have proposed that the U.S. Constitution should be changed so that the children of unauthorized immigrants do not automatically get citizenship if they are born in this country.\nDo you favor, oppose, or neither favor nor oppose this proposal? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q35 | Immigration |
| Question | What should happen to immigrants who were brought to the U.S. illegally as children and have lived here for at least 10 years and graduated high school here? Should they be sent back where they came from, or should they be allowed to live and work in the United States? |
| Value Labels | -2. DK/RF\n1. Should be sent back where they came from\n2. Should be allowed to live and work in the US |
| Q36 | Immigration |
| Question | Do you favor, oppose, or neither favor nor oppose building a wall on the U.S. border with Mexico? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q37 | Unrest |
| Question | During the past few months, would you say that most of the actions taken by protestors to get the things they want have been violent, or have most of these actions by protestors been peaceful, or have these actions been equally violent and peaceful? |
| Value Labels | -2. DK/RF\n1. Mostly violent\n2. Mostly peaceful\n3. Equally violent and peaceful |
| Q38 | Government |
| Question | Do you think it is better when one party controls both the presidency and Congress, better when control is split between the Democrats and Republicans, or doesn’t it matter? |
| Value Labels | -2. DK/RF\n1. Better when one party controls both\n2. Better when control is split\n3. It doesn’t matter |
| Q39 | Government |
| Question | Would you say the government is pretty much run by a few big interests looking out for themselves or that it is run for the benefit of all the people? |
| Value Labels | -2. DK/RF\n1. Run by a few big interests\n2. For the benefit of all the people |
| Q40 | Government |
| Question | Do you think that people in government waste a lot of the money we pay in taxes, waste some of it, or don’t waste very much of it? |
| Value Labels | -2. DK/RF\n1. Waste a lot\n2. Waste some\n3. Don’t waste very much |
| Q41 | Election Integrity |
| Question | Do you favor, oppose, or neither favor nor oppose allowing convicted felons to vote once they complete their sentence? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q42 | Democratic Norms |
| Question | How important is it that news organizations are free to criticize political leaders? |
| Value Labels | -2. DK/RF\n1. Not important\n2. Moderately important\n3. Important |
| Q43 | Democratic Norms |
| Question | How important is it that the executive, legislative, and judicial branches of government keep one another from having too much power? |
| Value Labels | -2. DK/RF\n1. Not important\n2. Moderately important\n3. Important |
| Q44 | Democratic Norms |
| Question | How important is it that elected officials face serious consequences if they engage in misconduct? |
| Value Labels | -2. DK/RF\n1. Not important\n2. Moderately important\n3. Important |
| Q45 | Democratic Norms |
| Question | How important is it that people agree on basic facts even if they disagree politically? |
| Value Labels | -2. DK/RF\n1. Not important\n2. Moderately important\n3. Important |
| Q46 | Democratic Norms |
| Question | Would it be helpful, harmful, or neither helpful nor harmful if U.S. presidents could work on the country’s problems without paying attention to what Congress and the courts say? |
| Value Labels | -2. DK/RF\n1. Helpful\n2. Harmful\n3. Neither helpful nor harmful |
| Q47 | Democratic Norms |
| Question | Do you favor, oppose, or neither favor nor oppose elected officials restricting journalists’ access to information about government decision-making? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q48 | Gender Resentment |
| Question | ‘Many women interpret innocent remarks or acts as being sexist.’\nDo you agree, neither agree nor disagree, or disagree with this statement? |
| Value Labels | -2. DK/RF/technical error\n1. Agree\n2. Neither agree nor disagree\n3. Disagree |
| Q49 | Gender Resentment |
| Question | ‘Women seek to gain power by getting control over men.’\nDo you agree, neither agree nor disagree, or disagree with this statement? |
| Value Labels | -2. DK/RF/technical error\n1. Agree\n2. Neither agree nor disagree\n3. Disagree |
",
+ "bbox": [
+ 173,
+ 183,
+ 857,
+ 900
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "page_number",
+ "text": "20",
+ "bbox": [
+ 488,
+ 935,
+ 509,
+ 946
+ ],
+ "page_idx": 19
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 87,
+ 856,
+ 859
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "page_number",
+ "text": "21",
+ "bbox": [
+ 488,
+ 935,
+ 506,
+ 946
+ ],
+ "page_idx": 20
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 89,
+ 856,
+ 898
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "page_number",
+ "text": "22",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 21
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 89,
+ 856,
+ 901
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "page_number",
+ "text": "23",
+ "bbox": [
+ 488,
+ 935,
+ 506,
+ 946
+ ],
+ "page_idx": 22
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 89,
+ 856,
+ 898
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "page_number",
+ "text": "24",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 87,
+ 856,
+ 858
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "page_number",
+ "text": "25",
+ "bbox": [
+ 488,
+ 935,
+ 506,
+ 946
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 87,
+ 857,
+ 872
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "page_number",
+ "text": "26",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 25
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 89,
+ 857,
+ 378
+ ],
+ "page_idx": 26
+ },
+ {
+ "type": "text",
+ "text": "D.2 Questionnaire for Breaking News Feedback",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 401,
+ 519,
+ 416
+ ],
+ "page_idx": 26
+ },
+ {
+ "type": "table",
+ "img_path": "images/7ce59a581d9686e1bdc3056be8b7786b1ee7c8b56671e62f94b9629946ea1c5e.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Q01 | Public Cognition (PC) |
| Question | I have heard of ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q02 | Public Cognition (PC) |
| Question | Many people around me use ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q03 | Public Cognition (PC) |
| Question | I have a deep understanding of ChatGPT's functions and applications. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q04 | Perceived Risks (PR) |
| Question | ChatGPT may lead to the widespread dissemination of false information. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q05 | Perceived Risks (PR) |
| Question | ChatGPT may reduce human thinking ability and creativity. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q06 | Perceived Risks (PR) |
| Question | The development of ChatGPT may replace certain jobs, and I am deeply concerned about this. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q07 | Perceived Benefits (PB) |
| Question | ChatGPT will definitely improve my work and study efficiency. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q08 | Perceived Benefits (PB) |
| Question | ChatGPT helps broaden my knowledge and provides me with new perspectives and ideas. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q09 | Perceived Benefits (PB) |
| Question | ChatGPT promotes technological innovation and development in related fields. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q10 | Trust (TR) |
| Question | I fully trust the team developing ChatGPT to manage and guide its development responsibly. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q11 | Trust (TR) |
| Question | I have strong confidence in the accuracy and reliability of the information generated by ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q12 | Trust (TR) |
| Question | I believe that the future application of ChatGPT will be effectively regulated. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q13 | Fairness (FA) |
| Question | The opportunities to use ChatGPT are distributed fairly among different groups of people. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q14 | Fairness (FA) |
| Question | I find the distribution of benefits brought by ChatGPT to be fair. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q15 | Fairness (FA) |
| Question | I believe that the decision-making process for the development and promotion of ChatGPT is fully transparent and adequately reflects public interests. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q16 | Public Acceptance (PA) |
| Question | Overall, I strongly welcome the emergence of ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q17 | Public Acceptance (PA) |
| Question | I am definitely willing to use ChatGPT in my work or studies. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q18 | Public Acceptance (PA) |
| Question | I strongly support increased investment in the research and development of AI technologies like ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
",
+ "bbox": [
+ 173,
+ 433,
+ 857,
+ 910
+ ],
+ "page_idx": 26
+ },
+ {
+ "type": "page_number",
+ "text": "27",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 26
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 171,
+ 87,
+ 856,
+ 843
+ ],
+ "page_idx": 27
+ },
+ {
+ "type": "page_number",
+ "text": "28",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 27
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 171,
+ 87,
+ 857,
+ 893
+ ],
+ "page_idx": 28
+ },
+ {
+ "type": "page_number",
+ "text": "29",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 28
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 88,
+ 854,
+ 199
+ ],
+ "page_idx": 29
+ },
+ {
+ "type": "text",
+ "text": "D.3 Questionnaire for National Economic Survey",
+ "text_level": 1,
+ "bbox": [
+ 171,
+ 218,
+ 529,
+ 234
+ ],
+ "page_idx": 29
+ },
+ {
+ "type": "table",
+ "img_path": "images/18070cfafc0e2c3f8693e46f9ae0e339394f7c42e0f7d13109ea8126d6c4d3ff.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Q01 | Food |
| Question | What is your average monthly expenditure on food (including dining out)? (Unit: CNY) |
| Value Labels | A. Below 500 CNY\nB. 501-650 CNY\nC. 651-800 CNY\nD. 801-1000 CNY\nE. Above 1000 CNY |
| Q02 | Food |
| Question | Do you think your current spending on food, tobacco, and alcohol is too high relative to your income? |
| Value Labels | A. Yes\nB. No\nC. Acceptable |
| Q03 | Clothing |
| Question | What is your average monthly expenditure on clothing (including apparel, shoes, and accessories)? (Unit: CNY) |
| Value Labels | A. Below 50 CNY\nB. 51-100 CNY\nC. 101-150 CNY\nD. 151-200 CNY\nE. Above 200 CNY |
| Q04 | Clothing |
| Question | How much economic pressure do you feel from clothing expenses? |
| Value Labels | A. Very low, almost no pressure\nB. Moderate, some pressure but manageable\nC. High, requires careful spending\nD. Very high, affects spending in other areas |
| Q05 | Household |
| Question | What is your average monthly housing expenditure? (Including rent, mortgage, property fees, maintenance, etc.) (Unit: CNY) |
| Value Labels | A. Below 200 CNY\nB. 201-500 CNY\nC. 501-800 CNY\nD. 801-1200 CNY\nE. Above 1200 CNY |
| Q06 | Household |
| Question | What percentage of your monthly income is spent on housing? (Including rent, mortgage, property fees, maintenance, etc.) |
| Value Labels | A. Below 10% \nB. 10%-20% \nC. 21%-30% \nD. 31%-40% \nE. Above 40% |
| Q07 | Daily Service |
| Question | What is your average monthly expenditure on daily necessities (personal care, house- hold items, cleaning supplies, etc.) and services (housekeeping, repairs, beauty, pet services, etc.)? (Unit: CNY) |
| Value Labels | A. Below 80 CNY \nB. 81-120 CNY \nC. 121-160 CNY \nD. 161-200 CNY \nE. Above 200 CNY |
| Q08 | Transportation & Communication |
| Question | What is your average monthly expenditure on transportation (public transport, taxis, fuel, parking, etc.) and communication (mobile and internet fees)? (Unit: CNY) |
| Value Labels | A. Below 200 CNY \nB. 201-300 CNY \nC. 301-400 CNY \nD. 401-500 CNY \nE. Above 500 CNY |
| Q09 | Education & Entertainment |
| Question | What is your average monthly expenditure on education (tuition, training, books, etc.) and cultural entertainment (movies, performances, games, fitness, cultural activities, etc.)? (Unit: CNY) |
| Value Labels | A. Below 100 CNY \nB. 101-200 CNY \nC. 201-300 CNY \nD. 301-400 CNY \nE. Above 400 CNY |
| Q10 | Education & Entertainment |
| Question | Can you easily afford your current education, cultural, and entertainment expenses? |
| Value Labels | A. Yes, spending does not affect other areas \nB. Barely, needs some control \nC. Not really, affects other expenditures \nD. No, it creates significant financial pressure |
| Q11 | Medical |
| Question | What is your average monthly expenditure on healthcare (medications, medical services, health management, etc.)? (Unit: CNY) |
| Value Labels | A. Below 100 CNY \nB. 101-200 CNY \nC. 201-300 CNY \nD. 301-400 CNY \nE. Above 400 CNY |
| Q12 | Medical |
| Question | Have you purchased private medical or health insurance for yourself or your family? |
| Value Labels | A. Yes \nB. Not yet, but planning to \nC. No, and no plans to |
| Q13 | Others |
| Question | Besides food, clothing, housing, daily necessities and services, transportation, education, culture, and healthcare, what is your average monthly expenditure on other areas (e.g., hobbies, charitable donations, investment, etc.)? (Unit: CNY) |
| Value Labels | A. Below 30 CNY\nB. 31-60 CNY\nC. 61-90 CNY\nD. 91-120 CNY\nE. Above 120 CNY |
| Q14 | Overall |
| Question | How would you evaluate the impact of your current consumption level on your household (or personal) financial situation? |
| Value Labels | A. Comfortable, can moderately increase spending\nB. Average, can maintain current spending\nC. Tight, need to control or reduce spending\nD. Very tight, affects quality of life |
| Q15 | Overall |
| Question | Do you feel that your consumption pressure is too high relative to your income level? |
| Value Labels | A. Yes\nB. No\nC. Not sure |
| Q16 | Overall |
| Question | If your income increases, which consumption areas would you most like to expand or improve? (Multiple choices allowed) |
| Value Labels | A. Food and alcohol\nB. Clothing\nC. Housing\nD. Daily necessities and services\nE. Transportation and communication\nF. Education, culture, and entertainment\nG. Healthcare\nH. Other goods and services |
| Q17 | Overall |
| Question | What is your consumption expectation for the next six months to a year? |
| Value Labels | A. Will continue to increase\nB. Will remain roughly the same\nC. Will moderately decrease\nD. Uncertain |
",
+ "bbox": [
+ 173,
+ 250,
+ 857,
+ 887
+ ],
+ "page_idx": 29
+ },
+ {
+ "type": "page_number",
+ "text": "30",
+ "bbox": [
+ 488,
+ 935,
+ 509,
+ 946
+ ],
+ "page_idx": 29
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 89,
+ 856,
+ 905
+ ],
+ "page_idx": 30
+ },
+ {
+ "type": "page_number",
+ "text": "31",
+ "bbox": [
+ 488,
+ 935,
+ 506,
+ 946
+ ],
+ "page_idx": 30
+ },
+ {
+ "type": "table",
+ "img_path": "",
+ "table_caption": [],
+ "table_footnote": [],
+ "bbox": [
+ 173,
+ 87,
+ 856,
+ 678
+ ],
+ "page_idx": 31
+ },
+ {
+ "type": "page_number",
+ "text": "32",
+ "bbox": [
+ 488,
+ 935,
+ 508,
+ 946
+ ],
+ "page_idx": 31
+ }
+]
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10157/9641b7f1-5ad9-4487-b9c3-7734cccafedc_model.json b/data/2025/2504_10xxx/2504.10157/9641b7f1-5ad9-4487-b9c3-7734cccafedc_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..2b89436ae459359bf30c0f66e9505a522e7da852
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/9641b7f1-5ad9-4487-b9c3-7734cccafedc_model.json
@@ -0,0 +1,3289 @@
+[
+ [
+ {
+ "type": "aside_text",
+ "bbox": [
+ 0.023,
+ 0.283,
+ 0.058,
+ 0.717
+ ],
+ "angle": 270,
+ "content": "arXiv:2504.10157v3 [cs.CL] 15 Jul 2025"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.197,
+ 0.123,
+ 0.803,
+ 0.196
+ ],
+ "angle": 0,
+ "content": "SocioVerse: A World Model for Social Simulation Powered by LLM Agents and A Pool of 10 Million Real-World Users"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.187,
+ 0.245,
+ 0.816,
+ 0.307
+ ],
+ "angle": 0,
+ "content": "Xinnong Zhang\\(^{1,2\\dagger}\\), Jiayu Lin\\(^{1,2\\dagger}\\), Xinyi Mou\\(^{2\\dagger}\\), Shiyue Yang\\(^{2}\\), Xiawei Liu\\(^{2}\\), Libo Sun\\(^{2}\\), Hanjia Lyu\\(^{3}\\), Yihang Yang\\(^{2}\\), Weihong Qi\\(^{4}\\), Yue Chen\\(^{2}\\), Guanying Li\\(^{2}\\), Ling Yan\\(^{5}\\), Yao Hu\\(^{5}\\), Siming Chen\\(^{2}\\), Yu Wang\\(^{2}\\), Xuanjing Huang\\(^{2}\\), Jiebo Luo\\(^{3}\\), Shiping Tang\\(^{2}\\), Libo Wu\\(^{1,2}\\), Baohua Zhou\\(^{2}\\), Zhongyu Wei\\(^{1,2}\\)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.283,
+ 0.306,
+ 0.714,
+ 0.363
+ ],
+ "angle": 0,
+ "content": "\\(^{1}\\)Shanghai Innovation Institute, \\(^{2}\\)Fudan University, \\(^{3}\\)University of Rochester, \\(^{4}\\)Indiana University, \\(^{5}\\)Xiaohongshu Inc. zywei@fudan.edu.cn \nSocioVerse: https://github.com/FudanDISC/SocioVerse"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.184,
+ 0.379,
+ 0.803,
+ 0.628
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.171,
+ 0.631,
+ 0.825,
+ 0.66
+ ],
+ "angle": 0,
+ "content": "Figure 1: An illustration of the SocioVerse in the case of Ukraine issue. The alignment challenges are well handled regarding environment, user, scenario, and behavior."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.46,
+ 0.675,
+ 0.538,
+ 0.691
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.23,
+ 0.705,
+ 0.769,
+ 0.887
+ ],
+ "angle": 0,
+ "content": "Social simulation is transforming traditional social science research by modeling human behavior through interactions between virtual individuals and their environments. With recent advances in large language models (LLMs), this approach has shown growing potential in capturing individual differences and predicting group behaviors. However, existing methods face alignment challenges related to the environment, target users, interaction mechanisms, and behavioral patterns. To this end, we introduce SocioVerse, an LLM-agent-driven world model for social simulation. Our framework features four powerful alignment components and a user pool of 10 million real individuals. To validate its effectiveness, we conducted large-scale simulation experiments across three distinct domains: politics, news, and economics. Results demonstrate that SocioVerse can reflect large-scale population dynamics while ensuring diversity, credibility, and representativeness through standardized procedures and minimal manual adjustments."
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.193,
+ 0.897,
+ 0.472,
+ 0.912
+ ],
+ "angle": 0,
+ "content": "These authors contribute equally to this work."
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.174,
+ 0.09,
+ 0.313,
+ 0.106
+ ],
+ "angle": 0,
+ "content": "1 Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.133,
+ 0.827,
+ 0.203
+ ],
+ "angle": 0,
+ "content": "The study of human behavior aims to understand how individuals and groups act in various social contexts and serves as a cornerstone of social science research. Traditionally, this has been accomplished using methods such as surveys, interviews, and observations [10, 18, 44]. However, these approaches often encounter challenges, including high costs, limited sample sizes, and ethical concerns. As a result, researchers have resorted to alternative methods for studying human behavior."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.208,
+ 0.825,
+ 0.32
+ ],
+ "angle": 0,
+ "content": "Social simulation has emerged as an effective method for addressing this issue, where researchers use agents to model human behavior, observe their reactions, and translate these findings into insights about human behavior [48, 50]. By assigning behavioral rules to autonomous agents, researchers can explore how micro-level decisions lead to emergent macro-level patterns through the agent-based models [11, 21]. This approach enables capturing specific groups' preferences on particular topics and forecasting potential social dynamics. Furthermore, recent advancements in large language models (LLMs) have significantly enhanced agents' reasoning and decision-making capabilities, enabling them to operate and interact within increasingly realistic and complex environments [3, 35, 37]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.326,
+ 0.827,
+ 0.396
+ ],
+ "angle": 0,
+ "content": "Recent studies have explored social simulation across various levels and scenarios, from mimicking well-known individuals and mirroring specific situations to modeling large-scale social dynamics [4, 29, 34, 36, 49, 60]. However, they share a common challenge: alignment between the simulated environment and the real world, which manifests across multiple dimensions and raises several key questions that remain to be addressed, as shown in Figure 1."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.402,
+ 0.63,
+ 0.415
+ ],
+ "angle": 0,
+ "content": "Q1. How to align the simulated environment with the real world?"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.416,
+ 0.827,
+ 0.513
+ ],
+ "angle": 0,
+ "content": "In the real world, new events occur every day and new content is continuously generated. The behavior of real users is rooted in these ever-evolving social contexts and policy agendas. However, the static knowledge of LLMs prevents them from aligning with the dynamic nature of the real-world social environment [2, 15]. There is a gap between the simulated context and the real world, which results in discrepancies between the simulation process and outcomes compared to those in reality. Therefore, it is necessary to establish an update mechanism to keep the simulated environment synchronized with the real world."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.519,
+ 0.611,
+ 0.533
+ ],
+ "angle": 0,
+ "content": "Q2. How to align simulated agents with target users precisely?"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.533,
+ 0.825,
+ 0.616
+ ],
+ "angle": 0,
+ "content": "The composition of users in the real world is both complex and diverse, making it impractical to enumerate all users in every scenario. Therefore, it is essential to identify target users whose distribution aligns with that of the users in the corresponding scenario, thereby accurately reflecting the real-world composition and relationships [17, 45]. Based on this, precise target user simulation also requires providing agents with a detailed and comprehensive description of the corresponding users, often involving the integration of high-fidelity demographic, contextual, and behavioral data."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.622,
+ 0.812,
+ 0.636
+ ],
+ "angle": 0,
+ "content": "Q3. How to align the interaction mechanism with the real world among different scenarios?"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.636,
+ 0.827,
+ 0.733
+ ],
+ "angle": 0,
+ "content": "The diversity of social interactions presents challenges in social simulation design, requiring deliberate choices regarding the number of individuals, social structures, interaction patterns, and message dissemination mechanisms, to align with the real world. This often results in independently constructed task-specific simulation pipelines performing repetitive work, which reduces their generalizability and scalability [26, 58]. Therefore, there is a need for unified simulation frameworks based on systematic categorization to standardize simulation components and facilitate extensibility across different social scenarios."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.739,
+ 0.652,
+ 0.754
+ ],
+ "angle": 0,
+ "content": "Q4. How to align the behavioral pattern with the real-world groups?"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.753,
+ 0.826,
+ 0.837
+ ],
+ "angle": 0,
+ "content": "When the environment perceived by agents, the user composition, and the interaction mechanisms are aligned with the real world, agents are expected to exhibit responses consistent with those of the corresponding real users. However, current LLMs exhibit inherent bias and limitations in such reasoning, failing to infer different types of user behaviors [16, 60]. Therefore, it is necessary to systematically collect behavior-driving factors across different user characteristics and adopt appropriate modeling approaches to effectively capture diverse behavior patterns."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.843,
+ 0.826,
+ 0.913
+ ],
+ "angle": 0,
+ "content": "In this paper, we propose SocioVerse, a world model for social simulation driven by LLM-based agents based on a large-scale real-world user pool. As shown in Figure 2, we design modular components to address the above questions. The Social Environment injects up-to-date and external real-world information into the simulation. The User Engine and Scenario Engine respectively reconstruct realistic user context and orchestrate the simulation process to align the simulation with"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.936,
+ 0.505,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "2"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.184,
+ 0.098,
+ 0.809,
+ 0.4
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.171,
+ 0.412,
+ 0.825,
+ 0.469
+ ],
+ "angle": 0,
+ "content": "Figure 2: An illustration of SocioVerse framework involving 4 powerful parts. The social environment provides an updated context for the simulation. During the simulation, the behavior engine takes the simulation setting, user profiles, and social information from the scenario engine, user engine, and social environment, respectively, and generates the results according to the query."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.51,
+ 0.825,
+ 0.54
+ ],
+ "angle": 0,
+ "content": "the real world. Given this rich contextual setup, the Behavior Engine then drives agents to reproduce human behaviors accordingly."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.544,
+ 0.825,
+ 0.616
+ ],
+ "angle": 0,
+ "content": "To support the framework, we construct a user pool of 10 million individuals by collecting real-world social media data to power the user engine. Comparable in scale to the entire populations of Hungary or Greece, this extensive pool enables diverse and large-scale social simulations. For any customized simulation task, various sampling strategies can be applied to extract target user groups from the pool to support the simulation process."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.62,
+ 0.827,
+ 0.704
+ ],
+ "angle": 0,
+ "content": "We conduct three simulations using the SocioVerse framework, each differing in research domain, user composition, and social environment: (a) presidential election prediction, (b) breaking news feedback, and (c) national economic survey. For each task, we compare the simulation results with real-world situations. Extensive and comprehensive experiments demonstrate that our framework serves as a robust foundation for building standardized and accurate large-scale social simulations. In summary, our key contributions are as follows:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.728,
+ 0.825,
+ 0.77
+ ],
+ "angle": 0,
+ "content": "- SocioVerse: We propose a world model for social simulation comprising four powerful alignment modules, enabling diverse and trustworthy social simulations (as illustrated in Figure 2)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.788,
+ 0.827,
+ 0.831
+ ],
+ "angle": 0,
+ "content": "- 10M User Pool: A user pool of 10 million individuals, constructed from real-world behavioral data, enables large-scale and diverse social simulations, ranging from small interest groups to large citizen communities."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.849,
+ 0.825,
+ 0.892
+ ],
+ "angle": 0,
+ "content": "- Three Illustrative Simulations: We demonstrate the framework's capabilities through three distinct scenarios: presidential election prediction, breaking news feedback, and a national economic survey, providing a foundation for future research."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.217,
+ 0.728,
+ 0.827,
+ 0.892
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.494,
+ 0.936,
+ 0.504,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "3"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.173,
+ 0.09,
+ 0.28,
+ 0.106
+ ],
+ "angle": 0,
+ "content": "2 Methods"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.173,
+ 0.122,
+ 0.317,
+ 0.136
+ ],
+ "angle": 0,
+ "content": "Overall Framework"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.149,
+ 0.828,
+ 0.22
+ ],
+ "angle": 0,
+ "content": "The SocioVerse framework follows a structured pipeline to achieve realistic social simulation results, as shown in Figure 2: (1) Social Environment collects updated information and contextual knowledge. Within the simulation environment, (2) User Engine aligns the simulated agents with target users, (3) Scenario Engine aligns the interaction structure with diverse scenarios, and (4) Behavior Engine aligns the behavioral pattern with real-world target groups."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.238,
+ 0.352,
+ 0.251
+ ],
+ "angle": 0,
+ "content": "2.1 Social Environment"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.264,
+ 0.827,
+ 0.307
+ ],
+ "angle": 0,
+ "content": "Function The social environment provides event-related context to align the simulation environment with real-world conditions. By integrating up-to-date events, social statistics, and preference content into LLM-based agents, it enhances the realism of the simulation and improve agent decision-making."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.323,
+ 0.825,
+ 0.366
+ ],
+ "angle": 0,
+ "content": "Components The social environment should encompass as much real-world social, cultural, and technological context as possible. It can be broadly categorized into three types: social structural information, social dynamic information, and personalized context."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.371,
+ 0.826,
+ 0.456
+ ],
+ "angle": 0,
+ "content": "Social Structure: Social structural information provides agents with a rich knowledge base encompassing demographic distributions, cultural norms, urban infrastructures, and collective behavior patterns [57]. This data allows agents to behave in a way that aligns with the typical characteristics of their assigned demographic or geographic profile. For example, by incorporating regional dialect preferences, work-life habits, and common social values, the simulation can more accurately reflect public discourse trends, mobility behaviors, and economic interactions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.461,
+ 0.825,
+ 0.532
+ ],
+ "angle": 0,
+ "content": "Social Dynamics: Social dynamics encompass time-sensitive content continuously generated in the real world, such as news events and policy changes. Typically, this engine maintains an up-to-date event base to continuously collect real-world event news from mainstream news, and all the news articles contain time stamps and event-related tags so that LLM-based agents can comb through the timeline of the events and react accordingly [37]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.537,
+ 0.826,
+ 0.607
+ ],
+ "angle": 0,
+ "content": "Personalized Context: In addition to the macro social environment, individuals also receive different personalized information feeds. Previous studies have explored that the recommendation system can enhance the behavior diversity of the agent [31, 56, 60]. Consequently, the preference content component constructs relevant posts and pushes them to agents according to their social interaction network and interesting topics."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.172,
+ 0.625,
+ 0.3,
+ 0.64
+ ],
+ "angle": 0,
+ "content": "2.2 User Engine"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.651,
+ 0.827,
+ 0.681
+ ],
+ "angle": 0,
+ "content": "Function The user engine aligns simulated agents with a rich set of real-world user samples, enabling the creation of complex target users within the simulation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.697,
+ 0.825,
+ 0.725
+ ],
+ "angle": 0,
+ "content": "Components To support diverse user composition and effective user retrieval and description, the user engine incorporates a large user pool and a wide range of user labels."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.171,
+ 0.731,
+ 0.827,
+ 0.87
+ ],
+ "angle": 0,
+ "content": "User Pools: The user pool is designed to collect extensive digital footprints of individuals across social media platforms, enabling a more comprehensive characterization of real-world behavioral patterns and expression tendencies. To this end, we constructed a user pool covering a variety of social media platforms, including \\(\\mathbf{X}^1\\) and Rednote| Value Labels | A. Below 10% \nB. 10%-20% \nC. 21%-30% \nD. 31%-40% \nE. Above 40% |
| Q07 | Daily Service |
| Question | What is your average monthly expenditure on daily necessities (personal care, house- hold items, cleaning supplies, etc.) and services (housekeeping, repairs, beauty, pet services, etc.)? (Unit: CNY) |
| Value Labels | A. Below 80 CNY \nB. 81-120 CNY \nC. 121-160 CNY \nD. 161-200 CNY \nE. Above 200 CNY |
| Q08 | Transportation & Communication |
| Question | What is your average monthly expenditure on transportation (public transport, taxis, fuel, parking, etc.) and communication (mobile and internet fees)? (Unit: CNY) |
| Value Labels | A. Below 200 CNY \nB. 201-300 CNY \nC. 301-400 CNY \nD. 401-500 CNY \nE. Above 500 CNY |
| Q09 | Education & Entertainment |
| Question | What is your average monthly expenditure on education (tuition, training, books, etc.) and cultural entertainment (movies, performances, games, fitness, cultural activities, etc.)? (Unit: CNY) |
| Value Labels | A. Below 100 CNY \nB. 101-200 CNY \nC. 201-300 CNY \nD. 301-400 CNY \nE. Above 400 CNY |
| Q10 | Education & Entertainment |
| Question | Can you easily afford your current education, cultural, and entertainment expenses? |
| Value Labels | A. Yes, spending does not affect other areas \nB. Barely, needs some control \nC. Not really, affects other expenditures \nD. No, it creates significant financial pressure |
| Q11 | Medical |
| Question | What is your average monthly expenditure on healthcare (medications, medical services, health management, etc.)? (Unit: CNY) |
| Value Labels | A. Below 100 CNY \nB. 101-200 CNY \nC. 201-300 CNY \nD. 301-400 CNY \nE. Above 400 CNY |
| Q12 | Medical |
| Question | Have you purchased private medical or health insurance for yourself or your family? |
| Value Labels | A. Yes \nB. Not yet, but planning to \nC. No, and no plans to |
| Q13 | Others |
"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.49,
+ 0.936,
+ 0.508,
+ 0.948
+ ],
+ "angle": 0,
+ "content": "31"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.174,
+ 0.088,
+ 0.857,
+ 0.679
+ ],
+ "angle": 0,
+ "content": "| Q01 | Voting Behavior |
| Question | ORDER OF MAJOR PARTY CANDIDATE NAMES |
| Value Labels | 1. Democrat first / Republican second2. Republican first / Democrat second |
| Q02 | Social Security |
| Question | Next I am going to read you a list of federal programs. For each one, I would like you to tell me whether you would like to see spending increased, decreased, or kept the same.What about Social Security? Should federal spending on Social Security be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q03 | Education |
| Question | What about public schools? Should federal spending on public schools be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q04 | Immigration |
| Question | What about tightening border security to prevent illegal immigration? Should federal spending on tightening border security to prevent illegal immigration be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q05 | Criminal Justice |
| Question | What about dealing with crime? Should federal spending on dealing with crime be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q06 | Social Welfare |
| Question | What about welfare programs? Should federal spending on welfare programs be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q07 | Infrastructure |
| Question | What about building and repairing highways? Should federal spending on building and repairing highways be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF
+1. Increased
+2. Decreased
+3. Kept the same |
| Q08 | Aid to Poor |
| Question | What about aid to the poor? Should federal spending on aid to the poor be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF
+1. Increased
+2. Decreased
+3. Kept the same |
| Q09 | Environment |
| Question | What about protecting the environment? Should federal spending on protecting the environment be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF
+1. Increased
+2. Decreased
+3. Kept the same |
| Q10 | Government |
| Question | How much do you feel that having elections makes the government pay attention to what the people think? |
| Value Labels | -2. DK/RF
+1. A good deal
+2. Some
+3. Not much |
| Q11 | Economy |
| Question | Which party do you think would do a better job of handling the nation's economy? |
| Value Labels | -2. DK/RF
+1. Democrats would do a better job
+2. Not much difference between them
+3. Republicans would do a better job |
| Q12 | Health Care |
| Question | Which party do you think would do a better job of handling health care? |
| Value Labels | -2. DK/RF
+1. Democrats would do a better job
+2. Not much difference between them
+3. Republicans would do a better job |
| Q13 | Immigration |
| Question | Which party do you think would do a better job of handling immigration? |
| Value Labels | -2. DK/RF
+1. Democrats would do a better job
+2. Not much difference between them
+3. Republicans would do a better job |
| Q14 | Taxes |
| Question | Which party do you think would do a better job of handling taxes? |
| Value Labels | -2. DK/RF
+1. Democrats would do a better job
+2. Not much difference between them
+3. Republicans would do a better job |
| Q15 | Environment |
| Question | Which party do you think would do a better job of handling the environment? |
| Value Labels | -2. DK/RF
+1. Democrats would do a better job
+2. Not much difference between them
+3. Republicans would do a better job |
| Q16 | Education |
| Question | Some people think the government should provide fewer services even in areas such as health and education in order to reduce spending.
+Other people feel it is important for the government to provide many more services even if it means an increase in spending.
+And, of course, some people have a neutral position.
+Which of the following best describes your view? |
| Value Labels | -2. DK/RF
+1. Government should provide fewer services
+2. Neutral
+3. Government should provide more services |
| Q17 | Defense |
| Question | Some people believe that we should spend less money for defense.
+Others feel that defense spending should be increased.
+And, of course, some people have a neutral position.
+Which of the following best describes your view? |
| Value Labels | -2. DK/RF
+1. Decrease defense spending
+2. Neutral
+3. Increase defense spending |
| Q18 | Health Care |
| Question | There is much concern about the rapid rise in medical and hospital costs.
+Some people feel there should be a government insurance plan which would cover all medical and hospital expenses for everyone.
+Others feel that all medical expenses should be paid by individuals through private insurance plans like Blue Cross or other company paid plans.
+And, of course, some people have a neutral position.
+Which of the following best describes your view? |
| Value Labels | -2. DK/RF
+1. Government insurance plan
+2. Neutral
+3. Private insurance plan |
| Q19 | Social Welfare |
| Question | Some people feel the government in Washington should see to it that every person has a job and a good standard of living.
+Others think the government should just let each person get ahead on their own.
+And, of course, some people have a neutral position.
+Which of the following best describes your view? |
| Value Labels | -2. DK/RF
+1. Government should see to jobs and standard of living
+2. Neutral
+3. Government should let each person get ahead on own |
| Q20 | Aid to Blacks |
| Question | Some people feel that the government in Washington should make every effort to improve the social and economic position of blacks.
+Others feel that the government should not make any special effort to help blacks because they should help themselves.
+And, of course, some people have a neutral position.
+Which of the following best describes your view? |
| Value Labels | -2. DK/RF
+1. Government should help blacks
+2. Neutral
+3. Blacks should help themselves |
| Q21 | Environment |
| Question | Some people think we need much tougher government regulations on business in order to protect the environment.
+Others think that current regulations to protect the environment are already too much of a burden on business.
+And, of course, some people have a neutral position.
+Which of the following best describes your view? |
| Value Labels | -2. DK/RF
+1. Tougher regulations on business needed to protect environment
+2. Neutral
+3. Regulations to protect environment already too much a burden on business |
| Q22 | Abortion |
| Question | Would you be pleased, upset, or neither pleased nor upset if the Supreme Court reduced abortion rights? |
| Value Labels | -2. DK/RF
+1. Pleased
+2. Upset
+3. Neither pleased nor upset |
| Q23 | Criminal Justice |
| Question | Do you favor or oppose the death penalty for persons convicted of murder? |
| Value Labels | -2. DK/RF
+1. Favor
+2. Oppose |
| Q24 | US Position in World |
| Question | Do you agree or disagree with this statement: ‘This country would be better off if we just stayed home and did not concern ourselves with problems in other parts of the world.’ |
| Value Labels | -2. DK/RF
+1. Agree
+2. Disagree |
| Q25 | US Position in World |
| Question | How willing should the United States be to use military force to solve international problems? |
| Value Labels | -2. DK/RF
+1. Willing
+2. Moderately willing
+3. Not willing |
| Q26 | Inequality |
| Question | Do you think the difference in incomes between rich people and poor people in the United States today is larger, smaller, or about the same as it was 20 years ago? |
| Value Labels | -2. DK/RF
+1. Larger
+2. Smaller
+3. About the same |
| Q27 | Environment |
| Question | Do you think the federal government should be doing more about rising temperatures, should be doing less, or is it currently doing the right amount? |
| Value Labels | -2. DK/RF
+1. Should be doing more
+2. Should be doing less
+3. Is currently doing the right amount |
| Q28 | Parental Leave |
| Question | Do you favor, oppose, or neither favor nor oppose requiring employers to offer paid leave to parents of new children? |
| Value Labels | -2. DK/RF
+1. Favor
+2. Oppose
+3. Neither favor nor oppose |
| Q29 | LGBTQ+ Rights |
| Question | Do you think business owners who provide wedding-related services should be allowed to refuse services to same-sex couples if same-sex marriage violates their religious beliefs, or do you think business owners should be required to provide services regardless of a couple's sexual orientation? |
| Value Labels | -2. DK/RF
+1. Should be allowed to refuse
+2. Should be required to provide services |
| Q30 | LGBTQ+ Rights |
| Question | Should transgender people - that is, people who identify themselves as the sex or gender different from the one they were born as - have to use the bathrooms of the gender they were born as, or should they be allowed to use the bathrooms of their identified gender? |
| Value Labels | -2. DK/RF
+1. Have to use the bathrooms of the gender they were born as
+2. Be allowed to use the bathrooms of their identified gender |
| Q31 | LGBTQ+ Rights |
| Question | Do you favor or oppose laws to protect gays and lesbians against job discrimination? |
| Value Labels | -2. DK/RF
+1. Favor
+2. Oppose |
| Q32 | LGBTQ+ Rights |
| Question | Do you think gay or lesbian couples should be legally permitted to adopt children? |
| Value Labels | -2. DK/RF
+1. Yes
+2. No |
| Q33 | LGBTQ+ Rights |
| Question | Which comes closest to your view? You can just tell me the number of your choice. |
| Value Labels | -2. DK/RF 1. Gay and lesbian couples should be allowed to legally marry
+2. Gay and lesbian couples should be allowed to form civil unions but not legally marry
+3. There should be no legal recognition of gay or lesbian couples' relationship |
| Q34 | Immigration |
| Question | Some people have proposed that the U.S. Constitution should be changed so that the children of unauthorized immigrants do not automatically get citizenship if they are born in this country.
+Do you favor, oppose, or neither favor nor oppose this proposal? |
| Value Labels | -2. DK/RF
+1. Favor
+2. Oppose
+3. Neither favor nor oppose |
| Q35 | Immigration |
| Question | What should happen to immigrants who were brought to the U.S. illegally as children and have lived here for at least 10 years and graduated high school here? Should they be sent back where they came from, or should they be allowed to live and work in the United States? |
| Value Labels | -2. DK/RF
+1. Should be sent back where they came from
+2. Should be allowed to live and work in the US |
| Q36 | Immigration |
| Question | Do you favor, oppose, or neither favor nor oppose building a wall on the U.S. border with Mexico? |
| Value Labels | -2. DK/RF
+1. Favor
+2. Oppose
+3. Neither favor nor oppose |
| Q37 | Unrest |
| Question | During the past few months, would you say that most of the actions taken by protestors to get the things they want have been violent, or have most of these actions by protestors been peaceful, or have these actions been equally violent and peaceful? |
| Value Labels | -2. DK/RF
+1. Mostly violent
+2. Mostly peaceful
+3. Equally violent and peaceful |
| Q38 | Government |
| Question | Do you think it is better when one party controls both the presidency and Congress, better when control is split between the Democrats and Republicans, or doesn’t it matter? |
| Value Labels | -2. DK/RF
+1. Better when one party controls both
+2. Better when control is split
+3. It doesn’t matter |
| Q39 | Government |
| Question | Would you say the government is pretty much run by a few big interests looking out for themselves or that it is run for the benefit of all the people? |
| Value Labels | -2. DK/RF
+1. Run by a few big interests
+2. For the benefit of all the people |
| Q40 | Government |
| Question | Do you think that people in government waste a lot of the money we pay in taxes, waste some of it, or don’t waste very much of it? |
| Value Labels | -2. DK/RF
+1. Waste a lot
+2. Waste some
+3. Don’t waste very much |
| Q41 | Election Integrity |
| Question | Do you favor, oppose, or neither favor nor oppose allowing convicted felons to vote once they complete their sentence? |
| Value Labels | -2. DK/RF
+1. Favor
+2. Oppose
+3. Neither favor nor oppose |
| Q42 | Democratic Norms |
| Question | How important is it that news organizations are free to criticize political leaders? |
| Value Labels | -2. DK/RF
+1. Not important
+2. Moderately important
+3. Important |
| Q43 | Democratic Norms |
| Question | How important is it that the executive, legislative, and judicial branches of government keep one another from having too much power? |
| Value Labels | -2. DK/RF
+1. Not important
+2. Moderately important
+3. Important |
| Q44 | Democratic Norms |
| Question | How important is it that elected officials face serious consequences if they engage in misconduct? |
| Value Labels | -2. DK/RF
+1. Not important
+2. Moderately important
+3. Important |
| Q45 | Democratic Norms |
| Question | How important is it that people agree on basic facts even if they disagree politically? |
| Value Labels | -2. DK/RF
+1. Not important
+2. Moderately important
+3. Important |
| Q46 | Democratic Norms |
| Question | Would it be helpful, harmful, or neither helpful nor harmful if U.S. presidents could work on the country’s problems without paying attention to what Congress and the courts say? |
| Value Labels | -2. DK/RF
+1. Helpful
+2. Harmful
+3. Neither helpful nor harmful |
| Q47 | Democratic Norms |
| Question | Do you favor, oppose, or neither favor nor oppose elected officials restricting journalists’ access to information about government decision-making? |
| Value Labels | -2. DK/RF
+1. Favor
+2. Oppose
+3. Neither favor nor oppose |
| Q48 | Gender Resentment |
| Question | ‘Many women interpret innocent remarks or acts as being sexist.’
+Do you agree, neither agree nor disagree, or disagree with this statement? |
| Value Labels | -2. DK/RF/technical error
+1. Agree
+2. Neither agree nor disagree
+3. Disagree |
| Q49 | Gender Resentment |
| Question | ‘Women seek to gain power by getting control over men.’
+Do you agree, neither agree nor disagree, or disagree with this statement? |
| Value Labels | -2. DK/RF/technical error
+1. Agree
+2. Neither agree nor disagree
+3. Disagree |
+
+# D.2 Questionnaire for Breaking News Feedback
+
+| Q01 | Public Cognition (PC) |
| Question | I have heard of ChatGPT. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q02 | Public Cognition (PC) |
| Question | Many people around me use ChatGPT. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q03 | Public Cognition (PC) |
| Question | I have a deep understanding of ChatGPT's functions and applications. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q04 | Perceived Risks (PR) |
| Question | ChatGPT may lead to the widespread dissemination of false information. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q05 | Perceived Risks (PR) |
| Question | ChatGPT may reduce human thinking ability and creativity. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q06 | Perceived Risks (PR) |
| Question | The development of ChatGPT may replace certain jobs, and I am deeply concerned about this. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q07 | Perceived Benefits (PB) |
| Question | ChatGPT will definitely improve my work and study efficiency. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q08 | Perceived Benefits (PB) |
| Question | ChatGPT helps broaden my knowledge and provides me with new perspectives and ideas. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q09 | Perceived Benefits (PB) |
| Question | ChatGPT promotes technological innovation and development in related fields. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q10 | Trust (TR) |
| Question | I fully trust the team developing ChatGPT to manage and guide its development responsibly. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q11 | Trust (TR) |
| Question | I have strong confidence in the accuracy and reliability of the information generated by ChatGPT. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q12 | Trust (TR) |
| Question | I believe that the future application of ChatGPT will be effectively regulated. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q13 | Fairness (FA) |
| Question | The opportunities to use ChatGPT are distributed fairly among different groups of people. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q14 | Fairness (FA) |
| Question | I find the distribution of benefits brought by ChatGPT to be fair. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q15 | Fairness (FA) |
| Question | I believe that the decision-making process for the development and promotion of ChatGPT is fully transparent and adequately reflects public interests. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q16 | Public Acceptance (PA) |
| Question | Overall, I strongly welcome the emergence of ChatGPT. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q17 | Public Acceptance (PA) |
| Question | I am definitely willing to use ChatGPT in my work or studies. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
| Q18 | Public Acceptance (PA) |
| Question | I strongly support increased investment in the research and development of AI technologies like ChatGPT. |
| Value Labels | 1. Disagree
+2. Partially disagree
+3. Neutral
+4. Partially agree
+5. Agree |
+
+# D.3 Questionnaire for National Economic Survey
+
+| Q01 | Food |
| Question | What is your average monthly expenditure on food (including dining out)? (Unit: CNY) |
| Value Labels | A. Below 500 CNY
+B. 501-650 CNY
+C. 651-800 CNY
+D. 801-1000 CNY
+E. Above 1000 CNY |
| Q02 | Food |
| Question | Do you think your current spending on food, tobacco, and alcohol is too high relative to your income? |
| Value Labels | A. Yes
+B. No
+C. Acceptable |
| Q03 | Clothing |
| Question | What is your average monthly expenditure on clothing (including apparel, shoes, and accessories)? (Unit: CNY) |
| Value Labels | A. Below 50 CNY
+B. 51-100 CNY
+C. 101-150 CNY
+D. 151-200 CNY
+E. Above 200 CNY |
| Q04 | Clothing |
| Question | How much economic pressure do you feel from clothing expenses? |
| Value Labels | A. Very low, almost no pressure
+B. Moderate, some pressure but manageable
+C. High, requires careful spending
+D. Very high, affects spending in other areas |
| Q05 | Household |
| Question | What is your average monthly housing expenditure? (Including rent, mortgage, property fees, maintenance, etc.) (Unit: CNY) |
| Value Labels | A. Below 200 CNY
+B. 201-500 CNY
+C. 501-800 CNY
+D. 801-1200 CNY
+E. Above 1200 CNY |
| Q06 | Household |
| Question | What percentage of your monthly income is spent on housing? (Including rent, mortgage, property fees, maintenance, etc.) |
| Value Labels | A. Below 10%
+B. 10%-20%
+C. 21%-30%
+D. 31%-40%
+E. Above 40% |
| Q07 | Daily Service |
| Question | What is your average monthly expenditure on daily necessities (personal care, house- hold items, cleaning supplies, etc.) and services (housekeeping, repairs, beauty, pet services, etc.)? (Unit: CNY) |
| Value Labels | A. Below 80 CNY
+B. 81-120 CNY
+C. 121-160 CNY
+D. 161-200 CNY
+E. Above 200 CNY |
| Q08 | Transportation & Communication |
| Question | What is your average monthly expenditure on transportation (public transport, taxis, fuel, parking, etc.) and communication (mobile and internet fees)? (Unit: CNY) |
| Value Labels | A. Below 200 CNY
+B. 201-300 CNY
+C. 301-400 CNY
+D. 401-500 CNY
+E. Above 500 CNY |
| Q09 | Education & Entertainment |
| Question | What is your average monthly expenditure on education (tuition, training, books, etc.) and cultural entertainment (movies, performances, games, fitness, cultural activities, etc.)? (Unit: CNY) |
| Value Labels | A. Below 100 CNY
+B. 101-200 CNY
+C. 201-300 CNY
+D. 301-400 CNY
+E. Above 400 CNY |
| Q10 | Education & Entertainment |
| Question | Can you easily afford your current education, cultural, and entertainment expenses? |
| Value Labels | A. Yes, spending does not affect other areas
+B. Barely, needs some control
+C. Not really, affects other expenditures
+D. No, it creates significant financial pressure |
| Q11 | Medical |
| Question | What is your average monthly expenditure on healthcare (medications, medical services, health management, etc.)? (Unit: CNY) |
| Value Labels | A. Below 100 CNY
+B. 101-200 CNY
+C. 201-300 CNY
+D. 301-400 CNY
+E. Above 400 CNY |
| Q12 | Medical |
| Question | Have you purchased private medical or health insurance for yourself or your family? |
| Value Labels | A. Yes
+B. Not yet, but planning to
+C. No, and no plans to |
| Q13 | Others |
| Question | Besides food, clothing, housing, daily necessities and services, transportation, education, culture, and healthcare, what is your average monthly expenditure on other areas (e.g., hobbies, charitable donations, investment, etc.)? (Unit: CNY) |
| Value Labels | A. Below 30 CNY
+B. 31-60 CNY
+C. 61-90 CNY
+D. 91-120 CNY
+E. Above 120 CNY |
| Q14 | Overall |
| Question | How would you evaluate the impact of your current consumption level on your household (or personal) financial situation? |
| Value Labels | A. Comfortable, can moderately increase spending
+B. Average, can maintain current spending
+C. Tight, need to control or reduce spending
+D. Very tight, affects quality of life |
| Q15 | Overall |
| Question | Do you feel that your consumption pressure is too high relative to your income level? |
| Value Labels | A. Yes
+B. No
+C. Not sure |
| Q16 | Overall |
| Question | If your income increases, which consumption areas would you most like to expand or improve? (Multiple choices allowed) |
| Value Labels | A. Food and alcohol
+B. Clothing
+C. Housing
+D. Daily necessities and services
+E. Transportation and communication
+F. Education, culture, and entertainment
+G. Healthcare
+H. Other goods and services |
| Q17 | Overall |
| Question | What is your consumption expectation for the next six months to a year? |
| Value Labels | A. Will continue to increase
+B. Will remain roughly the same
+C. Will moderately decrease
+D. Uncertain |
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10157/images/02108be3574c08f31dfa322bb6fe31f2cac9f0c26240bc43f385a3edcc216619.jpg b/data/2025/2504_10xxx/2504.10157/images/02108be3574c08f31dfa322bb6fe31f2cac9f0c26240bc43f385a3edcc216619.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..da453f3646cb53e00b5c36e4920ed48f1a6c0d2f
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/02108be3574c08f31dfa322bb6fe31f2cac9f0c26240bc43f385a3edcc216619.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56a5534777cef90574cd4eb36048aa295f73af9d9c6d03ea42f490dc3e17d9f5
+size 290134
diff --git a/data/2025/2504_10xxx/2504.10157/images/0cb012c6c55c49fd65fa15745af750ea86382e3c5992b62bd56df63efb6706fc.jpg b/data/2025/2504_10xxx/2504.10157/images/0cb012c6c55c49fd65fa15745af750ea86382e3c5992b62bd56df63efb6706fc.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e6cfa901c17f4a5558ee8e3ac464f69a348f1e6f
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/0cb012c6c55c49fd65fa15745af750ea86382e3c5992b62bd56df63efb6706fc.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc0cc3aa2e52fa9634c81220eb2a71717b5976533cc530b46dd78354f78cb141
+size 23707
diff --git a/data/2025/2504_10xxx/2504.10157/images/18070cfafc0e2c3f8693e46f9ae0e339394f7c42e0f7d13109ea8126d6c4d3ff.jpg b/data/2025/2504_10xxx/2504.10157/images/18070cfafc0e2c3f8693e46f9ae0e339394f7c42e0f7d13109ea8126d6c4d3ff.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ff7b70711f1f4ebfc60dfbbd15fc319c6e1c11a3
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/18070cfafc0e2c3f8693e46f9ae0e339394f7c42e0f7d13109ea8126d6c4d3ff.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab478f76529ed648d1e0565550e27e9e97a35fa1eb4386258777be392ab92e2a
+size 186968
diff --git a/data/2025/2504_10xxx/2504.10157/images/1c254baddeee82d257854c6de8cf0642cfab2a1bf3f197f0a8fb52bcd38fa176.jpg b/data/2025/2504_10xxx/2504.10157/images/1c254baddeee82d257854c6de8cf0642cfab2a1bf3f197f0a8fb52bcd38fa176.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..fd6fb47ee597485162c99978918a6e5bc5fec9f4
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/1c254baddeee82d257854c6de8cf0642cfab2a1bf3f197f0a8fb52bcd38fa176.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3a89a04f917ac9a5b4c208d8c88312d233e5263a997cbefc41d7e624d8e3fe0
+size 39671
diff --git a/data/2025/2504_10xxx/2504.10157/images/2e6b8fa3fd5779995f19c798aae290400cdd80c75f2319f40980aedd30481058.jpg b/data/2025/2504_10xxx/2504.10157/images/2e6b8fa3fd5779995f19c798aae290400cdd80c75f2319f40980aedd30481058.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..850ce656c7c0c75ba4d6e562584e254006962c37
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/2e6b8fa3fd5779995f19c798aae290400cdd80c75f2319f40980aedd30481058.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09fcc66d93d6e3a56ca163ad9dd3e8090433fcdec2d17137ea0816997d9c94b3
+size 37898
diff --git a/data/2025/2504_10xxx/2504.10157/images/36082be4a11f38a78da16a370c6d153ea25495bc0786e11c4d30ffd3e799783c.jpg b/data/2025/2504_10xxx/2504.10157/images/36082be4a11f38a78da16a370c6d153ea25495bc0786e11c4d30ffd3e799783c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..bae4575f527544136558fa1110036a40c8fb740e
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/36082be4a11f38a78da16a370c6d153ea25495bc0786e11c4d30ffd3e799783c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5068d6fb387fc23a58ed083751437d5494de77e42c5ca70ef82b05f7dc587f6
+size 4743
diff --git a/data/2025/2504_10xxx/2504.10157/images/41b8f7e83330617a874286474fbb177b1c17d9e2e45f3b39659735bbc7457011.jpg b/data/2025/2504_10xxx/2504.10157/images/41b8f7e83330617a874286474fbb177b1c17d9e2e45f3b39659735bbc7457011.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d8a35b80327c8296872323a55b2a1f3c01abac15
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/41b8f7e83330617a874286474fbb177b1c17d9e2e45f3b39659735bbc7457011.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb12548993d9a6795156c49f6d4e766b90003858e7a2a9ec7f80633e7336ebac
+size 18121
diff --git a/data/2025/2504_10xxx/2504.10157/images/577ebeddce61d3ac540945a4740d2e0fe0a1598ad0904fd729d144e3d7a8e1bf.jpg b/data/2025/2504_10xxx/2504.10157/images/577ebeddce61d3ac540945a4740d2e0fe0a1598ad0904fd729d144e3d7a8e1bf.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..069dbb7851bbb0358bd0506f221184df296997f1
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/577ebeddce61d3ac540945a4740d2e0fe0a1598ad0904fd729d144e3d7a8e1bf.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cc412508e504514d7469a5c9f66962c79f821bfb4088a68caded13183ce9da7
+size 202303
diff --git a/data/2025/2504_10xxx/2504.10157/images/5a4f9f269826f82d19138a4c9b1f45390e15057486c0858c22877f315d993a5d.jpg b/data/2025/2504_10xxx/2504.10157/images/5a4f9f269826f82d19138a4c9b1f45390e15057486c0858c22877f315d993a5d.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3f1aea4e63f34902b30372c8691c3f6323ed5130
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/5a4f9f269826f82d19138a4c9b1f45390e15057486c0858c22877f315d993a5d.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1dec3626320becada29aa3a8a7388a7954ac9655fc554367578524e7d80e0fc
+size 283934
diff --git a/data/2025/2504_10xxx/2504.10157/images/5b6943b2f52d4b323fee259ac4a2f465c671d7812940c7effe8f60763112e09e.jpg b/data/2025/2504_10xxx/2504.10157/images/5b6943b2f52d4b323fee259ac4a2f465c671d7812940c7effe8f60763112e09e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..eacada1566f1d7706496ea2a6eed0031873c0a1d
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/5b6943b2f52d4b323fee259ac4a2f465c671d7812940c7effe8f60763112e09e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6289e3d8cb4564028b78272a7f5dfa15c5e9dde0988cc85be51885b0a7e4f520
+size 247733
diff --git a/data/2025/2504_10xxx/2504.10157/images/5cd70d2f9aa944b30cd8014a3f27307a2828ab5935fadc5daf426be35f3eedea.jpg b/data/2025/2504_10xxx/2504.10157/images/5cd70d2f9aa944b30cd8014a3f27307a2828ab5935fadc5daf426be35f3eedea.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a63d3612fc1892a1e2574133690d4656e5358385
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/5cd70d2f9aa944b30cd8014a3f27307a2828ab5935fadc5daf426be35f3eedea.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:858cb7d099bb833a8ac3c8b1738bb8aee1d74fac253bf2eca2679ee187f1c06e
+size 12930
diff --git a/data/2025/2504_10xxx/2504.10157/images/6b41c5eea28ebef44e10cd877369420c1e194556283d8d7959e64dfa68ee1f47.jpg b/data/2025/2504_10xxx/2504.10157/images/6b41c5eea28ebef44e10cd877369420c1e194556283d8d7959e64dfa68ee1f47.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e8056422f001cb6d19962397aa9b60e163953270
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/6b41c5eea28ebef44e10cd877369420c1e194556283d8d7959e64dfa68ee1f47.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b85aacc058bf278c82d7cc217b6134efabb6357ec571046f70cf28612d40800
+size 232098
diff --git a/data/2025/2504_10xxx/2504.10157/images/70d85cebae8909029afc6c6fbdf59529d8448c63ca93e6fe9200e65826808979.jpg b/data/2025/2504_10xxx/2504.10157/images/70d85cebae8909029afc6c6fbdf59529d8448c63ca93e6fe9200e65826808979.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..63c72ff8875184fc326aa853e8ac326682b81400
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/70d85cebae8909029afc6c6fbdf59529d8448c63ca93e6fe9200e65826808979.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d6ea684255fd209306094f857118c03bded6146ecaab33377d5ce27b8e8864f
+size 32871
diff --git a/data/2025/2504_10xxx/2504.10157/images/7904c50696c6c4b3ef575c985f044bc7eab44f178d18b3c0d1f745b398df9e69.jpg b/data/2025/2504_10xxx/2504.10157/images/7904c50696c6c4b3ef575c985f044bc7eab44f178d18b3c0d1f745b398df9e69.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c51eda120f988e002b934d06359bcf1e081ca5f2
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/7904c50696c6c4b3ef575c985f044bc7eab44f178d18b3c0d1f745b398df9e69.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9738de4321a356845241c32cd47a18aa4d9f765fb666eec1c854c54bafe5d036
+size 4121
diff --git a/data/2025/2504_10xxx/2504.10157/images/7ab3b60a9c281cf0675ca7cbaa1ac451358842d500a4f62bb1f322d99ada9ba4.jpg b/data/2025/2504_10xxx/2504.10157/images/7ab3b60a9c281cf0675ca7cbaa1ac451358842d500a4f62bb1f322d99ada9ba4.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3a02c0f017a4a53684fc7d502bd3707b90f31a38
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/7ab3b60a9c281cf0675ca7cbaa1ac451358842d500a4f62bb1f322d99ada9ba4.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:538689e8cfed429ddb58869ecbf6ad25469f29ef4c7fd47a826383538b053573
+size 15097
diff --git a/data/2025/2504_10xxx/2504.10157/images/7c9a05886551ed993c5c977d179fdfeb24745f20e17ac999b1da310b35dbc695.jpg b/data/2025/2504_10xxx/2504.10157/images/7c9a05886551ed993c5c977d179fdfeb24745f20e17ac999b1da310b35dbc695.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a4a663803183babe187f26632dc0378172de3b6b
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/7c9a05886551ed993c5c977d179fdfeb24745f20e17ac999b1da310b35dbc695.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d26f94847f1df3f059fa54c048fd7a8cef86246584e831b7659f032c294648e9
+size 9193
diff --git a/data/2025/2504_10xxx/2504.10157/images/7cb07d45660942f4481e7ecfce1c3d40f9a6ce0877b27cea63d127d5b346085f.jpg b/data/2025/2504_10xxx/2504.10157/images/7cb07d45660942f4481e7ecfce1c3d40f9a6ce0877b27cea63d127d5b346085f.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d37e66eb5b8e70472001d703a81bcaf272981bd1
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/7cb07d45660942f4481e7ecfce1c3d40f9a6ce0877b27cea63d127d5b346085f.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:161926b5a41698b077ae2d9289ed24f71c101a18143ba49ff83060fc9d5eae53
+size 33696
diff --git a/data/2025/2504_10xxx/2504.10157/images/7ce59a581d9686e1bdc3056be8b7786b1ee7c8b56671e62f94b9629946ea1c5e.jpg b/data/2025/2504_10xxx/2504.10157/images/7ce59a581d9686e1bdc3056be8b7786b1ee7c8b56671e62f94b9629946ea1c5e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c4b1255bbff9f258a0e5e1d6bc95e398a4b6998b
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/7ce59a581d9686e1bdc3056be8b7786b1ee7c8b56671e62f94b9629946ea1c5e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dac2be1d15565c1ffb367164e416840deeac457a573bc48964bebb87080766ec
+size 115977
diff --git a/data/2025/2504_10xxx/2504.10157/images/8d8724c8324cb94a67c34cbf4205324ec05b3f0fe6a1795102f63c1f6d883257.jpg b/data/2025/2504_10xxx/2504.10157/images/8d8724c8324cb94a67c34cbf4205324ec05b3f0fe6a1795102f63c1f6d883257.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3ab1ec5d73d7ec3762da731e023dc432339faad6
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/8d8724c8324cb94a67c34cbf4205324ec05b3f0fe6a1795102f63c1f6d883257.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d28bba9964b5fd2accbb5e2ffda817d6f01fa5335385dfc03c03ffdd4e6b44e
+size 186867
diff --git a/data/2025/2504_10xxx/2504.10157/images/8e14d1c2825d1d0816a823a907c78e20f63f7b4e274ee343a9155a827632e1a0.jpg b/data/2025/2504_10xxx/2504.10157/images/8e14d1c2825d1d0816a823a907c78e20f63f7b4e274ee343a9155a827632e1a0.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f169e409596439e3423b8a9e39074f33e5b49cc1
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/8e14d1c2825d1d0816a823a907c78e20f63f7b4e274ee343a9155a827632e1a0.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5954662c420e24264a2f294508c63857140c6278c28bf46b692836ee96ed141
+size 269594
diff --git a/data/2025/2504_10xxx/2504.10157/images/91ad3cd680ab7c20191622101061881bd683e063e4f7da28fa9d5ce86b319768.jpg b/data/2025/2504_10xxx/2504.10157/images/91ad3cd680ab7c20191622101061881bd683e063e4f7da28fa9d5ce86b319768.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..4a762e25311169cf7aeb3d82a74f0998d2c66453
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/91ad3cd680ab7c20191622101061881bd683e063e4f7da28fa9d5ce86b319768.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2999964a6943f0e60bf533950c870b2c376580650f0ddcff2ebeaef46f070789
+size 267721
diff --git a/data/2025/2504_10xxx/2504.10157/images/95378075e345f87be207b28094b42ea42f66abffcb8dd1545b7c107a69d3e97d.jpg b/data/2025/2504_10xxx/2504.10157/images/95378075e345f87be207b28094b42ea42f66abffcb8dd1545b7c107a69d3e97d.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..98c268b6e5103b31ae3ca86f4de4ff960b7bc8ce
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/95378075e345f87be207b28094b42ea42f66abffcb8dd1545b7c107a69d3e97d.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1464dddbde0fbf6b9f9b42fb6543a052f92b077f3b843b4eb5e52dad708188af
+size 28892
diff --git a/data/2025/2504_10xxx/2504.10157/images/9839e0e42e838abe4b61415c929efb4951f849533401bc247dd8cf0d2900ba80.jpg b/data/2025/2504_10xxx/2504.10157/images/9839e0e42e838abe4b61415c929efb4951f849533401bc247dd8cf0d2900ba80.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..db70fa797387ba5691827badcd9afa23c6b3003c
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/9839e0e42e838abe4b61415c929efb4951f849533401bc247dd8cf0d2900ba80.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53dd3ae075761916df7b20e67f35f198acd23ae841cf144ee6db9735845aea6e
+size 14561
diff --git a/data/2025/2504_10xxx/2504.10157/images/9971298eafb0097555473200ebe95969169d99fbf1d407eea813867188612681.jpg b/data/2025/2504_10xxx/2504.10157/images/9971298eafb0097555473200ebe95969169d99fbf1d407eea813867188612681.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d8f4bd3249cfba4309d3f44008311d65a1794923
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/9971298eafb0097555473200ebe95969169d99fbf1d407eea813867188612681.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9908147701823e8cc752b21234aee771f1a3b878732bbe3cb71ddda084fc5641
+size 43516
diff --git a/data/2025/2504_10xxx/2504.10157/images/9ab297241e7b5ad5640f2f7c89ea694c7a2b607437aa0355f16e8a8ee7812b35.jpg b/data/2025/2504_10xxx/2504.10157/images/9ab297241e7b5ad5640f2f7c89ea694c7a2b607437aa0355f16e8a8ee7812b35.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3b814655f6618aa4def4ddee261e2bf3eaf0d489
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/9ab297241e7b5ad5640f2f7c89ea694c7a2b607437aa0355f16e8a8ee7812b35.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2def6d3a86de65d73f779f20e568d7f47c74cd9f0c10964c0b2e4e4a71516416
+size 18334
diff --git a/data/2025/2504_10xxx/2504.10157/images/a66412febde24e12e177a5c22dc9307635f04919238d13f32d8145fd1596adf3.jpg b/data/2025/2504_10xxx/2504.10157/images/a66412febde24e12e177a5c22dc9307635f04919238d13f32d8145fd1596adf3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e2c92e908c9da020b09309ff3f576de58ec94211
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/a66412febde24e12e177a5c22dc9307635f04919238d13f32d8145fd1596adf3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4b6f55774271a94ab75d3aa2c14c67e92d92218f985ce72cda5c985ad526362
+size 216955
diff --git a/data/2025/2504_10xxx/2504.10157/images/af88ad03ac5ce1c47c53dd172206932141236f13f7154e620648f8b4af53663e.jpg b/data/2025/2504_10xxx/2504.10157/images/af88ad03ac5ce1c47c53dd172206932141236f13f7154e620648f8b4af53663e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..1eb48cc3f03094b86807ac5c0c2f19afcfd3cda9
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/af88ad03ac5ce1c47c53dd172206932141236f13f7154e620648f8b4af53663e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dd5d336e3f27304deb008747e867a0c24ee95e4d37f608d678310e54a618da1
+size 120679
diff --git a/data/2025/2504_10xxx/2504.10157/images/afe8fd5e150abb10daff5d2cd9105ca719a3f43016065f6085c46dd68212ec23.jpg b/data/2025/2504_10xxx/2504.10157/images/afe8fd5e150abb10daff5d2cd9105ca719a3f43016065f6085c46dd68212ec23.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..b2a15208c0960b393131b20907379aba865270d6
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/afe8fd5e150abb10daff5d2cd9105ca719a3f43016065f6085c46dd68212ec23.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86a3ccd0b3d4f1c773bd8e137db688d7de17827bbbbf95b2e56c7fb9b18bec06
+size 82125
diff --git a/data/2025/2504_10xxx/2504.10157/images/b3698fd8a7cc0afbf47a6e31691ed9fd5974990404ef71422f83c8297bb7cea7.jpg b/data/2025/2504_10xxx/2504.10157/images/b3698fd8a7cc0afbf47a6e31691ed9fd5974990404ef71422f83c8297bb7cea7.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..93528f6c34b15fe62eba228b0231bebafe34e82b
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/b3698fd8a7cc0afbf47a6e31691ed9fd5974990404ef71422f83c8297bb7cea7.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b7c8f51d081439c8f89e1533ff0a48fae9ae63f792a63d34eb7d5571ba4b028
+size 39107
diff --git a/data/2025/2504_10xxx/2504.10157/images/b549f4ea6be4bd3460a79b7004922964d2d10d6fea789b0c2253d82729469c6e.jpg b/data/2025/2504_10xxx/2504.10157/images/b549f4ea6be4bd3460a79b7004922964d2d10d6fea789b0c2253d82729469c6e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..73631e8769c3262565cdc5621ea9c4f3e8d6c91b
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/b549f4ea6be4bd3460a79b7004922964d2d10d6fea789b0c2253d82729469c6e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8f13994204268a9220d623d5314c02fb2fe35d898d5efa91b68cafaf334fce8
+size 187575
diff --git a/data/2025/2504_10xxx/2504.10157/images/c20d35dd2c26c2ee2cdb820d4f671e70e3eae3a204e23a512d96c6000ab32e42.jpg b/data/2025/2504_10xxx/2504.10157/images/c20d35dd2c26c2ee2cdb820d4f671e70e3eae3a204e23a512d96c6000ab32e42.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..68b2c36722e016bcc65a926ba7b70dff2bc6173b
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/c20d35dd2c26c2ee2cdb820d4f671e70e3eae3a204e23a512d96c6000ab32e42.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5dd1703afeb5a6dca9e6916fe8e64b80e77ca8f29f603d6ee6b32a927018f5fb
+size 231552
diff --git a/data/2025/2504_10xxx/2504.10157/images/c455658df89756889d84385989fcf4d04f7baa808ab9d0760fe1501340dd9f50.jpg b/data/2025/2504_10xxx/2504.10157/images/c455658df89756889d84385989fcf4d04f7baa808ab9d0760fe1501340dd9f50.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..5ce13afefd8d76dedbe487432f115ad6d98d70bb
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/c455658df89756889d84385989fcf4d04f7baa808ab9d0760fe1501340dd9f50.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d88021c0b88ef58c4f4710e78c54d1d9d558b95de78fb7d351ec0b2c8df4628
+size 5939
diff --git a/data/2025/2504_10xxx/2504.10157/images/c8315fc00f5be85637036238e2bbe85316d8d30a637370c74716c310a2613d7d.jpg b/data/2025/2504_10xxx/2504.10157/images/c8315fc00f5be85637036238e2bbe85316d8d30a637370c74716c310a2613d7d.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..1c3eb348e17556efbc19db9087cc33c25f9634e8
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/c8315fc00f5be85637036238e2bbe85316d8d30a637370c74716c310a2613d7d.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58bb665710155d4706404effa2c01d7fd3544c91fd420581f1d15afbd75c3dda
+size 66168
diff --git a/data/2025/2504_10xxx/2504.10157/images/c89d917b41e40f4b606b5a27431a400adae43e74cb416cbd6208d68554aee7ab.jpg b/data/2025/2504_10xxx/2504.10157/images/c89d917b41e40f4b606b5a27431a400adae43e74cb416cbd6208d68554aee7ab.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d983770c764c5f6fe8e01f0d804aeba6e7aa8362
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/c89d917b41e40f4b606b5a27431a400adae43e74cb416cbd6208d68554aee7ab.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f225dd497e4c5079eb3a1155f86a8f34132a61619bccc8f64af9c5f995ca605
+size 4058
diff --git a/data/2025/2504_10xxx/2504.10157/images/cb26ecd99c58640d8eb4effc287a4e8bee0c4214555edb59f3d149882199fb82.jpg b/data/2025/2504_10xxx/2504.10157/images/cb26ecd99c58640d8eb4effc287a4e8bee0c4214555edb59f3d149882199fb82.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..34132c50711c99650c1be22eed09aa4f90724f49
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/cb26ecd99c58640d8eb4effc287a4e8bee0c4214555edb59f3d149882199fb82.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3eb2f04c62f25e87422ffdbfee37692b7bb54f9ae0b551ebed28d8633f0c466f
+size 19407
diff --git a/data/2025/2504_10xxx/2504.10157/images/ee37a034cc7bfd54903875bca0d8ebdb66151851f1d8c1aae3fe11feeefe474d.jpg b/data/2025/2504_10xxx/2504.10157/images/ee37a034cc7bfd54903875bca0d8ebdb66151851f1d8c1aae3fe11feeefe474d.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..6acec4a8febaee48dec83a05df09cad7222bc096
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/ee37a034cc7bfd54903875bca0d8ebdb66151851f1d8c1aae3fe11feeefe474d.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6c6d229ff88d42a4f0f113c40a10bde93437b04d6b3308017086b2e2e014be8
+size 113002
diff --git a/data/2025/2504_10xxx/2504.10157/images/f30fe370c7d43ea99207a6f091595cf5559b34c28ba881f238d60524e7804d01.jpg b/data/2025/2504_10xxx/2504.10157/images/f30fe370c7d43ea99207a6f091595cf5559b34c28ba881f238d60524e7804d01.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..9187afffae3d4aed2d7543cb1cb74e5f337b7866
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/f30fe370c7d43ea99207a6f091595cf5559b34c28ba881f238d60524e7804d01.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9f9545b5c3bed8fcc14d8871944a33e71af15bed22c162e70ac8aad264d1275
+size 41886
diff --git a/data/2025/2504_10xxx/2504.10157/images/ff781687759cc0943011e2da7ef1096cc2baed0a7ee1909c62cdf62fa3b74e60.jpg b/data/2025/2504_10xxx/2504.10157/images/ff781687759cc0943011e2da7ef1096cc2baed0a7ee1909c62cdf62fa3b74e60.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..35da2a4bb7b346dbb489bba62ddf358afbded800
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/images/ff781687759cc0943011e2da7ef1096cc2baed0a7ee1909c62cdf62fa3b74e60.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7f0da4719e501b2deab90a79d43c0edd36a7c8e7ec9dde952a826e24ee1c7e9
+size 6226
diff --git a/data/2025/2504_10xxx/2504.10157/layout.json b/data/2025/2504_10xxx/2504.10157/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..b3c75e6ccaa4039e2697f2256dfa157c4768d207
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10157/layout.json
@@ -0,0 +1,11165 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 120,
+ 97,
+ 491,
+ 155
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 120,
+ 97,
+ 491,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 97,
+ 491,
+ 155
+ ],
+ "type": "text",
+ "content": "SocioVerse: A World Model for Social Simulation Powered by LLM Agents and A Pool of 10 Million Real-World Users"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": "Xinnong Zhang"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2\\dagger}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Jiayu Lin"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2\\dagger}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Xinyi Mou"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2\\dagger}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Shiyue Yang"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Xiawei Liu"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Libo Sun"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Hanjia Lyu"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{3}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Yihang Yang"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Weihong Qi"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{4}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Yue Chen"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Guanying Li"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Ling Yan"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{5}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Yao Hu"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{5}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Siming Chen"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Yu Wang"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Xuanjing Huang"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Jiebo Luo"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{3}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Shiping Tang"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Libo Wu"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Baohua Zhou"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "text",
+ "content": ", Zhongyu Wei"
+ },
+ {
+ "bbox": [
+ 114,
+ 194,
+ 499,
+ 243
+ ],
+ "type": "inline_equation",
+ "content": "^{1,2}"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "text",
+ "content": "Shanghai Innovation Institute, "
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "text",
+ "content": "Fudan University, "
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "inline_equation",
+ "content": "^{3}"
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "text",
+ "content": "University of Rochester, "
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "inline_equation",
+ "content": "^{4}"
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "text",
+ "content": "Indiana University, "
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "inline_equation",
+ "content": "^{5}"
+ },
+ {
+ "bbox": [
+ 173,
+ 242,
+ 436,
+ 287
+ ],
+ "type": "text",
+ "content": "Xiaohongshu Inc. zywei@fudan.edu.cn \nSocioVerse: https://github.com/FudanDISC/SocioVerse"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 112,
+ 300,
+ 491,
+ 497
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 112,
+ 300,
+ 491,
+ 497
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 300,
+ 491,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 300,
+ 491,
+ 497
+ ],
+ "type": "image",
+ "image_path": "ee37a034cc7bfd54903875bca0d8ebdb66151851f1d8c1aae3fe11feeefe474d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 104,
+ 499,
+ 504,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 499,
+ 504,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 499,
+ 504,
+ 522
+ ],
+ "type": "text",
+ "content": "Figure 1: An illustration of the SocioVerse in the case of Ukraine issue. The alignment challenges are well handled regarding environment, user, scenario, and behavior."
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 281,
+ 534,
+ 329,
+ 547
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 281,
+ 534,
+ 329,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 281,
+ 534,
+ 329,
+ 547
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 140,
+ 558,
+ 470,
+ 702
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 558,
+ 470,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 558,
+ 470,
+ 702
+ ],
+ "type": "text",
+ "content": "Social simulation is transforming traditional social science research by modeling human behavior through interactions between virtual individuals and their environments. With recent advances in large language models (LLMs), this approach has shown growing potential in capturing individual differences and predicting group behaviors. However, existing methods face alignment challenges related to the environment, target users, interaction mechanisms, and behavioral patterns. To this end, we introduce SocioVerse, an LLM-agent-driven world model for social simulation. Our framework features four powerful alignment components and a user pool of 10 million real individuals. To validate its effectiveness, we conducted large-scale simulation experiments across three distinct domains: politics, news, and economics. Results demonstrate that SocioVerse can reflect large-scale population dynamics while ensuring diversity, credibility, and representativeness through standardized procedures and minimal manual adjustments."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 14,
+ 224,
+ 35,
+ 567
+ ],
+ "type": "aside_text",
+ "angle": 270,
+ "lines": [
+ {
+ "bbox": [
+ 14,
+ 224,
+ 35,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 14,
+ 224,
+ 35,
+ 567
+ ],
+ "type": "text",
+ "content": "arXiv:2504.10157v3 [cs.CL] 15 Jul 2025"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 118,
+ 710,
+ 288,
+ 722
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 710,
+ 288,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 710,
+ 288,
+ 722
+ ],
+ "type": "text",
+ "content": "These authors contribute equally to this work."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 106,
+ 71,
+ 191,
+ 83
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 71,
+ 191,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 71,
+ 191,
+ 83
+ ],
+ "type": "text",
+ "content": "1 Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 105,
+ 506,
+ 160
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 105,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 105,
+ 506,
+ 160
+ ],
+ "type": "text",
+ "content": "The study of human behavior aims to understand how individuals and groups act in various social contexts and serves as a cornerstone of social science research. Traditionally, this has been accomplished using methods such as surveys, interviews, and observations [10, 18, 44]. However, these approaches often encounter challenges, including high costs, limited sample sizes, and ethical concerns. As a result, researchers have resorted to alternative methods for studying human behavior."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 164,
+ 504,
+ 253
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 164,
+ 504,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 164,
+ 504,
+ 253
+ ],
+ "type": "text",
+ "content": "Social simulation has emerged as an effective method for addressing this issue, where researchers use agents to model human behavior, observe their reactions, and translate these findings into insights about human behavior [48, 50]. By assigning behavioral rules to autonomous agents, researchers can explore how micro-level decisions lead to emergent macro-level patterns through the agent-based models [11, 21]. This approach enables capturing specific groups' preferences on particular topics and forecasting potential social dynamics. Furthermore, recent advancements in large language models (LLMs) have significantly enhanced agents' reasoning and decision-making capabilities, enabling them to operate and interact within increasingly realistic and complex environments [3, 35, 37]."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 258,
+ 506,
+ 313
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 258,
+ 506,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 258,
+ 506,
+ 313
+ ],
+ "type": "text",
+ "content": "Recent studies have explored social simulation across various levels and scenarios, from mimicking well-known individuals and mirroring specific situations to modeling large-scale social dynamics [4, 29, 34, 36, 49, 60]. However, they share a common challenge: alignment between the simulated environment and the real world, which manifests across multiple dimensions and raises several key questions that remain to be addressed, as shown in Figure 1."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 385,
+ 328
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 385,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 385,
+ 328
+ ],
+ "type": "text",
+ "content": "Q1. How to align the simulated environment with the real world?"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 104,
+ 329,
+ 506,
+ 406
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 329,
+ 506,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 329,
+ 506,
+ 406
+ ],
+ "type": "text",
+ "content": "In the real world, new events occur every day and new content is continuously generated. The behavior of real users is rooted in these ever-evolving social contexts and policy agendas. However, the static knowledge of LLMs prevents them from aligning with the dynamic nature of the real-world social environment [2, 15]. There is a gap between the simulated context and the real world, which results in discrepancies between the simulation process and outcomes compared to those in reality. Therefore, it is necessary to establish an update mechanism to keep the simulated environment synchronized with the real world."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 411,
+ 373,
+ 422
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 373,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 373,
+ 422
+ ],
+ "type": "text",
+ "content": "Q2. How to align simulated agents with target users precisely?"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 422,
+ 504,
+ 487
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 422,
+ 504,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 422,
+ 504,
+ 487
+ ],
+ "type": "text",
+ "content": "The composition of users in the real world is both complex and diverse, making it impractical to enumerate all users in every scenario. Therefore, it is essential to identify target users whose distribution aligns with that of the users in the corresponding scenario, thereby accurately reflecting the real-world composition and relationships [17, 45]. Based on this, precise target user simulation also requires providing agents with a detailed and comprehensive description of the corresponding users, often involving the integration of high-fidelity demographic, contextual, and behavioral data."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 496,
+ 503
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 496,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 496,
+ 503
+ ],
+ "type": "text",
+ "content": "Q3. How to align the interaction mechanism with the real world among different scenarios?"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 104,
+ 503,
+ 506,
+ 580
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 503,
+ 506,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 503,
+ 506,
+ 580
+ ],
+ "type": "text",
+ "content": "The diversity of social interactions presents challenges in social simulation design, requiring deliberate choices regarding the number of individuals, social structures, interaction patterns, and message dissemination mechanisms, to align with the real world. This often results in independently constructed task-specific simulation pipelines performing repetitive work, which reduces their generalizability and scalability [26, 58]. Therefore, there is a need for unified simulation frameworks based on systematic categorization to standardize simulation components and facilitate extensibility across different social scenarios."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 399,
+ 597
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 399,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 399,
+ 597
+ ],
+ "type": "text",
+ "content": "Q4. How to align the behavioral pattern with the real-world groups?"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 596,
+ 505,
+ 662
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 596,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 596,
+ 505,
+ 662
+ ],
+ "type": "text",
+ "content": "When the environment perceived by agents, the user composition, and the interaction mechanisms are aligned with the real world, agents are expected to exhibit responses consistent with those of the corresponding real users. However, current LLMs exhibit inherent bias and limitations in such reasoning, failing to infer different types of user behaviors [16, 60]. Therefore, it is necessary to systematically collect behavior-driving factors across different user characteristics and adopt appropriate modeling approaches to effectively capture diverse behavior patterns."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 104,
+ 667,
+ 505,
+ 723
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 667,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 667,
+ 505,
+ 723
+ ],
+ "type": "text",
+ "content": "In this paper, we propose SocioVerse, a world model for social simulation driven by LLM-based agents based on a large-scale real-world user pool. As shown in Figure 2, we design modular components to address the above questions. The Social Environment injects up-to-date and external real-world information into the simulation. The User Engine and Scenario Engine respectively reconstruct realistic user context and orchestrate the simulation process to align the simulation with"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "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": "2"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 1
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 112,
+ 77,
+ 495,
+ 316
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 112,
+ 77,
+ 495,
+ 316
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 77,
+ 495,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 77,
+ 495,
+ 316
+ ],
+ "type": "image",
+ "image_path": "af88ad03ac5ce1c47c53dd172206932141236f13f7154e620648f8b4af53663e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 104,
+ 326,
+ 504,
+ 371
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 326,
+ 504,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 326,
+ 504,
+ 371
+ ],
+ "type": "text",
+ "content": "Figure 2: An illustration of SocioVerse framework involving 4 powerful parts. The social environment provides an updated context for the simulation. During the simulation, the behavior engine takes the simulation setting, user profiles, and social information from the scenario engine, user engine, and social environment, respectively, and generates the results according to the query."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 403,
+ 504,
+ 427
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 403,
+ 504,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 403,
+ 504,
+ 427
+ ],
+ "type": "text",
+ "content": "the real world. Given this rich contextual setup, the Behavior Engine then drives agents to reproduce human behaviors accordingly."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 104,
+ 430,
+ 504,
+ 487
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 430,
+ 504,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 430,
+ 504,
+ 487
+ ],
+ "type": "text",
+ "content": "To support the framework, we construct a user pool of 10 million individuals by collecting real-world social media data to power the user engine. Comparable in scale to the entire populations of Hungary or Greece, this extensive pool enables diverse and large-scale social simulations. For any customized simulation task, various sampling strategies can be applied to extract target user groups from the pool to support the simulation process."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 491,
+ 506,
+ 557
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 491,
+ 506,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 491,
+ 506,
+ 557
+ ],
+ "type": "text",
+ "content": "We conduct three simulations using the SocioVerse framework, each differing in research domain, user composition, and social environment: (a) presidential election prediction, (b) breaking news feedback, and (c) national economic survey. For each task, we compare the simulation results with real-world situations. Extensive and comprehensive experiments demonstrate that our framework serves as a robust foundation for building standardized and accurate large-scale social simulations. In summary, our key contributions are as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 132,
+ 576,
+ 506,
+ 706
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 8,
+ "blocks": [
+ {
+ "bbox": [
+ 132,
+ 576,
+ 504,
+ 609
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 576,
+ 504,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 576,
+ 504,
+ 609
+ ],
+ "type": "text",
+ "content": "- SocioVerse: We propose a world model for social simulation comprising four powerful alignment modules, enabling diverse and trustworthy social simulations (as illustrated in Figure 2)."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 132,
+ 624,
+ 506,
+ 658
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 624,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 624,
+ 506,
+ 658
+ ],
+ "type": "text",
+ "content": "- 10M User Pool: A user pool of 10 million individuals, constructed from real-world behavioral data, enables large-scale and diverse social simulations, ranging from small interest groups to large citizen communities."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 132,
+ 672,
+ 504,
+ 706
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 672,
+ 504,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 672,
+ 504,
+ 706
+ ],
+ "type": "text",
+ "content": "- Three Illustrative Simulations: We demonstrate the framework's capabilities through three distinct scenarios: presidential election prediction, breaking news feedback, and a national economic survey, providing a foundation for future research."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "sub_type": "text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 308,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 308,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 308,
+ 750
+ ],
+ "type": "text",
+ "content": "3"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 2
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 171,
+ 83
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 171,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 171,
+ 83
+ ],
+ "type": "text",
+ "content": "2 Methods"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 96,
+ 194,
+ 107
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 194,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 96,
+ 194,
+ 107
+ ],
+ "type": "text",
+ "content": "Overall Framework"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 118,
+ 506,
+ 174
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 118,
+ 506,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 118,
+ 506,
+ 174
+ ],
+ "type": "text",
+ "content": "The SocioVerse framework follows a structured pipeline to achieve realistic social simulation results, as shown in Figure 2: (1) Social Environment collects updated information and contextual knowledge. Within the simulation environment, (2) User Engine aligns the simulated agents with target users, (3) Scenario Engine aligns the interaction structure with diverse scenarios, and (4) Behavior Engine aligns the behavioral pattern with real-world target groups."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 215,
+ 198
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 215,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 215,
+ 198
+ ],
+ "type": "text",
+ "content": "2.1 Social Environment"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 209,
+ 506,
+ 243
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 209,
+ 506,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 209,
+ 506,
+ 243
+ ],
+ "type": "text",
+ "content": "Function The social environment provides event-related context to align the simulation environment with real-world conditions. By integrating up-to-date events, social statistics, and preference content into LLM-based agents, it enhances the realism of the simulation and improve agent decision-making."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 104,
+ 255,
+ 504,
+ 289
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 255,
+ 504,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 255,
+ 504,
+ 289
+ ],
+ "type": "text",
+ "content": "Components The social environment should encompass as much real-world social, cultural, and technological context as possible. It can be broadly categorized into three types: social structural information, social dynamic information, and personalized context."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 293,
+ 505,
+ 361
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 293,
+ 505,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 293,
+ 505,
+ 361
+ ],
+ "type": "text",
+ "content": "Social Structure: Social structural information provides agents with a rich knowledge base encompassing demographic distributions, cultural norms, urban infrastructures, and collective behavior patterns [57]. This data allows agents to behave in a way that aligns with the typical characteristics of their assigned demographic or geographic profile. For example, by incorporating regional dialect preferences, work-life habits, and common social values, the simulation can more accurately reflect public discourse trends, mobility behaviors, and economic interactions."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 365,
+ 504,
+ 421
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 365,
+ 504,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 365,
+ 504,
+ 421
+ ],
+ "type": "text",
+ "content": "Social Dynamics: Social dynamics encompass time-sensitive content continuously generated in the real world, such as news events and policy changes. Typically, this engine maintains an up-to-date event base to continuously collect real-world event news from mainstream news, and all the news articles contain time stamps and event-related tags so that LLM-based agents can comb through the timeline of the events and react accordingly [37]."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 104,
+ 425,
+ 505,
+ 480
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 425,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 425,
+ 505,
+ 480
+ ],
+ "type": "text",
+ "content": "Personalized Context: In addition to the macro social environment, individuals also receive different personalized information feeds. Previous studies have explored that the recommendation system can enhance the behavior diversity of the agent [31, 56, 60]. Consequently, the preference content component constructs relevant posts and pushes them to agents according to their social interaction network and interesting topics."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 183,
+ 506
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 183,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 183,
+ 506
+ ],
+ "type": "text",
+ "content": "2.2 User Engine"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 515,
+ 506,
+ 539
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 515,
+ 506,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 515,
+ 506,
+ 539
+ ],
+ "type": "text",
+ "content": "Function The user engine aligns simulated agents with a rich set of real-world user samples, enabling the creation of complex target users within the simulation."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 552,
+ 504,
+ 574
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 552,
+ 504,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 552,
+ 504,
+ 574
+ ],
+ "type": "text",
+ "content": "Components To support diverse user composition and effective user retrieval and description, the user engine incorporates a large user pool and a wide range of user labels."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 104,
+ 578,
+ 506,
+ 689
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 578,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 578,
+ 506,
+ 689
+ ],
+ "type": "text",
+ "content": "User Pools: The user pool is designed to collect extensive digital footprints of individuals across social media platforms, enabling a more comprehensive characterization of real-world behavioral patterns and expression tendencies. To this end, we constructed a user pool covering a variety of social media platforms, including "
+ },
+ {
+ "bbox": [
+ 104,
+ 578,
+ 506,
+ 689
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{X}^1"
+ },
+ {
+ "bbox": [
+ 104,
+ 578,
+ 506,
+ 689
+ ],
+ "type": "text",
+ "content": " and Rednote| Q01 | Voting Behavior |
| Question | ORDER OF MAJOR PARTY CANDIDATE NAMES |
| Value Labels | 1. Democrat first / Republican second2. Republican first / Democrat second |
| Q02 | Social Security |
| Question | Next I am going to read you a list of federal programs. For each one, I would like you to tell me whether you would like to see spending increased, decreased, or kept the same.What about Social Security? Should federal spending on Social Security be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q03 | Education |
| Question | What about public schools? Should federal spending on public schools be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q04 | Immigration |
| Question | What about tightening border security to prevent illegal immigration? Should federal spending on tightening border security to prevent illegal immigration be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q05 | Criminal Justice |
| Question | What about dealing with crime? Should federal spending on dealing with crime be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q06 | Social Welfare |
| Question | What about welfare programs? Should federal spending on welfare programs be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF1. Increased2. Decreased3. Kept the same |
| Q07 | Infrastructure |
| Question | What about building and repairing highways? Should federal spending on building and repairing highways be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF\n1. Increased\n2. Decreased\n3. Kept the same |
| Q08 | Aid to Poor |
| Question | What about aid to the poor? Should federal spending on aid to the poor be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF\n1. Increased\n2. Decreased\n3. Kept the same |
| Q09 | Environment |
| Question | What about protecting the environment? Should federal spending on protecting the environment be increased, decreased, or kept the same? |
| Value Labels | -2. DK/RF\n1. Increased\n2. Decreased\n3. Kept the same |
| Q10 | Government |
| Question | How much do you feel that having elections makes the government pay attention to what the people think? |
| Value Labels | -2. DK/RF\n1. A good deal\n2. Some\n3. Not much |
| Q11 | Economy |
| Question | Which party do you think would do a better job of handling the nation's economy? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q12 | Health Care |
| Question | Which party do you think would do a better job of handling health care? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q13 | Immigration |
| Question | Which party do you think would do a better job of handling immigration? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q14 | Taxes |
| Question | Which party do you think would do a better job of handling taxes? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q15 | Environment |
| Question | Which party do you think would do a better job of handling the environment? |
| Value Labels | -2. DK/RF\n1. Democrats would do a better job\n2. Not much difference between them\n3. Republicans would do a better job |
| Q16 | Education |
| Question | Some people think the government should provide fewer services even in areas such as health and education in order to reduce spending.\nOther people feel it is important for the government to provide many more services even if it means an increase in spending.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Government should provide fewer services\n2. Neutral\n3. Government should provide more services |
| Q17 | Defense |
| Question | Some people believe that we should spend less money for defense.\nOthers feel that defense spending should be increased.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Decrease defense spending\n2. Neutral\n3. Increase defense spending |
| Q18 | Health Care |
| Question | There is much concern about the rapid rise in medical and hospital costs.\nSome people feel there should be a government insurance plan which would cover all medical and hospital expenses for everyone.\nOthers feel that all medical expenses should be paid by individuals through private insurance plans like Blue Cross or other company paid plans.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Government insurance plan\n2. Neutral\n3. Private insurance plan |
| Q19 | Social Welfare |
| Question | Some people feel the government in Washington should see to it that every person has a job and a good standard of living.\nOthers think the government should just let each person get ahead on their own.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Government should see to jobs and standard of living\n2. Neutral\n3. Government should let each person get ahead on own |
| Q20 | Aid to Blacks |
| Question | Some people feel that the government in Washington should make every effort to improve the social and economic position of blacks.\nOthers feel that the government should not make any special effort to help blacks because they should help themselves.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Government should help blacks\n2. Neutral\n3. Blacks should help themselves |
| Q21 | Environment |
| Question | Some people think we need much tougher government regulations on business in order to protect the environment.\nOthers think that current regulations to protect the environment are already too much of a burden on business.\nAnd, of course, some people have a neutral position.\nWhich of the following best describes your view? |
| Value Labels | -2. DK/RF\n1. Tougher regulations on business needed to protect environment\n2. Neutral\n3. Regulations to protect environment already too much a burden on business |
| Q22 | Abortion |
| Question | Would you be pleased, upset, or neither pleased nor upset if the Supreme Court reduced abortion rights? |
| Value Labels | -2. DK/RF\n1. Pleased\n2. Upset\n3. Neither pleased nor upset |
| Q23 | Criminal Justice |
| Question | Do you favor or oppose the death penalty for persons convicted of murder? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose |
| Q24 | US Position in World |
| Question | Do you agree or disagree with this statement: ‘This country would be better off if we just stayed home and did not concern ourselves with problems in other parts of the world.’ |
| Value Labels | -2. DK/RF\n1. Agree\n2. Disagree |
| Q25 | US Position in World |
| Question | How willing should the United States be to use military force to solve international problems? |
| Value Labels | -2. DK/RF\n1. Willing\n2. Moderately willing\n3. Not willing |
| Q26 | Inequality |
| Question | Do you think the difference in incomes between rich people and poor people in the United States today is larger, smaller, or about the same as it was 20 years ago? |
| Value Labels | -2. DK/RF\n1. Larger\n2. Smaller\n3. About the same |
| Q27 | Environment |
| Question | Do you think the federal government should be doing more about rising temperatures, should be doing less, or is it currently doing the right amount? |
| Value Labels | -2. DK/RF\n1. Should be doing more\n2. Should be doing less\n3. Is currently doing the right amount |
| Q28 | Parental Leave |
| Question | Do you favor, oppose, or neither favor nor oppose requiring employers to offer paid leave to parents of new children? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q29 | LGBTQ+ Rights |
| Question | Do you think business owners who provide wedding-related services should be allowed to refuse services to same-sex couples if same-sex marriage violates their religious beliefs, or do you think business owners should be required to provide services regardless of a couple's sexual orientation? |
| Value Labels | -2. DK/RF\n1. Should be allowed to refuse\n2. Should be required to provide services |
| Q30 | LGBTQ+ Rights |
| Question | Should transgender people - that is, people who identify themselves as the sex or gender different from the one they were born as - have to use the bathrooms of the gender they were born as, or should they be allowed to use the bathrooms of their identified gender? |
| Value Labels | -2. DK/RF\n1. Have to use the bathrooms of the gender they were born as\n2. Be allowed to use the bathrooms of their identified gender |
| Q31 | LGBTQ+ Rights |
| Question | Do you favor or oppose laws to protect gays and lesbians against job discrimination? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose |
| Q32 | LGBTQ+ Rights |
| Question | Do you think gay or lesbian couples should be legally permitted to adopt children? |
| Value Labels | -2. DK/RF\n1. Yes\n2. No |
| Q33 | LGBTQ+ Rights |
| Question | Which comes closest to your view? You can just tell me the number of your choice. |
| Value Labels | -2. DK/RF 1. Gay and lesbian couples should be allowed to legally marry\n2. Gay and lesbian couples should be allowed to form civil unions but not legally marry\n3. There should be no legal recognition of gay or lesbian couples' relationship |
| Q34 | Immigration |
| Question | Some people have proposed that the U.S. Constitution should be changed so that the children of unauthorized immigrants do not automatically get citizenship if they are born in this country.\nDo you favor, oppose, or neither favor nor oppose this proposal? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q35 | Immigration |
| Question | What should happen to immigrants who were brought to the U.S. illegally as children and have lived here for at least 10 years and graduated high school here? Should they be sent back where they came from, or should they be allowed to live and work in the United States? |
| Value Labels | -2. DK/RF\n1. Should be sent back where they came from\n2. Should be allowed to live and work in the US |
| Q36 | Immigration |
| Question | Do you favor, oppose, or neither favor nor oppose building a wall on the U.S. border with Mexico? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q37 | Unrest |
| Question | During the past few months, would you say that most of the actions taken by protestors to get the things they want have been violent, or have most of these actions by protestors been peaceful, or have these actions been equally violent and peaceful? |
| Value Labels | -2. DK/RF\n1. Mostly violent\n2. Mostly peaceful\n3. Equally violent and peaceful |
| Q38 | Government |
| Question | Do you think it is better when one party controls both the presidency and Congress, better when control is split between the Democrats and Republicans, or doesn’t it matter? |
| Value Labels | -2. DK/RF\n1. Better when one party controls both\n2. Better when control is split\n3. It doesn’t matter |
| Q39 | Government |
| Question | Would you say the government is pretty much run by a few big interests looking out for themselves or that it is run for the benefit of all the people? |
| Value Labels | -2. DK/RF\n1. Run by a few big interests\n2. For the benefit of all the people |
| Q40 | Government |
| Question | Do you think that people in government waste a lot of the money we pay in taxes, waste some of it, or don’t waste very much of it? |
| Value Labels | -2. DK/RF\n1. Waste a lot\n2. Waste some\n3. Don’t waste very much |
| Q41 | Election Integrity |
| Question | Do you favor, oppose, or neither favor nor oppose allowing convicted felons to vote once they complete their sentence? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q42 | Democratic Norms |
| Question | How important is it that news organizations are free to criticize political leaders? |
| Value Labels | -2. DK/RF\n1. Not important\n2. Moderately important\n3. Important |
| Q43 | Democratic Norms |
| Question | How important is it that the executive, legislative, and judicial branches of government keep one another from having too much power? |
| Value Labels | -2. DK/RF\n1. Not important\n2. Moderately important\n3. Important |
| Q44 | Democratic Norms |
| Question | How important is it that elected officials face serious consequences if they engage in misconduct? |
| Value Labels | -2. DK/RF\n1. Not important\n2. Moderately important\n3. Important |
| Q45 | Democratic Norms |
| Question | How important is it that people agree on basic facts even if they disagree politically? |
| Value Labels | -2. DK/RF\n1. Not important\n2. Moderately important\n3. Important |
| Q46 | Democratic Norms |
| Question | Would it be helpful, harmful, or neither helpful nor harmful if U.S. presidents could work on the country’s problems without paying attention to what Congress and the courts say? |
| Value Labels | -2. DK/RF\n1. Helpful\n2. Harmful\n3. Neither helpful nor harmful |
| Q47 | Democratic Norms |
| Question | Do you favor, oppose, or neither favor nor oppose elected officials restricting journalists’ access to information about government decision-making? |
| Value Labels | -2. DK/RF\n1. Favor\n2. Oppose\n3. Neither favor nor oppose |
| Q48 | Gender Resentment |
| Question | ‘Many women interpret innocent remarks or acts as being sexist.’\nDo you agree, neither agree nor disagree, or disagree with this statement? |
| Value Labels | -2. DK/RF/technical error\n1. Agree\n2. Neither agree nor disagree\n3. Disagree |
| Q49 | Gender Resentment |
| Question | ‘Women seek to gain power by getting control over men.’\nDo you agree, neither agree nor disagree, or disagree with this statement? |
| Value Labels | -2. DK/RF/technical error\n1. Agree\n2. Neither agree nor disagree\n3. Disagree |
",
+ "image_path": "a66412febde24e12e177a5c22dc9307635f04919238d13f32d8145fd1596adf3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 312,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 312,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 312,
+ 750
+ ],
+ "type": "text",
+ "content": "20"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 19
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 69,
+ 524,
+ 681
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 524,
+ 681
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "text",
+ "content": "21"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 20
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 71,
+ 524,
+ 712
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 71,
+ 524,
+ 712
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "22"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 21
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 71,
+ 524,
+ 714
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 71,
+ 524,
+ 714
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "text",
+ "content": "23"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 22
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 71,
+ 524,
+ 712
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 71,
+ 524,
+ 712
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "24"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 23
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 69,
+ 524,
+ 680
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 524,
+ 680
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "text",
+ "content": "25"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 24
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 69,
+ 525,
+ 691
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 525,
+ 691
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "26"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 25
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 71,
+ 525,
+ 300
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 71,
+ 525,
+ 300
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 318,
+ 330
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 318,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 318,
+ 330
+ ],
+ "type": "text",
+ "content": "D.2 Questionnaire for Breaking News Feedback"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 343,
+ 525,
+ 721
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 525,
+ 721
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 525,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 525,
+ 721
+ ],
+ "type": "table",
+ "html": "| Q01 | Public Cognition (PC) |
| Question | I have heard of ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q02 | Public Cognition (PC) |
| Question | Many people around me use ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q03 | Public Cognition (PC) |
| Question | I have a deep understanding of ChatGPT's functions and applications. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q04 | Perceived Risks (PR) |
| Question | ChatGPT may lead to the widespread dissemination of false information. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q05 | Perceived Risks (PR) |
| Question | ChatGPT may reduce human thinking ability and creativity. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q06 | Perceived Risks (PR) |
| Question | The development of ChatGPT may replace certain jobs, and I am deeply concerned about this. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q07 | Perceived Benefits (PB) |
| Question | ChatGPT will definitely improve my work and study efficiency. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q08 | Perceived Benefits (PB) |
| Question | ChatGPT helps broaden my knowledge and provides me with new perspectives and ideas. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q09 | Perceived Benefits (PB) |
| Question | ChatGPT promotes technological innovation and development in related fields. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q10 | Trust (TR) |
| Question | I fully trust the team developing ChatGPT to manage and guide its development responsibly. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q11 | Trust (TR) |
| Question | I have strong confidence in the accuracy and reliability of the information generated by ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q12 | Trust (TR) |
| Question | I believe that the future application of ChatGPT will be effectively regulated. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q13 | Fairness (FA) |
| Question | The opportunities to use ChatGPT are distributed fairly among different groups of people. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q14 | Fairness (FA) |
| Question | I find the distribution of benefits brought by ChatGPT to be fair. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q15 | Fairness (FA) |
| Question | I believe that the decision-making process for the development and promotion of ChatGPT is fully transparent and adequately reflects public interests. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q16 | Public Acceptance (PA) |
| Question | Overall, I strongly welcome the emergence of ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q17 | Public Acceptance (PA) |
| Question | I am definitely willing to use ChatGPT in my work or studies. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
| Q18 | Public Acceptance (PA) |
| Question | I strongly support increased investment in the research and development of AI technologies like ChatGPT. |
| Value Labels | 1. Disagree\n2. Partially disagree\n3. Neutral\n4. Partially agree\n5. Agree |
",
+ "image_path": "7ce59a581d9686e1bdc3056be8b7786b1ee7c8b56671e62f94b9629946ea1c5e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "27"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 26
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 105,
+ 69,
+ 524,
+ 668
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 105,
+ 69,
+ 524,
+ 668
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "28"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 27
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 105,
+ 69,
+ 525,
+ 708
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 105,
+ 69,
+ 525,
+ 708
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "29"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 28
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 70,
+ 523,
+ 158
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 523,
+ 158
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 173,
+ 324,
+ 186
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 324,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 324,
+ 186
+ ],
+ "type": "text",
+ "content": "D.3 Questionnaire for National Economic Survey"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 198,
+ 525,
+ 703
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 525,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 525,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 525,
+ 703
+ ],
+ "type": "table",
+ "html": "| Q01 | Food |
| Question | What is your average monthly expenditure on food (including dining out)? (Unit: CNY) |
| Value Labels | A. Below 500 CNY\nB. 501-650 CNY\nC. 651-800 CNY\nD. 801-1000 CNY\nE. Above 1000 CNY |
| Q02 | Food |
| Question | Do you think your current spending on food, tobacco, and alcohol is too high relative to your income? |
| Value Labels | A. Yes\nB. No\nC. Acceptable |
| Q03 | Clothing |
| Question | What is your average monthly expenditure on clothing (including apparel, shoes, and accessories)? (Unit: CNY) |
| Value Labels | A. Below 50 CNY\nB. 51-100 CNY\nC. 101-150 CNY\nD. 151-200 CNY\nE. Above 200 CNY |
| Q04 | Clothing |
| Question | How much economic pressure do you feel from clothing expenses? |
| Value Labels | A. Very low, almost no pressure\nB. Moderate, some pressure but manageable\nC. High, requires careful spending\nD. Very high, affects spending in other areas |
| Q05 | Household |
| Question | What is your average monthly housing expenditure? (Including rent, mortgage, property fees, maintenance, etc.) (Unit: CNY) |
| Value Labels | A. Below 200 CNY\nB. 201-500 CNY\nC. 501-800 CNY\nD. 801-1200 CNY\nE. Above 1200 CNY |
| Q06 | Household |
| Question | What percentage of your monthly income is spent on housing? (Including rent, mortgage, property fees, maintenance, etc.) |
| Value Labels | A. Below 10% \nB. 10%-20% \nC. 21%-30% \nD. 31%-40% \nE. Above 40% |
| Q07 | Daily Service |
| Question | What is your average monthly expenditure on daily necessities (personal care, house- hold items, cleaning supplies, etc.) and services (housekeeping, repairs, beauty, pet services, etc.)? (Unit: CNY) |
| Value Labels | A. Below 80 CNY \nB. 81-120 CNY \nC. 121-160 CNY \nD. 161-200 CNY \nE. Above 200 CNY |
| Q08 | Transportation & Communication |
| Question | What is your average monthly expenditure on transportation (public transport, taxis, fuel, parking, etc.) and communication (mobile and internet fees)? (Unit: CNY) |
| Value Labels | A. Below 200 CNY \nB. 201-300 CNY \nC. 301-400 CNY \nD. 401-500 CNY \nE. Above 500 CNY |
| Q09 | Education & Entertainment |
| Question | What is your average monthly expenditure on education (tuition, training, books, etc.) and cultural entertainment (movies, performances, games, fitness, cultural activities, etc.)? (Unit: CNY) |
| Value Labels | A. Below 100 CNY \nB. 101-200 CNY \nC. 201-300 CNY \nD. 301-400 CNY \nE. Above 400 CNY |
| Q10 | Education & Entertainment |
| Question | Can you easily afford your current education, cultural, and entertainment expenses? |
| Value Labels | A. Yes, spending does not affect other areas \nB. Barely, needs some control \nC. Not really, affects other expenditures \nD. No, it creates significant financial pressure |
| Q11 | Medical |
| Question | What is your average monthly expenditure on healthcare (medications, medical services, health management, etc.)? (Unit: CNY) |
| Value Labels | A. Below 100 CNY \nB. 101-200 CNY \nC. 201-300 CNY \nD. 301-400 CNY \nE. Above 400 CNY |
| Q12 | Medical |
| Question | Have you purchased private medical or health insurance for yourself or your family? |
| Value Labels | A. Yes \nB. Not yet, but planning to \nC. No, and no plans to |
| Q13 | Others |
| Question | Besides food, clothing, housing, daily necessities and services, transportation, education, culture, and healthcare, what is your average monthly expenditure on other areas (e.g., hobbies, charitable donations, investment, etc.)? (Unit: CNY) |
| Value Labels | A. Below 30 CNY\nB. 31-60 CNY\nC. 61-90 CNY\nD. 91-120 CNY\nE. Above 120 CNY |
| Q14 | Overall |
| Question | How would you evaluate the impact of your current consumption level on your household (or personal) financial situation? |
| Value Labels | A. Comfortable, can moderately increase spending\nB. Average, can maintain current spending\nC. Tight, need to control or reduce spending\nD. Very tight, affects quality of life |
| Q15 | Overall |
| Question | Do you feel that your consumption pressure is too high relative to your income level? |
| Value Labels | A. Yes\nB. No\nC. Not sure |
| Q16 | Overall |
| Question | If your income increases, which consumption areas would you most like to expand or improve? (Multiple choices allowed) |
| Value Labels | A. Food and alcohol\nB. Clothing\nC. Housing\nD. Daily necessities and services\nE. Transportation and communication\nF. Education, culture, and entertainment\nG. Healthcare\nH. Other goods and services |
| Q17 | Overall |
| Question | What is your consumption expectation for the next six months to a year? |
| Value Labels | A. Will continue to increase\nB. Will remain roughly the same\nC. Will moderately decrease\nD. Uncertain |
",
+ "image_path": "18070cfafc0e2c3f8693e46f9ae0e339394f7c42e0f7d13109ea8126d6c4d3ff.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 312,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 312,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 312,
+ 750
+ ],
+ "type": "text",
+ "content": "30"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 29
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 71,
+ 524,
+ 717
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 71,
+ 524,
+ 717
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 310,
+ 750
+ ],
+ "type": "text",
+ "content": "31"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 30
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 69,
+ 524,
+ 537
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 106,
+ 69,
+ 524,
+ 537
+ ],
+ "lines": [],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body",
+ "lines_deleted": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 741,
+ 311,
+ 750
+ ],
+ "type": "text",
+ "content": "32"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 31
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/data/2025/2504_10xxx/2504.10160/b4495eb2-d145-42fa-9b61-d633ab2f00d9_content_list.json b/data/2025/2504_10xxx/2504.10160/b4495eb2-d145-42fa-9b61-d633ab2f00d9_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..a572f7f6ec6afe8d74e456ce333d6d73d165d2e3
--- /dev/null
+++ b/data/2025/2504_10xxx/2504.10160/b4495eb2-d145-42fa-9b61-d633ab2f00d9_content_list.json
@@ -0,0 +1,3556 @@
+[
+ {
+ "type": "text",
+ "text": "MT-R1-Zero: Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 89,
+ 821,
+ 130
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Zhaopeng Feng $^{1}$ Shaosheng Cao $^{2\\dagger}$ Jiahan Ren $^{1}$ Jiayuan Su $^{1}$",
+ "bbox": [
+ 213,
+ 146,
+ 781,
+ 164
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Ruizhe Chen $^{1}$ Yan Zhang $^{1}$ Zhe Xu $^{2}$ Yao Hu $^{2}$ Jian Wu $^{1}$ Zuozhu Liu $^{1\\dagger}$",
+ "bbox": [
+ 154,
+ 164,
+ 843,
+ 180
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "$^{1}$ Zhejiang University $^{2}$ Xiaohongshu Inc.",
+ "bbox": [
+ 324,
+ 180,
+ 670,
+ 197
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "{zhaopeng.23,zuozhuliu}@intl.zju.edu.cn",
+ "bbox": [
+ 302,
+ 198,
+ 695,
+ 212
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "{caoshaosheng,qiete,xiahou}@xiaohongshu.com",
+ "bbox": [
+ 282,
+ 214,
+ 717,
+ 229
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 260,
+ 260,
+ 339,
+ 275
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Large-scale reinforcement learning (RL) methods have proven highly effective in enhancing the reasoning abilities of large language models (LLMs), particularly for tasks with verifiable solutions such as mathematics and coding. However, applying this idea to machine translation (MT), where outputs are flexibly formatted and difficult to automatically evaluate with explicit rules, remains underexplored. In this work, we introduce MT-R1-Zero, the first open-source adaptation of the R1-Zero RL framework for MT without supervised fine-tuning or cold-start. We propose a rule-metric mixed reward mechanism to guide LLMs towards improved translation quality via emergent reasoning. On the WMT 24 English-Chinese benchmark, our MT-R1-Zero3B-Mix achieves competitive performance, surpassing TowerInstruct-7B-v0.2 by an average of 1.26 points. Meanwhile, our MT-R1-Zero7B-Mix attains a high average score of 62.25 across all metrics, placing it on par with advanced proprietary models such as GPT-4o and Claude-3.5-Sonnet, while the MT-R1-Zero7B-Sem variant achieves state-of-the-art scores on semantic metrics. Moreover, our work exhibits strong generalization capabilities on out-of-distribution MT tasks, robustly supporting multilingual and low-resource settings. Extensive analysis of model behavior across different initializations and reward metrics offers pioneering insight into the critical role of reward design, LLM adaptability, training dynamics, and emergent reasoning patterns within the R1-Zero paradigm for MT. Our code is available at https://github.com/fzp0424/MT-R1-Zero.",
+ "bbox": [
+ 142,
+ 288,
+ 460,
+ 800
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 Introduction",
+ "text_level": 1,
+ "bbox": [
+ 114,
+ 824,
+ 258,
+ 839
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Large-scale Reinforcement Learning (RL) has empowered Large Language Models (LLMs) with strong reasoning capabilities (OpenAI, 2024; Team,",
+ "bbox": [
+ 112,
+ 850,
+ 489,
+ 897
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/6f616ff4dcdccce81b29cf948d93e5785e99354f0a92b7b75e4b01182a86d233.jpg",
+ "image_caption": [
+ "Figure 1: Performance comparison of contemporary LLM-based translation systems on the WMT 24 EN-ZH test set, plotted by average score across BLEU, COMETKiwi, and XCOMET versus model size."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 517,
+ 259,
+ 873,
+ 462
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "2025a,b), demonstrating significant success in tasks such as mathematical reasoning or coding in which answers can be clearly verified. In particular, DeepSeek-R1-Zero (DeepSeek-AI et al., 2025) introduced a pure rule-based RL approach that directly fosters emergent reasoning ability without requirements on structured Chain-of-Thought (CoT) data (Wei et al., 2022; Cui et al., 2025) or sophisticated techniques such as Monte Carlo Tree Search (MCTS) (Silver et al., 2016; Luo et al., 2024; Qi et al., 2024; Guan et al., 2025). However, the applicability of these methods to machine translation (MT) remains challenging and underexplored, as MT outputs are flexibly generated and hard to evaluate automatically with explicit rules.",
+ "bbox": [
+ 507,
+ 549,
+ 884,
+ 790
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Recent work has launched attempts to empower LLMs for MT with reasoning capabilities (Chen et al., 2025; Liu et al., 2025). Early studies investigate explicit reasoning methods for improved translation, such as finetuning with CoT (Wang et al., 2024a) or MCTS (Zhao et al., 2024), where advanced multi-step pipelines with self-correction or long-thought agentic mechanisms are further ex",
+ "bbox": [
+ 507,
+ 793,
+ 884,
+ 921
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "aside_text",
+ "text": "arXiv:2504.10160v1 [cs.CL] 14 Apr 2025",
+ "bbox": [
+ 21,
+ 307,
+ 60,
+ 724
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_footnote",
+ "text": "† Corresponding author.",
+ "bbox": [
+ 141,
+ 906,
+ 295,
+ 921
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_number",
+ "text": "1",
+ "bbox": [
+ 492,
+ 942,
+ 502,
+ 954
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "plored (Feng et al., 2024b; Wang et al., 2024b,a). Another line of work leverages RL to empower LLMs for MT through process reward models or supervised finetuning (SFT) with manually annotated CoT data (Feng et al., 2025; He et al., 2025). However, these methods often depend on manually designed or synthetically generated structured CoT data, rely on complex search algorithms, or require explicit multi-stage prompting, leaving the potential of pure RL-based approaches largely unexplored. Furthermore, the performance reported in these studies often lags behind state-of-the-art (SoTA) open-source or proprietary models.",
+ "bbox": [
+ 112,
+ 84,
+ 492,
+ 294
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Developing pure RL methods to directly enhance the reasoning ability of LLMs for better translation requires answering three key questions: 1) Feasibility: How to design R1-Zero-like RL pipelines with effective reward signals to directly solve MT tasks without binary rule-based rewards; 2) Reasoning capability: Could pure RL training cultivate emergent reasoning abilities and induce models to generate explicit thinking patterns for MT, such as multi-step CoT or verification/reflection; 3) Generalizability: Could the training paradigm generalize across different models (e.g., pre-trained base models, instruction-tuned models, or models pretrained on translation data) or diverse downstream settings (e.g., out-of-distribution, multilingual or low-resource scenarios).",
+ "bbox": [
+ 115,
+ 294,
+ 490,
+ 549
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In this work, we introduce MT-R1-Zero, the first open-source implementation that extends the RL-Zero-like RL training paradigm to MT. We propose a rule-metric mixed reward mechanism that adapts the original rule-based reward concept to effectively guide training in MT scenarios. We explore different rewards optimizing over lexical (Lex), semantic (Sem), and Lex-Sem mixed (Mix) objectives to guide LLMs towards improved translation quality via emergent reasoning. Our experiments demonstrate the efficacy of this approach: as RL training progresses, our MT-R1-Zero-3B-Mix achieves competitive performance, surpassing TowerInstruct-7B-v0.2 by an average of 1.26 points across all metrics (BLEU, COMETKiwi, XCOMET) on the WMT 24 English-Chinese (EN-ZH) benchmark. Meanwhile, our MT-R1-Zero-7B-Mix surpasses LLaMA-3.1-70B by an average of 1.24 points and Qwen2.5-72B by 0.48 points, even on par with top proprietary models such as GPT-4o and Claude-3.5-Sonnet. The MT-R1-Zero further demonstrate promising generalizability across multilingual and low-resource settings.",
+ "bbox": [
+ 115,
+ 551,
+ 490,
+ 921
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Extensive experiments further provide key findings and insight into the adaptation of R1-Zero paradigm to MT. First, we empirically demonstrate that the choice of metric reward plays a pivotal role in steering RL optimization and translation style (semantic or lexical) (Finding 1). Further analysis reveals that MT-R1-Zero induces diverse emergent reasoning patterns, including dynamic language-of-thought transition during translation (Findings 2 and 3). We also identify distinct RL adaptability of different base LLMs (Finding 4). Ablation studies suggest that the pure RL process alone can lead to substantial translation improvements, independent of thinking morbidity (Section 6). Our core contributions are as follows:",
+ "bbox": [
+ 507,
+ 84,
+ 885,
+ 325
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- We present the first open-source implementation of the DeepSeek-R1-Zero paradigm for MT, achieving superior performance across indomain, OOD and generalization MT tasks.",
+ "- Our analysis reveals key findings and recipes for effective R1-Zero adaptation to MT, including reward metric selection, emergent reasoning patterns, training dynamics and LLM adaptability.",
+ "- Extensive experiments and ablations show that pure RL serves as the primary driver of MT improvements, with minimal dependence on forced reasoning or output length, highlighting the significant potential of RL for diverse translation applications and broader language tasks."
+ ],
+ "bbox": [
+ 509,
+ 336,
+ 885,
+ 582
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2 Related Work",
+ "text_level": 1,
+ "bbox": [
+ 509,
+ 590,
+ 665,
+ 606
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "LLM Reasoning with Post-training. Recent research indicates that scaling test-time computation can significantly enhance the ability of LLMs to tackle complex reasoning tasks (OpenAI, 2024; Zeng et al., 2024; Xiang et al., 2025). Many approaches rely on sophisticated techniques such as step-level process reward models (PRMs) that provide granular feedback (Lightman et al., 2024; Yuan et al., 2024; Snell et al., 2024) or MCTS to explore potential reasoning paths (Feng et al., 2023; Qi et al., 2024; Guan et al., 2025). A recent alternative, DeepSeek-R1-Zero (DeepSeek-AI et al., 2025), demonstrated that large-scale pure RL, guided only by formatting rules and correctness of final predictions (rule-based reward), can motivate LLMs to develop self-emergent reasoning processes for complex reasoning tasks. Subsequent work (Hu et al., 2025; Face, 2025) successfully replicated this training paradigm in open-source",
+ "bbox": [
+ 507,
+ 615,
+ 885,
+ 922
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "2",
+ "bbox": [
+ 492,
+ 942,
+ 505,
+ 954
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "models, focusing on mathematical domains. Xie et al. (2025) further demonstrated the effectiveness and generalization capabilities of the R1-Zero paradigm using logic reasoning game problems, while Huang et al. (2025) explored its potential for vision reasoning. Despite its potential, the application of the R1-Zero RL paradigm to complex generation tasks like MT, in which the accuracy/quality of outputs is not rule-based and difficult to validate automatically, remains an open question.",
+ "bbox": [
+ 112,
+ 84,
+ 489,
+ 244
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "LLM Reasoning for MT. Leveraging reasoning to improve MT has garnered increasing attention, as systematically explored in Chen et al. (2025) and Liu et al. (2025). Previous work have designed multi-step processes for MT, e.g., Feng et al. (2024b) introduced an API-based self-correcting framework, and Wang et al. (2024b) employed multi-task training followed by a multistage inference phase. Wang et al. (2024a) integrated a similar procedure into inference-time CoT, using a multi-agent mechanism to synthesize long CoT prompts for English-Chinese literary translation. Efforts have also focused on reward modeling for MT reasoning. Feng et al. (2025) constructed implicit process reward models for translation and explored their effectiveness when combined with test-time search. Recent study further evaluated explicit reasoning for MT using CoT fine-tuning and MCTS to expand test-time computation (Zhao et al., 2024). He et al. (2025) demonstrated that models can acquire reasoning-based translation capabilities through multi-stage training with manually constructed CoT templates.",
+ "bbox": [
+ 115,
+ 247,
+ 489,
+ 615
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "However, these existing methods often necessitate manually designed or synthetically generated structured CoT data, rely on complex search algorithms (MCTS), or require explicit multi-stage prompting (self-correction). The effectiveness of large-scale pure RL training paradigms such as R1-Zero remains unexplored. Furthermore, the performance reported in these studies often lags behind state-of-the-art open-source or proprietary models.",
+ "bbox": [
+ 112,
+ 618,
+ 489,
+ 763
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3 Method",
+ "text_level": 1,
+ "bbox": [
+ 112,
+ 778,
+ 218,
+ 794
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In this section, we present our method that trains a translation model with pure RL using a hybrid reward model. Unlike tasks with fixed correct answers, translation allows for multiple valid outputs, making the evaluation more complicated. In this work, we introduce a rule-metric mixed reward that integrates reasoning format checking with multi-",
+ "bbox": [
+ 112,
+ 808,
+ 489,
+ 921
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "ple translation quality assessment metrics, which is used within the Group Relative Policy Optimization (GRPO) (Shao et al., 2024) algorithm to ensure stable and efficient RL training.",
+ "bbox": [
+ 507,
+ 84,
+ 884,
+ 149
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3.1 Rule-Metric Mixed Reward",
+ "text_level": 1,
+ "bbox": [
+ 507,
+ 160,
+ 773,
+ 174
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In RL, the reward is the main signal that drives model training. DeepSeek-R1-Zero (DeepSeek-AI et al., 2025) employs simple rule-based rewards that check whether the final answer is correct and whether the response follows a specific format. This works well for tasks with fixed format correct answers such as math or coding. However, there is often no single \"correct\" output for MT, impeding the design of rule-based rewards. Fortunately, the MT community has developed many evaluation metrics to measure translation quality. Recent advancements in automated MT evaluation metrics have shown promise in aligning automated assessments with human translation quality judgments (Freitag et al., 2022, 2023). Thus, we design a rule-metric mixed reward, which consists of two parts: a Format Reward that checks output structure, and a Metric Reward that evaluates translation quality. We use a structured prompt template similar to that in DeepSeek-R1-Zero:",
+ "bbox": [
+ 507,
+ 181,
+ 884,
+ 502
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Template for MT-R1-Zero",
+ "text_level": 1,
+ "bbox": [
+ 536,
+ 514,
+ 744,
+ 530
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "A conversation between User and Assistant. The User asks for a translation from {src_language} to {tgt_language}, and the Assistant solves it. The Assistant first thinks about the reasoning process in the mind and then provides the user with the final translation. The reasoning process and final translation are enclosed within