CogSearch: A Cognitive-Aligned Multi-Agent Framework for Proactive Decision Support in E-Commerce Search

CogSearch: A Cognitive-Aligned Multi-Agent Framework for Proactive Decision Support in E-Commerce Search Zhouwei Zhai ∗ zhaizhouwei1@jd.com JD.com Beijing, China Mengxiang Chen chenmengxiang9@jd.com JD.com Beijing, China Haoyun Xia xiahaoyun1@jd.com JD.com Beijing, China Jin Li lijin.257@jd.com JD.com Beijing, China Renquan Zhou zhourenquan.1@jd.com JD.com Beijing, China Min Yang yangmin.aurora@jd.com JD.com Beijing, China Abstract Modern e-commerce search engines, largely rooted in passive retrieval- and-ranking models, frequently fail to support complex decision- making, leaving users overwhelmed by cognitive friction. In this paper, we introduce CogSearch, a novel cognitive-oriented multi- agent framework that reimagines e-commerce search as a proactive decision support system. By synergizing four specialized agents, CogSearch mimics human cognitive workflows: it decomposes in- tricate user intents, fuses heterogeneous knowledge across internal and external sources, and delivers highly actionable insights. Our offline benchmarks validate CogSearch’s excellence in consulta- tive and complex search scenarios. Extensive online A/B testing on JD.com demonstrates the system’s transformative impact: it reduced decision costs by 5% and achieved a 0.41% increase in overall UCVR, with a remarkable 30% surge in conversion for decision-heavy queries. CogSearch represents a fundamental shift in information retrieval, moving beyond traditional relevance-centric paradigms toward a future of holistic, collaborative decision intelligence. CCS Concepts • Computing methodologies→Multi-agent systems;• Infor- mation systems→ Novelty in information retrieval. Keywords E-commerce Search, Multi-Agent Systems, Large Language Models, Cognitive Search ACM Reference Format: Zhouwei Zhai, Mengxiang Chen, Haoyun Xia, Jin Li, Renquan Zhou, and Min Yang. 2018. CogSearch: A Cognitive-Aligned Multi-Agent Framework for Proactive Decision Support in E-Commerce Search. In Proceedings of Make sure to enter the correct conference title from your rights confirmation email ∗ The corresponding author. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. Conference acronym ’X, Woodstock, NY © 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-X-X/2018/06 https://doi.org/X.X (Conference acronym ’X). ACM, New York, NY, USA, 5 pages. https://doi. org/X.X 1 Introduction E-commerce search serves as the core hub connecting user needs with product supply. Its fundamental objective extends beyond merely returning relevant products; it aims to assist users through- out the entire cognitive process from “need articulation” to “decision formation.” Industrial search systems have long evolved around the “retrieval-ranking” paradigm. The retrieval stage has progressed from statistical models like LSI[4], BM25[12], and TF-IDF[1] to deep semantic matching driven by approaches like RocketQA[11]. The ranking stage enhances precision through techniques such as knowledge distillation[9], DIN attention mechanisms[20], and personalized modeling[18], forming a closed-loop optimization process[14]. Recently, Large Language Models have further aug- mented system performance via query rewriting[3,10], generative retrieval[2,8,16], and relevance fine-tuning[5,15,17]. However, existing research still faces critical bottlenecks: the passive response mode hinders the system’s ability to handle complex, multi-attribute implicit reasoning needs, and reliance solely on platform-internal data limits cross-platform information integration. More crucially, systems often function as “result displays” rather than “decision facilitators,” leading to excessive cognitive load and comparison costs for users during high-complexity decision-making. The core challenge in addressing these issues lies in breaking the limitations of static indexing to achieve deep convergence of user intent within dynamic dialogues and transforming fragmented product attributes into structured decision-making bases. Inspired by Simon’s decision- making model[13], this paper proposes CogSearch, a cognitively- aligned multi-agent framework. This framework upgrades search from “keyword matching” to “active decision assistance” through the collaborative operation of four specialized agents: Planner, Ex- ecutor, Guider, and Decider. Our core contributions are: •Architectural Innovation: We propose the CogSearch multi- agent framework. The Planner decomposes complex intents, the Executor integrates multi-source information, the Guider steers demand convergence, and the Decider provides struc- tured purchasing recommendations. This achieves precise alignment between the search process and the user’s shop- ping cognitive chain. arXiv:2603.11927v1 [cs.MA] 12 Mar 2026 Conference acronym ’X, June 03–05, 2018, Woodstock, NYZhouwei et al. •Large-Scale Industrial Validation: CogSearch has been fully deployed on JD.com, one of China’s largest self-operated e-commerce platforms. Offline experiments demonstrate sig- nificant performance gains in handling complex queries. Fur- thermore, online A/B testing confirms that our framework reduces user decision costs by 5%, yields a 0.41% lift in overall User Conversion Rate (UCVR), and achieves a remarkable 30% UCVR increase for decision-heavy queries. 2 The CogSearch Framework 2.1 System Overview Inspired by Simon’s three-stage decision-making model[13], CogSearch fundamentally redefines e-commerce search as a collaborative cog- nitive decision-making process, as opposed to the traditional retrieval- ranking pipeline, as illustrated in Figure 1. Figure 1: CogSearch Framework Overview The design logic of CogSearch deeply aligns with the user’s shopping cognitive chain: •Planner Agent: Solves the problem of demand understand- ing. Based on the user’s current query, search history, and historical click context, it comprehends user intent and per- forms task planning, generating a task graph stored in the Memory System. Specific tasks include product search tasks, web search tasks, and tool/service invocation. •Executor Agent: Solves the problem of multi-source infor- mation gathering. Based on the Planning Agent’s output, it executes multi-source retrieval tasks – encompassing in- platform product retrieval (returning a list of eligible prod- ucts), web-scale information retrieval, and tool/service invo- cation (e.g., weather, logistics APIs). •Guider Agent: Solves the problem of proactive decision guid- ance. Leveraging retrieved product results combined with web-sourced information, it generates personalized dynamic filters and purchasing strategies. Furthermore, it proactively identifies latent needs to generate follow-up questions for demand convergence or stimulate new user requirements. •Decider Agent: Solves the problem of integrated decision- making. Synthesizing the user profile, historical behavioral context, interaction signals from the Navigation Agent (e.g., filter clicks), and multi-source information (product attributes + web guides/strategies), it employs LLM-based reasoning to output final product recommendations accompanied by structured decision rationale. Memory System acts as the hub for storing, updating, and main- taining all contextual information for the agents. This includes contextual memories, agent states, and the task graph. 2.2 Planner Agent: Intent-Driven Task Generation As the cognitive orchestration engine of CogSearch, the Planner Agent transcends traditional query rewriting by formulating the search process as a structured reasoning problem. Its primary ob- jective is to bridge the semantic gap between ambiguous user re- quirements and precise system execution capabilities through the generation of an executable Task Graph. Formally, we define the planner’s input as a dynamic context tuple: C 푡 =푞 푡 ,H 푠푒푎푟푐ℎ <푡 ,H 푐푙푖푐푘 <푡 (1) where푞 푡 denotes the current user query, whileH 푠푒푎푟푐ℎ <푡 andH 푐푙푖푐푘 <푡 represent the historical query sequence and item interaction con- text, respectively. The Planner leverages the reasoning capacity of Large Language Models to mapC 푡 into a Directed Acyclic Graph (DAG), denoted as:G=(V,E). The vertex setV=푣 1 ,푣 2 , ...,푣 푛 represents atomic cognitive tasks derived from the user’s latent intent. To facilitate comprehensive decision support, we define the action space ofVacross three functional dimensions: (1) Prod- uctSearch (푇 푝푟표푑 ) for retrieving candidate items from the internal E-commerce inventory; (2) WebSearch (푇 푤푒푏 ) for acquiring open- world knowledge (e.g., reviews, usage guides) to support cross- platform information fusion; and (3) ToolInvocation (푇 푡표푙 ) for exe- cuting specific functional operations via the Model Context Protocol (MCP), such as price comparison or logistics querying. The edge setEexplicitly models the semantic dependencies between these tasks. A directed edge(푣 푖 ,푣 푗 ) ∈ Eis generated if and only if the execution of task푣 푗 is conditional on the output context of푣 푖 , sat- isfying the constraint퐼푛푝푢푡(푣 푗 ) ⊆ 푂푢푡푝푢푡(푣 푖 ). The generation of Gis modeled as a conditional probability distribution푃 휃 (G|C 푡 ), where the Planner decomposes complex queries into a topological sequence of sub-tasks. 2.3 Executor Agent: Multi-Source Acquisition 2.3.1ProductSearch Agent. This agent transforms structured queries into ranked product candidates via a cognitively augmented hy- brid retrieval architecture. To balance precision and diversity: 1) Keyword-Vector Fusion: Concurrently invokes platform APIs (lexi- cal recall) and contrastive learning-based semantic embedding mod- els (intent generalization). 2) Multi-Source Enrichment: Coordinates ProductAttr and ReviewSum sub-agents to aggregate attributes and generate compressed pros/cons summaries from descriptions and reviews. 2.3.2WebSearch Agent. Overcoming platform-bound information gaps, this agent integrates real-time cross-domain data via a noise- robust “Expand-Retrieve-Filter” pipeline: 1)Query Expansion: LLM generates synonymic variants of web search sub-tasks.2) Parallel Retrieval: Fetches raw results from search APIs. 3)Multi-Criteria Filtering: Scores results using: 푆푐표푟푒(푑)= 훼 · 푅푒푙(푑,푞)+ 훽· 퐴푢푡ℎ(푑)+훾 · 퐹푟푒푠ℎ(푑) where 훼,훽,훾 calibrate relevance, authority, and freshness. CogSearch: A Cognitive-Aligned Multi-Agent Framework for Proactive Decision Support in E-Commerce SearchConference acronym ’X, June 03–05, 2018, Woodstock, NY 2.3.3 Tool Integration Agent. Dynamically invokes external APIs (e.g., weather impact on logistics, price trend forecasting) using LLM-generated function calls when contextual triggers are detected in the cognitive loop. 2.4 Guider Agent: Cognitive-Aligned Guidance 2.4.1 Dynamic Cognitive Facet Generation. Traditional facets are often pre-defined and rigid. Drawing inspiration from cognitive psy- chology [23], the Guider Agent employs a Clarify-Guide-Contextualize (CGC) mechanism to generate dynamic, hierarchical filters. By ana- lyzing the latent semantic variance between the retrieved item set Iand the user’s cognitive stateS 푢 , the agent formulates a scoring function to select the most discriminative attributes 푎 ∗ : 푎 ∗ = arg max 푎∈A InfoGain(푎|I,S 푢 ,C) whereCrepresents the multi-turn interaction context. This "brain- inspired" hierarchical reasoning [24] ensures that the generated facets (e.g., "Battery Life for Vlogging" vs. "Weight for Hiking") align with the user’s real-time decision priorities, significantly reducing the cognitive load of information filtering. 2.4.2 Personalized Purchase Strategy. To address the "paradox of choice," the Guider Agent synthesizes multi-source intelligence—including item specifications from the Executor, real-time web knowledge (e.g., expert reviews), and the user’s historical preferences—to gen- erate a Personalized Purchase Strategy. The strategy is formulated as a structured knowledge prompt: •Expert Anchoring: Identifies key performance indicators (KPIs) for the specific category. •Trade-off Analysis: Explicitly highlights pros and cons based on the user’s "add-to-cart" history and budget constraints. By transforming raw data into actionable "Buying Guides," the agent shifts the system from a passive list provider to a professional shopping consultant. 2.4.3 Query Suggestions. The Guider Agent proactively manages the search trajectory by generating dual-purpose suggestions: (1)Convergent Queries: Aimed at "zooming in" on specific at- tributes when the user shows hesitation (e.g., "Noise-cancelling headphones under 200g"). (2)Stimulative Queries: Designed to "expand" the search space based on latent needs identified via web-enhanced cross- platform signals (e.g., "Compatible stabilizers for your new mirrorless camera"). This dual-steering approach ensures that the search process is not merely a reactive loop but a guided journey toward an optimal decision. 2.5 Decider Agent: Integrative Reasoning and Decision-Making The Decider Agent serves as the terminal cognitive nexus of the CogSearch framework. Unlike traditional ranking modules that output probabilistic scores, this agent acts as a reasoning engine designed to reach "cognitive closure." Grounded in Simon’s three- stage decision model (Intelligence-Design-Choice)[13] and Multi- Criteria Decision Making theory[7], the Decider Agent synthesizes heterogeneous signals into actionable, interpretative purchasing advice. Context Aggregation. The agent first constructs a holistic deci- sion contextS 푐푡푥 by aggregating outputs from upstream agents. LetPdenote the candidate items and summarized reviews from the ProductSearch Agent,Wthe real-time external evidence from the WebSearch Agent, andHthe interaction trajectory from the Guider Agent. Combined with the user’s static profileUand dy- namic constraintsC(budget, preferences) defined by the Planner Agent, the unified context is formalized as: S 푐푡푥 =Φ(P,W,H,U,C) whereΦrepresents the context fusion mechanism. This integra- tion ensures the decision is not isolated but deeply rooted in both internal platform data and the open web ecosystem. Multi-Criteria Decision Execution. We leverage the reasoning capabilities of large language models to perform MCDM. To miti- gate the hallucination risks common in direct generation, we inject a structured evaluation protocolI 푒푣푎푙 into the prompt. The agent evaluates candidates based on a composite utility vectorv spanning four dimensions: (1) Functional Performance: Assessing core specifications and quality sentiment. (2) Economic Viability: Analyzing price-to-budget alignment. (3)Reliability: Cross-verifying internal reviews with external web credibility signals. (4)Constraint Adherence: Checking compliance with Planner- derived hard constraints. Decision & Rationale Generation. The final output is not merely a ranked list but a persuasive recommendation pair(푑 ∗ ,R), where 푑 ∗ is the optimal item andRis the natural language rationale explaining why the item fits the user’s specific scenario: (푑 ∗ ,R) ← LLM(S 푐푡푥 ,I 푒푣푎푙 ) By explicitly modeling the trade-offs between performance, cost, and reliability, the Decider Agent bridges the gap between system retrieval and user decision-making, providing a transparent path from information seeking to transaction. 3 Experiments 3.1 Experimental Setup 3.1.1Dataset. To bridge the gap in standardized evaluation for cog- nitive decision-making in e-commerce search, we propose ECCD- Bench. This benchmark is curated from anonymized JD.com query logs and features 10k expert-annotated query-response pairs. It spans three task categories designed to assess diverse cognitive abilities: (1) Simple Needs, involving explicit product or category searches; (2) Complex Needs, requiring advanced reasoning, han- dling of multiple constraints, negation, or semantic ambiguity; and (3) Consultative Queries, focusing on advisory interactions such as product Q&A, comparative analysis, and general information seeking with latent shopping intent. 3.1.2Evaluation Metrics. The proposed model is evaluated through both offline benchmarks and online A/B testing. Offline Metrics: (1) Accuracy@K (ACC@K): Measures the capa- bility to retrieve relevant products aligning with user intent within Conference acronym ’X, June 03–05, 2018, Woodstock, NYZhouwei et al. the top K positions. We report ACC@5 to mirror the real-world dis- play. (2) User Demand Satisfaction (UDS): A human-centric metric (scale 1-3) where experts rate system responses across completeness, correctness, and usefulness. Online Metrics: (1) Decision Cost (DC): The average number of interactions (searches plus clicks) per transaction. A lower DC signifies higher decision-making efficiency.(2) User Conversion Rate (UCVR): UCVR is defined as the ratio of the total number of completed orders to the total number of Unique Visitors (UV) within a specific timeframe. 3.2 Offline Experiments We conducted offline evaluations using the current JD.com search system (an LLM-enhanced retrieval-ranking model) as the baseline. To validate CogSearch’s core agents and modules, ablation experi- ments were performed on the ECCD-Bench dataset (Table 1). Re- sults show that CogSearch significantly outperforms the online sys- tem in ACC@5 and UDS for complex and consultative queries. This suggests that CogSearch effectively decodes implicit user intent and multifaceted constraints. By mimicking professional shopping assistants through dynamic information synthesis and reasoning, CogSearch excels in complex scenarios, bridging the semantic gap that limits traditional paradigms. Table 1: ECCD-Bench Offline Evaluation Method Simple Needs Complex Needs Consultative Queries (ACC@5 / UDS)(ACC@5 / UDS)(ACC@5 / UDS) online system(base)0.87 / 2.640.48 / 1.920.60 / 1.81 CogSearch (Full)0.87 / 2.700.83 / 2.790.65 / 2.88 w/o Planner0.86 / 2.650.52 / 2.050.61 / 2.42 w/o WebSearch0.87 / 2.680.81 / 2.720.58 / 2.15 w/o Guider0.87 / 2.640.80 / 2.230.65 / 2.10 w/o Decider0.85 / 2.620.75 / 2.380.63 / 2.40 w/o Memory0.86 / 2.610.68 / 2.350.62 / 2.45 Ablation study findings: • w/o Planner: ACC@5 and UDS for complex queries plum- met by 37.3% and 26.5%, respectively, while simple queries remain stable. This confirms the Planner’s necessity in bridg- ing semantic gaps via intent decomposition. •w/o WebSearch: UDS for consultative queries drops by 25.3%, highlighting the importance of web-scale information fusion for advisory decision-making. • w/o Guider: Complex and consultative UDS fall by 20% and 27%. Replacing the Guider with static filters increases user cognitive load, demonstrating the value of dynamic, personalized guidance. •w/o Decider: A universal decline in UDS underscores that multi-criteria reasoning and rationales are essential for a robust decision loop, transcending simple list re-ranking. •w/o Memory: All metrics decrease, validating the memory system’s role in ensuring contextual coherence and persis- tent intent alignment. 3.3 Deployment and Online A/B Testing We integrated CogSearch into the JD.com search platform via an opt- in "AI Search" entry. To satisfy strict online latency requirements, the Planner and Guider modules utilize Qwen3-4B[19], while the Decider employs Qwen3-8B[19]. These models were fine-tuned via knowledge distillation from DeepSeek-R1[6]. We optimized online inference using INT8 quantization, caching mechanisms, and parallel scheduling. Although the average end-to-end latency is 1.3s (compared to 800ms for traditional search), we implemented streaming output to maintain a responsive user experience. We conducted a two-week online A/B test with 10% of live traffic. The experimental group was granted access to the CogSearch in- terface, while the control group was restricted to traditional search. Results show a significant 5% reduction in Decision Cost (DC) and an average 0.41% increase in User Conversion Rate (UCVR) for the experimental group. Notably, for decision-heavy intents—such as professional product interpretation, gift selection, and functional comparisons—UCVR improved by over 30%. Behavioral analysis reveals that users leverage CogSearch for more complex, personalized needs; the proportion of long queries (>10 words) increased by 25% relative to the control, reflecting a shift toward natural language expression. These results demonstrate that CogSearch effectively reduces the interaction cost of repeated searches and clicks by providing precise, decision-oriented support. CogSearch has since been fully deployed on the JD platform. 4 Conclusion we presented CogSearch, a system that successfully transforms e- commerce search into a proactive and cognitively aligned decision- making environment. Through our multi-agent framework—which integrates intent decomposition, cross-source information fusion, and interactive steering—we effectively minimize cognitive friction and simplify the consumer journey. Our large-scale deployment on JD.com yields irrefutable empirical results: a 5% decrease in user decision costs and a statistically significant lift in average UCVR by 0.41%, soaring to a 30% increase for complex, decision-heavy queries. Such performance underscores both the technical efficacy and the commercial potential of our approach. Future work will prioritize evolving CogSearch into a lifelong, multimodal shopping companion by integrating sensory perception and long-term cognitive memory, while simultaneously optimiz- ing the computational efficiency of multi-agent reasoning without compromising depth. Presenter Biography Zhouwei Zhai is a Scientist at JD.com, focusing on LLM-powered search systems and AI Agents. At JD.com, he spearheaded the transition towards LLM-augmented e-commerce search and led the end-to-end construction of the platform’s next-generation AI Search Assistant. References [1]Akiko Aizawa. 2003. An information-theoretic perspective of tf–idf measures. Information Processing & Management 39, 1 (2003), 45–65. [2]Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen-tau Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: Generating substrings as document identifiers. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 35. 31668–31683. [3]Aijun Dai, Zhenyu Zhu, Haiqing Hu, Guoyu Tang, Lin Liu, and Sulong Xu. 2024. Enhancing E-Commerce Query Rewriting: A Large Language Model Approach with Domain-Specific Pre-Training and Reinforcement Learning. In Proceedings of CogSearch: A Cognitive-Aligned Multi-Agent Framework for Proactive Decision Support in E-Commerce SearchConference acronym ’X, June 03–05, 2018, Woodstock, NY the 33rd ACM International Conference on Information and Knowledge Management. 4439–4445. [4]Scott Deerwester, Susan T Dumais, George W Furnas, Thomas K Landauer, and Richard Harshman. 1990. Indexing by latent semantic analysis. Journal of the American society for information science 41, 6 (1990), 391–407. [5]Chenhe Dong, Shaowei Yao, Pengkun Jiao, Jianhui Yang, Yiming Jin, Zerui Huang, Xiaojiang Zhou, Dan Ou, Haihong Tang, and Bo Zheng. 2025. TaoSR1: The thinking model for e-commerce relevance search. arXiv preprint arXiv:2508.12365 (2025). [6]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruiyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al.2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025). [7]Ralph L Keeney and Howard Raiffa. 1993. Decisions with multiple objectives: preferences and value trade-offs. Cambridge university press. [8]Mingming Li, Huimu Wang, Zuxu Chen, Guangtao Nie, Yiming Qiu, Guoyu Tang, Lin Liu, and Jingwei Zhuo. 2024. Generative retrieval with preference optimization for e-commerce search. arXiv preprint arXiv:2407.19829 (2024). [9]Ziyang Liu, Chaokun Wang, Hao Feng, Lingfei Wu, and Liqun Yang. 2022. Knowl- edge distillation based contextual relevance matching for e-commerce product search. arXiv preprint arXiv:2210.01701 (2022). [10] Wenjun Peng, Guiyang Li, Yue Jiang, Zilong Wang, Dan Ou, Xiaoyi Zeng, Derong Xu, Tong Xu, and Enhong Chen. 2024. Large language model based long-tail query rewriting in taobao search. In Companion Proceedings of the ACM Web Conference 2024. 20–28. [11] Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxi- ang Dong, Hua Wu, and Haifeng Wang. 2020. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2010.08191 (2020). [12]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. [13] Herbert A Simon. 1955. A behavioral model of rational choice. The quarterly journal of economics (1955), 99–118. [14]Shivaramakrishnan Kalpetta Subramaniam. 2025. AI-Based E-commerce Search Optimization. Journal Of Engineering And Computer Sciences 4, 6 (2025), 307–316. [15]Tian Tang, Zhixing Tian, Zhenyu Zhu, Chenyang Wang, Haiqing Hu, Guoyu Tang, Lin Liu, and Sulong Xu. 2025. LREF: A Novel LLM-based Relevance Framework for E-commerce Search. In Companion Proceedings of the ACM on Web Conference 2025. 468–475. [16]Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Donald Metzler, and William W Cohen. 2022. Transformer memory as a differentiable search index. NeurIPS 35 (2022), 21831–21843. [17]Paul Thomas, Seth Spielman, Nick Craswell, and Bhaskar Mitra. 2024. Large language models can accurately predict searcher preferences. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1930–1940. [18]Su Yan, Wei Lin, Tianshu Wu, Daorui Xiao, Xu Zheng, Bo Wu, and Kaipeng Liu. 2018. Beyond keywords and relevance: a personalized ad retrieval framework in e-commerce sponsored search. In Proceedings of the 2018 World Wide Web Conference. 1919–1928. [19]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al.2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025). [20]Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1059–1068.