Paper deep dive
GenRec: A Preference-Oriented Generative Framework for Large-Scale Recommendation
Yanyan Zou, Junbo Qi, Lunsong Huang, Yu Li, Kewei Xu, Jiabao Gao, Binglei Zhao, Xuanhua Yang, Sulong Xu, Shengjie Li
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/18/2026, 1:40:02 AM
Summary
GenRec is a preference-oriented generative framework for large-scale recommendation that addresses challenges in generative retrieval (GR) by introducing Page-wise Next Token Prediction (PW-NTP) for training, an asymmetric linear Token Merger for efficient long-sequence modeling, and GRPO-SR (Group Relative Policy Optimization with Supervised Regularization) for preference alignment. Deployed on the JD App, it significantly improves click and transaction metrics.
Entities (6)
Relation Signals (4)
GenRec → deployedon → JD.com
confidence 100% · GenRec, a preference-oriented generative framework deployed on the JD App
GenRec → utilizes → PW-NTP
confidence 95% · For training objective, we propose Page-wise NTP task
GenRec → utilizes → GRPO-SR
confidence 95% · we introduce GRPO-SR, a reinforcement learning method
GenRec → utilizes → Token Merger
confidence 95% · an asymmetric linear Token Merger compresses multi-token Semantic IDs
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Generative Retrieval (GR) offers a promising paradigm for recommendation through next-token prediction (NTP). However, scaling it to large-scale industrial systems introduces three challenges: (i) within a single request, the identical model inputs may produce inconsistent outputs due to the pagination request mechanism; (ii) the prohibitive cost of encoding long user behavior sequences with multi-token item representations based on semantic IDs, and (iii) aligning the generative policy with nuanced user preference signals. We present GenRec, a preference-oriented generative framework deployed on the JD App that addresses above challenges within a single decoder-only architecture. For training objective, we propose Page-wise NTP task, which supervises over an entire interaction page rather than each interacted item individually, providing denser gradient signal and resolving the one-to-many ambiguity of point-wise training. On the prefilling side, an asymmetric linear Token Merger compresses multi-token Semantic IDs in the prompt while preserving full-resolution decoding, reducing input length by ~2X with negligible accuracy loss. To further align outputs with user satisfaction, we introduce GRPO-SR, a reinforcement learning method that pairs Group Relative Policy Optimization with NLL regularization for training stability, and employs Hybrid Rewards combining a dense reward model with a relevance gate to mitigate reward hacking. In month-long online A/B tests serving production traffic, GenRec achieves 9.5% improvement in click count and 8.7% in transaction count over the existing pipeline.
Tags
Links
- Source: https://arxiv.org/abs/2604.14878v1
- Canonical: https://arxiv.org/abs/2604.14878v1
Trouble viewing inline? Open PDF directly →
Full Text
30,927 characters extracted from source content.
Expand or collapse full text
GenRec: A Preference-Oriented Generative Framework for Large-Scale Recommendation Yanyan Zou ∗† JD.com Beijing, China zoe.yyzou@gmail.com Junbo Qi ∗ Waseda University Tokyo, Japan junboqi@toki.waseda.jp Lunsong Huang ∗ JD.com Beijing, China huanglunsong1@jd.com Yu Li ∗ JD.com Beijing, China liyu.liz@jd.com Kewei Xu ∗ JD.com Beijing, China xukewei3@jd.com Jiahao Gao ∗ JD.com Beijing, China gaojiahao.20@jd.com Binglei Zhao JD.com Beijing, China zhaobinglei1@jd.com Xuanhua Yang JD.com Beijing, China yangxuanhua1@jd.com Sulong Xu JD.com Beijing, China xusulong@jd.com Shengjie Li ‡ JD.com Beijing, China lishengjie1@jd.com Abstract Generative Retrieval (GR) offers a promising paradigm for recom- mendation through next-token prediction (NTP). However, scaling it to large-scale industrial systems introduces three challenges: (i) within a single request, the identical model inputs may produce inconsistent outputs due to the pagination request mechanism; (i) the prohibitive cost of encoding long user behavior sequences with multi-token item representations based on semantic IDs, and (i) aligning the generative policy with nuanced user preference signals. We present GenRec, a preference-oriented generative framework deployed on the JD App 1 that addresses above challenges within a single decoder-only architecture. For training objective, we propose Page-wise NTP task, which supervises over an entire interaction page rather than each interacted item individually, providing denser gradient signal and resolving the one-to-many ambiguity of point- wise training. On the prefilling side, an asymmetric linear Token Merger compresses multi-token Semantic IDs in the prompt while preserving full-resolution decoding, reducing input length by∼2× with negligible accuracy loss. To further align outputs with user satisfaction, we introduce GRPO-SR, a reinforcement learning method that pairs Group Relative Policy Optimization with NLL ∗ Core contributors. † Project leader. ‡ Corresponding author. 1 https://w.jd.com This work is licensed under a Creative Commons Attribution 4.0 International License. SIGIR ’26, Melbourne, VIC, Australia © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2599-9/2026/07 https://doi.org/10.1145/3805712.3808437 regularization for training stability, and employs Hybrid Rewards combining a dense reward model with a relevance gate to mitigate reward hacking. In month-long online A/B tests serving production traffic, GenRec achieves 9.5% improvement in click count and 8.7% in transaction count over the existing pipeline. CCS Concepts • Information systems→ Recommender systems. Keywords Generative Retrieval; Large-scale Recommender System; Super- vised Fine-tuning; Preference Alignment ACM Reference Format: Yanyan Zou, Junbo Qi, Lunsong Huang, Yu Li, Kewei Xu, Jiahao Gao, Bin- glei Zhao, Xuanhua Yang, Sulong Xu, and Shengjie Li. 2026. GenRec: A Preference-Oriented Generative Framework for Large-Scale Recommen- dation. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26), July 20– 24, 2026, Melbourne, VIC, Australia. ACM, New York, NY, USA, 6 pages. https://doi.org/10.1145/3805712.3808437 1 Introduction Modern recommender systems typically adopt a retrieve-and-rank architecture [11,13]. Recent progresses [10,16,21] have shown the effectiveness of generative retrieval paradigm. By reformulat- ing retrieval task as a conditional sequence generation problem, this approach directly generates target items from the entire cor- pus. Nevertheless, we observe that deploying such a method in large-scale industrial recommender systems remains challenging. First, to handle high-volume traffic and ensure user experience, arXiv:2604.14878v1 [cs.IR] 16 Apr 2026 SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYanyan Zou et al. industrial systems typically employ the pagination request mecha- nism. Within each paginated request, a user might exhibit multiple positive interactions (e.g., click, transaction), leading to identical model inputs yet multiple valid output in generative task following the vanilla NTP paradigm. Second, the long user historical behav- ior sequences can lead to substantial computational overhead and increased online inference latency. Third, naively aligning the gen- erative model with personalized objectives could result in reward hacking and performance degradation. To address above challenges, we propose GenRec, a generative retrieval-based recommendation framework, unifying user intent understanding and item retrieval within a single decoder-only architecture. It integrates SID-based representations and preference alignment via Group Relative Policy Optimization with Supervised Regularization (GRPO-SR), enabling robust optimization under large-scale real-world users’ feedback. The main contributions are summarized as follows: • We introduce a Page-Wise NTP supervised fine-tuning (SFT) strategy to capture holistic user interaction patterns and resolve the one-to-many ambiguity of vanilla NTP. •We propose an asymmetric representation architecture utiliz- ing a linear Token Merger in prefilling side. This mechanism compresses the prompt embedding sequence to efficiently model long user behaviors, while decoding side are unmerged SIDs to ensure fine-grained item retrieval. •We develop a reinforcement learning (RL) method GRPO- SR with Hybrid Rewards for preference alignment, which combines a dense reward model with a gating mechanism that suppresses Reward Hacking. We further introduce Negative Log-Likelihood (NLL) regularization to stabilize training and preserve real-world user behavior patterns. •We empirically demonstrate scaling laws in generative rec- ommendation and validate the framework through large- scale deployment in the JD App, achieving 9.5% click and 8.7% transaction improvements. 2 Methodology 2.1 Preliminary We formulate the retrieval task as a unified conditional sequence generation problem. Given theH=푣 1 , . . .,푣 푛 , which denotes the user’s historical behavior sequence arranged in chronological order (e.g.푣 푛 represents the most recent interaction), the model predicts the subsequent itemY, representing a potential user interest. To be specific, the method operates over discrete Semantic Iden- tifiers (SIDs) [16]. A multi-modal model (i.e., Qwen2.5-VL [1]) is employed to jointly encode both visual appearance and textual descriptions of each item into a continuous representation to cap- ture more comprehensive item information. Following existing practices [27], we fine-tune the embedding model using domain- specific collaborative pairs to ensure the learned embeddings cap- ture recommendation-oriented semantics. Then, RQ K-means is utilized to discretize the refined embeddings. This process iter- atively clusters the residual vectors, mapping each item푣 푖 to a hierarchical tuple of cluster indices: SID(푣 푖 )=푠 1 푖 ,푠 2 푖 ,푠 3 푖 .(1) 2.2 User-Centric Page-Wise NTP SFT Previous generative recommendation methods [16,27] train and infer under the same point-wise protocol: predicting a single next item given a user history. Due to the pagination request mechanism of large-scale industrial recommender systems, we argue that this creates a fundamental label ambiguity: when the same historyH is paired with퐾distinct positive items푣 (푘) 퐾 푘=1 , the model must maximize Í 푘 log푃 휃 (푣 (푘) | H)over an identical prefix, effectively fitting a uniform mixture over all valid continuations. This flattened distribution both inflates gradient variance and dilutes per-item probability mass, degrading top-퐾precision. The root causes a cardinality mismatch: a single session naturally yields multiple engagement signals, yet vanilla point-wise NTP collapses them into isolated input–label pairs, discarding intra-session structure. To resolve this, we decouple the training and inference formu- lations. Following Eq 1, we formulate the input as a composite prompt: 푆 푢 =[SID(푣) : 푣 ∈H] ≻ ,(2) Page-wise Supervision. We further design a page-wise next token prediction (i.e., PW-NTP) strategy for SFT. The target sequence is a page-wise list of items the user interacted within the current page, like ordered itemsO, clicked itemsCand exposed itemsE, ordered by interaction intensity: 푌 page =[SID(푣) : 푣 ∈ O∪C∪E] ≻ (3) The training objective is the standard autoregressive SFT loss over the full response sequence 푌 page : L SFT =− |푌 page | ∑︁ 푡=1 log푃 휃 (푦 푡 | 푆 푢 ,푦 <푡 ).(4) By supervising over the entire page rather than a single item, each forward pass provides a denser learning signal and resolves the one-to-many ambiguity inherent in point-wise training. Point-wise Beam Search. At serving time, the model generates beam-width items per query via beam search, following the standard point-wise protocol. This asymmetry is by design: list-wise training provides richer supervision of model gradients, while point-wise inference maintains compatibility with the production beam search pipeline for online serving requirements. Decoder-only Architecture with Token Merger. To better leverage and reuse the inference optimization techniques developed in the Large Language Model community, we directly adopt a decoder- only transformer architecture, as depicted in Figure 1. However, multi-token SIDs triple the input sequence length of the item part, posing severe latency challenges. We address this via a linear Token Merger. Since the SID triplet푠 1 푖 ,푠 2 푖 ,푠 3 푖 is derived from a single item푣 푖 , we concatenate and project their embeddings into a unified vectorh 푣 푖 via a linear layer in the prompt part: h 푣 푖 = Linear(Concat(e(푠 1 푖 ),e(푠 2 푖 ),e(푠 3 푖 ))).(5) This design compresses item SIDs into latent vectors, reducing prompt length by∼2×to accommodate long user sequences within strict inference budgets. Special tokens (e.g.,<sep>) are kept un- merged to serve as explicit indicators of structural separation. Cru- cially, this optimization is confined to the prefilling phase, while the GenRec: A Preference-Oriented Generative Framework for Large-Scale RecommendationSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia Decoder Layers Merger ... ... LM Head Input Output No Compressed Items for Training&Decoding ... ... Figure 1: Model architecture of GenRec. High-dimensional items are quantized into Semantic IDs. To enhance efficiency, an Linear token Merger projects the concatenated embeddings of an item’s SIDs into a unified latent vector on the prefilling side. Other tokens (e.g. <sep>) remain uncompressed. decoding process and generative objective adhere to the original semantic token sequence. 2.3 Preference Alignment via RL While page-wise SFT captures behavioral regularities from histor- ical logs, it lacks explicit optimization for user satisfaction and is inherently brittle against sparse and non-stationary real-world feedback. To address these limitations, we introduce a RL method GRPO-SR, building on Group Relative Policy Optimization (GRPO) [17]. Unlike PW-NTP SFT stage, the RL stage aligns with the point- wise beam search inference protocol: each rollout generates a single item sequence per query, ensuring consistency between RL train- ing and online serving. Our method optimizes relative preferences among multiple generated candidates rather than absolute reward values, thereby improving robustness in industrial settings. Reward Formulation. Raw engagement signals (e.g., clicks) are too sparse to provide effective policy gradients. We instead employ a SIM-based model [14] to estimate a continuous preference score 푟 pref 푖 ∈ [0,1]for each rollout candidate표 푖 . A key failure mode is reward hacking: the policy produces syntactically valid SID combi- nations that receive non-trivial푟 pref yet are semantically irrelevant. We suppress this with a gate mechanismG 푖 = I(푠 푖 > 휏)where휏is a small constant, yielding the hybrid reward: 푟 푖 =G 푖 · 푟 pref 푖 .(6) The dense preference model, while smoother than binary labels, may still under-estimate rewards for items the user actually engaged with. LetD + =O∪Cdenote the set of ordered and clicked items from the interaction page. We calibrate the reward within each rollout group by anchoring positive items to the group maximum: ̃ 푟 푖 = 1− I(표 푖 ∈ D + ) · 푟 푖 + I(표 푖 ∈ D + )· 푟 max ,(7) where푟 푚푎푥 is the highest푟 푖 in the group. This guarantees that the predicted items that hit real-world user’ positive behaviors always receive top-tier rewards, preventing the reward model’s estimation bias from down-weighting genuinely preferred items. GRPO-SR Objective. We propose a composite objective that har- monizes group-relative policy optimization with supervised stabil- ity. The loss function is defined as: L GRPO−SR (휃)= − E 푆 푢 ∼푇,표 푖 퐺 푖=1 ∼휋 휃 (·|푆 푢 ) " 1 퐺 퐺 ∑︁ 푖=1 1 |표 푖 | |표 푖 | ∑︁ 푡=1 휋 휃 (표 푖,푡 | 푆 푢 ,표 푖,<푡 ) sg 휋 휃 (표 푖,푡 | 푆 푢 ,표 푖,<푡 ) ˆ 퐴 푖,푡 # − 훼 · E 푣∼D + " |푣| ∑︁ 푡=1 log휋 휃 (푣 푡 | 푆 푢 ,푣 <푡 ) # (8) where푇is training set, and ˆ 퐴 푖,푡 is the advantage derived from group- relative rewards. The first term leverages a importance sampling 휋 휃 /sg(휋 휃 )to enable stable, one-step policy updates. The second term, weighted by훼, imposes a negative log-likelihood constraint over positive trajectoriesD + . Distinct from standard KL-divergence penalties, this NLL regularizer explicitly anchors the policy to the real-world users’ behavior, mitigating over-optimization towards. SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYanyan Zou et al. ModelHR@1HR@10N@10HR@50N@50HaR↓ Traditional Methods BERT4Rec0.03150.09680.04120.18320.0689- SASRec0.03830.10480.04920.19760.0776- Generative Methods TIGER0.05180.16600.08030.35560.140915.46% LC-Rec0.09470.36690.21460.62260.27177.80% Ours GenRec0.11890.44560.26350.71920.32474.96% w/o TM 0.1193 0.4467 0.2653 0.7201 0.3276 4.89% Table 1: Next-item vs. next-sequence prediction performance. “TM” denotes Token Merger. Best results are in bold, and the second best are underlined. 3 Evaluation 3.1 Experimental Setup The training and testing datasets are collected from a large-scale recommender platform from the JD.com, covering around 560 mil- lion user interaction sequences over a one-month period. We take the data from the last day for testing and the remaining for train- ing. For SFT, we consider three metrics for evaluation: HitRate (HR@K) [4], NDCG (N@K) [7] and Hallucination Rate (HaR, the percentage of invalid SIDs of the generated results.). For the RL ex- periments study, we utilize the highest푟 SIM score of the generated K items as Reward Metrics (R@K). We consider both traditional and generative methods as our baselines, including BERT4Rec [18], SASRec [8], TIGER [16], and LC-Rec [26]. It is worthy noting that both TIGER and LC-Rec are trained via vanilla point-wise NTP task. The Qwen2.5 [15] decoder-only architecture is adopted as our back- bone. To be specific, our model is further trained on the Qwen2.5 variants of 1.5B, 3B and 7B. To train the generative models, we conducted distributed training across 8 NVIDIA H100 GPUs. The AdamW [9] is employed as the optimizer with a linear warm-up phase over the first 1% of training steps, followed by cosine learning rate decay. 3.2 Evaluation Results 3.2.1Effectiveness of the Model Structure and SFT Framework. We take the Qwen2.5 3B as our main backbone. To make fair compar- ison, we reproduce the LC-Rec using the same Qwen2.5 variant. Recall that LC-Rec is trained following vanilla next-token prediction task. Table 1 illustrates the offline results of various methods on the large-scale industrial dataset. Our method (denoted as “GenRec") achieve better performance on HR and N yet lower HaR compared to both traditional and generative methods. When comparing the variant of our method with the full token as the input of the decoder module (i.e., removing the token merger module, denoted as “w/o TM"), both the performance and the valid generation rate remain comparable to the one with token merger. This demonstrates the effectiveness of the simple token merger where the decisive infor- mation of input are preserved while the input token length of the decoder module is reduced by half. Furthermore, to investigate the effectiveness of our proposed PW- NTP task, we compare LC-Rec (trained with vanilla NTP), which is Model SizeHR@1HR@10N@10HR@50N@50HaR.↓ 1.5B0.10770.41030.24840.65270.18855.34% 3B0.11890.44560.26350.71920.32474.96% 7B0.1221 0.4483 0.2649 0.7216 0.32695.42% Table 2: Model scaling performance. Best results are in bold, and the second best are underlined. equal to the variant of GenRec without token merger module as well as trained via vanilla NTP. As illustrated in Table 1 and Figure 2a, GenRec consistently outperforms LC-Rec across all metrics, as well as achieves a better converged loss, demonstrating the effectiveness of the proposed PW-NTP supervision. We attribute such significant improvement to two factors: (1) vanilla NTP creates a one-to-many ambiguity where identical input contexts correspond to multiple valid labels, increasing optimiza- tion difficulty and gradient variance; (2) PW-NTP aggregates super- vision signals across sequential targets, providing denser learning signal per forward pass and accelerating convergence. Notably, PW- NTP also reduces the hallucination rate by over 50%, suggesting that joint prediction encourages more coherent item generation. 3.2.2 Scaling with Model Size. We investigate model capacity by training Qwen2.5 variants (1.5B, 3B, and 7B) on identical data. As shown in Figure 2b and Table 2, training loss consistently decreases with scale. However, the performance gain from 3B to 7B is mar- ginal compared to the 1.5B-to-3B leap, despite a∼2.3×parameter increase. Architectural analysis reveals a distinct structural differ- ence: the 3B model is deeper but narrower (36 layers, 2048 hidden size) compared to the 7B variant (28 layers, 3584 hidden size) [15]. This suggests that for generative recommendation, the increased depth in the 3B model may allow for more effective modeling of complex user-item interactions through additional non-linear trans- formations, partially compensating for the lower representational capacity (width). This aligns with the “capacity density” hypoth- esis [20], indicating that optimizing depth over width may yield better efficiency for this domain. 3.2.3RL Alignment. We study how RL helps align model outputs with user preferences. By systematically varying the number of generated rollouts and recording the mean reward among candi- dates, we use Reward Metric of푟 SIM as a metric to evaluate the quality of preference alignment. Table 3 demonstrates that the RL-aligned model consistently surpasses the SFT baseline across all inference budgets. The largest improvement appears at Reward@1 (+18.01% relative gain) with full GRPO-SR, indicating that RL effectively reshapes the output distribution toward high-reward candidates. However, removing Gcauses substantial drops in HR@50 and HaR despite marginal reward gains, which is a clear sign of reward hacking where the model exploits SIM’s requirement for valid SIDs that represent the real items while sacrificing overall quality. 3.3 Online A/B Testing We deploy both base SFT and RL-Aligned models on JD’s home- page feed recommendation platform with 10% traffic each over one GenRec: A Preference-Oriented Generative Framework for Large-Scale RecommendationSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia 05000100001500020000250003000035000 Steps 1.25 1.50 1.75 2.00 2.25 2.50 2.75 Loss NTP PW-NTP (a) Task formulation 05000100001500020000250003000035000 Steps 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 Loss 1.5B 3B 7B (b) Model scaling Figure 2: SFT loss curves. (a) Page-wise NTP converges faster than NTP. (b) Larger models achieve lower loss with diminishing returns beyond 3B. ModelHR@50 Reward Metrics HaR↓ R@1 R@10 R@50 Baseline GenRec (Base SFT model)0.71920.10270.15190.17764.96% Policy Gradient Variants GRPO0.7248 0.11770.16500.18616.03% GRPO-SR0.7438 0.1212 0.1679 0.18922.68% Reward Variants GRPO w/o G0.69750.10450.16080.1797 1.75% GRPO-SR w/o G0.70160.10670.15980.18131.96% Table 3: RL performance. We analyze the impact of different method and the gating mechanism (G). Best results are in bold, and the second best are underlined. SettingExposure Rate Click Count Transaction Count GenRec (Base SFT model)48.7%+8.5%+7.3% + GRPO-SR alignment57.3%+9.5%+8.7% Table 4: Online A/B results: RL alignment resolves click- conversion misalignment. month. As illustrated in Table 4, both versions achieve significant (two-sided test with푝<0.05) improvements on click and trans- action counts. To be specific, for the long-tail items, the proposed method leads to a 10% increase in exposure rate, a 16% increase in click count, and a 13% increase in transaction count. GenRec with GRPO-SR alignment is now fully deployed in production. 4 Related Work Motivated by Large Language Models (LLMs), recommender sys- tems are shifting from traditional discriminative to generative mod- eling [3,6,23,24]. Redefining recommendation as sequence-to- sequence generation, studies optimize backbones: HSTU[22] boosts efficiency via gated linear recurrence instead of attention, MTGR[5] builds an industrial framework on it to balance scalability and pre- cision, and OneTrans [25] enhances generalization through multi- task learning and knowledge transfer. Treating items as tokens [2] causes large vocabularies and severe cold-start problems. To ad- dress this, Semantic IDs (SIDs) via vector quantization are explored: TIGER [16] uses RQ-VAE for cold-start transfer, LETTER [19] opti- mizes codebooks end-to-end, and OneRec [27] adopts iterative RQ K-means for efficient hierarchical IDs. Recent work pursues unified representations [12] to integrate semantic tokens’ generalization and atomic IDs’ specificity. 5 Conclusion This paper presents GenRec, a generative retrival model with pref- erence alignment. To address industrial challenges, we synergize Multimodal Semantic IDs and an Token Merger for efficient rep- resentation learning, while employing Page-wise Generative SFT and GRPO-SR to ensure alignment with hierarchical business objec- tives. Large-scale deployment confirms that this differentiable par- adigm significantly outperforms traditional multi-stage pipelines, validating its potential as a scalable solution for next-generation recommendation. We will investigate the reasoning ability of such framework for future work. Acknowledgments This work is sponsored by Beijing Nova Program (No.20250484857). Presenter Biography. Yanyan Zou is an applied scientist in Rec- ommendation Platform at JD.com since 2020, launching cutting- edge AI models into practical productions. Her research interests primarily lie in the areas of large language model and recommen- dation, with around 20 papers published in top-tier conferences (e.g., ACL, EMNLP, AAAI). She received her B.Engr. degree in 2015 from Xiamen University, China, as well as her Ph.D. degree from Singapore University of Technology and Design in 2020. References [1]Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al.2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025). SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaYanyan Zou et al. [2]Runjin Chen, Mingxuan Ju, Ngoc Bui, Dimosthenis Antypas, Stanley Cai, Xi- aopeng Wu, Leonardo Neves, Zhangyang Wang, Neil Shah, and Tong Zhao. 2024. Enhancing Item Tokenization for Generative Recommendation through Self-Improvement. arXiv:2412.17171 [cs.LG] https://arxiv.org/abs/2412.17171 [3]Sunhao Dai, Jiakai Tang, Jiahua Wu, Kun Wang, Yuxuan Zhu, Bingjun Chen, Bangyang Hong, Yu Zhao, Cong Fu, Kangle Wu, Yabo Ni, Anxiang Zeng, Wen- jie Wang, Xu Chen, Jun Xu, and See-Kiong Ng. 2025. OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking System. arXiv:2509.18091 [cs.IR] https://arxiv.org/abs/2509.18091 [4]Mukund Deshpande and George Karypis. 2004. Item-based top-n recommenda- tion algorithms. ACM Transactions on Information Systems (TOIS) 22, 1 (2004), 143–177. [5] Ruidong Han, Bin Yin, Shangyu Chen, He Jiang, Fei Jiang, Xiang Li, Chi Ma, Mincong Huang, Xiaoguang Li, Chunzhen Jing, Yueming Han, MengLei Zhou, Lei Yu, Chuan Liu, and Wei Lin. 2025. MTGR: Industrial-Scale Generative Rec- ommendation Framework in Meituan. In Proceedings of the 34th ACM Interna- tional Conference on Information and Knowledge Management (CIKM ’25). ACM, 5731–5738. doi:10.1145/3746252.3761565 [6] Peiyu Hu, Wayne Lu, and Jia Wang. 2025. From IDs to Semantics: A Genera- tive Framework for Cross-Domain Recommendation with Adaptive Semantic Tokenization. arXiv:2511.08006 [cs.IR] https://arxiv.org/abs/2511.08006 [7] Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated gain-based evaluation of IR techniques. ACM Transactions on Information Systems (TOIS) 20, 4 (2002), 422–446. [8]Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In 2018 IEEE international conference on data mining (ICDM). IEEE, 197–206. [9] Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014). [10] Xiaoxi Li, Jiajie Jin, Yujia Zhou, Yuyao Zhang, Peitian Zhang, Yutao Zhu, and Zhicheng Dou. 2025. From matching to generation: A survey on generative information retrieval. ACM Transactions on Information Systems 43, 3 (2025), 1–62. [11]Yang Li, Kangbo Liu, Ranjan Satapathy, Suhang Wang, and Erik Cambria. 2024. Recent Developments in Recommender Systems: A Survey [Review Article]. IEEE Computational Intelligence Magazine 19, 2 (2024), 78–95. doi:10.1109/MCI.2024. 3363984 [12]Guanyu Lin, Zhigang Hua, Tao Feng, Shuang Yang, Bo Long, and Jiaxuan You. 2025. Unified Semantic and ID Representation Learning for Deep Recommenders. arXiv:2502.16474 [cs.IR] https://arxiv.org/abs/2502.16474 [13] Ruihui Mu. 2018. A Survey of Recommender Systems Based on Deep Learning. IEEE Access 6 (2018), 69009–69022. doi:10.1109/ACCESS.2018.2880197 [14]Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 2685–2692. [15]Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tianyi Tang, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. 2025. Qwen2.5 Technical Report. arXiv:2412.15115 [cs.CL] https://arxiv.org/abs/2412.15115 [16] Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al. 2023. Recommender systems with generative retrieval. Advances in Neural Information Processing Systems 36 (2023), 10299–10315. [17]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeek- Math: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300 [cs.CL] https://arxiv.org/abs/2402.03300 [18] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential Recommendation with Bidirectional Encoder Repre- sentations from Transformers. In CIKM. [19]Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2025. Learnable Item Tokenization for Generative Recommendation. arXiv:2405.07314 [cs.IR] https://arxiv.org/abs/2405.07314 [20]Chaojun Xiao, Jie Cai, Weilin Zhao, Biyuan Lin, Guoyang Zeng, Jie Zhou, Zhi Zheng, Xu Han, Zhiyuan Liu, and Maosong Sun. 2025. Densing law of llms. Nature Machine Intelligence (2025), 1–11. [21] Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, Shuanglong Li, et al.2025. Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations. arXiv preprint arXiv:2503.02453 (2025). [22] Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, Yinghai Lu, and Yu Shi. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. arXiv:2402.17152 [cs.LG] https://arxiv.org/abs/2402.17152 [23] Jun Zhang, Yi Li, Yue Liu, Changping Wang, Yuan Wang, Yuling Xiong, Xun Liu, Haiyang Wu, Qian Li, Enming Zhang, Jiawei Sun, Xin Xu, Zishuai Zhang, Ruoran Liu, Suyuan Huang, Zhaoxin Zhang, Zhengkai Guo, Shuojin Yang, Meng-Hao Guo, Huan Yu, Jie Jiang, and Shi-Min Hu. 2025. GPR: Towards a Generative Pre-trained One-Model Paradigm for Large-Scale Advertising Recommendation. arXiv:2511.10138 [cs.IR] https://arxiv.org/abs/2511.10138 [24] Junjie Zhang, Beichen Zhang, Wenqi Sun, Hongyu Lu, Wayne Xin Zhao, Yu Chen, and Ji-Rong Wen. 2025. Slow Thinking for Sequential Recommendation. arXiv:2504.09627 [cs.IR] https://arxiv.org/abs/2504.09627 [25]Zhaoqi Zhang, Haolei Pei, Jun Guo, Tianyu Wang, Yufei Feng, Hui Sun, Shaowei Liu, and Aixin Sun. 2025. OneTrans: Unified Feature Interaction and Sequence Modeling with One Transformer in Industrial Recommender. arXiv:2510.26104 [cs.IR] https://arxiv.org/abs/2510.26104 [26] Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448. [27] Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Shiyao Wang, Weifeng Ding, Wuchao Li, Xinchen Luo, Xingmei Wang, Zexuan Cheng, Zixing Zhang, Bin Zhang, Boxuan Wang, Chaoyi Ma, Chengru Song, Chenhui Wang, Di Wang, Dongxue Meng, Fan Yang, Fangyu Zhang, Feng Jiang, Fuxing Zhang, Gang Wang, Guowang Zhang, Han Li, Hengrui Hu, Hezheng Lin, Hongtao Cheng, Hongyang Cao, Huanjie Wang, Jiaming Huang, Jiapeng Chen, Jiaqiang Liu, Jinghui Jia, Kun Gai, Lantao Hu, Liang Zeng, Liao Yu, Qiang Wang, Qidong Zhou, Shengzhe Wang, Shihui He, Shuang Yang, Shujie Yang, Sui Huang, Tao Wu, Tiantian He, Tingting Gao, Wei Yuan, Xiao Liang, Xiaoxiao Xu, Xugang Liu, Yan Wang, Yi Wang, Yiwu Liu, Yue Song, Yufei Zhang, Yunfan Wu, Yunfeng Zhao, and Zhanyu Liu. 2025. OneRec Technical Report. arXiv:2506.13695 [cs.IR] https://arxiv.org/abs/2506.13695