betaFeedback

WorksComputer Science

KuaiRecV2: Benchmarking Large-Scale Continual Learning for Diversified and Multi-task Recommendation

Abstract

Recommender systems, which retrieve items of interest to users based on their interaction histories, have recently witnessed several key technological advancements, benefiting from the semantic representation space and model-data scaling. Nevertheless, practitioners and researchers have frequently encountered critical inconsistencies between simple offline evaluations and complex online results, and this gap still exists in the era of large models. In this work, we address several key factors that contribute to this inconsistency: 1) the large-scale continual learning challenge that assumes a practical demand to maintain long-term model effectiveness under dynamic item pools and distributional shifts; 2) the combinatorial diversity challenge that requires the generation strategy to solve the accuracy-diversity trade-off in both the finer token level and the holistic item level; and 3) the multi-task reward balancing challenge, which is particularly critical for modern generative models that can accurately model the sequence distribution but are less controllable for the multi-task demands. To address these challenges and further pave the way for the development of more realistic recommendation solutions, we introduce KuaiRecV2, a deliberately constructed benchmark with industrial-level datasets from Kuaishou. Specifically, we provide a large-scale interaction dataset with million-scale multi-modal item information and long-term interaction records across 30 days, with a dynamic video candidate pool. Additionally, we provide a rigorous benchmark that evaluates retrieval accuracy, diversity, and multi-task performance in one rubric system, where an effective advancement occurs if and only if the new method improves all metrics. The code is available at https://github.com/cxli2333/KuaiRecV2

Introduction

Recommender systems [1, 2] capture user characteristics and preferences from their interaction histories and filter new content of interest that fulfills users’ demands for daily information consumption. Due to the similarity in the sequential modeling formulation, recommendation technologies have witnessed significant development in expressiveness and accuracy, as they benefit from the advancements of large language models and multi-modal models [3, 4]. In addition to the LLM-enhanced recommendation paradigm [5], recent discussions on the potential of LLM-style recommendation [6, 7] also reveal that scaling the sequential recommender on generalized semantic representations may achieve state-of-the-art performance, with generative retrievers [8, 9, 10, 11, 12] and generative rankers [13, 14, 15, 16] as representatives. The great expectation under these new methodologies is a more guaranteed performance boost with the scaling of data and model size, i.e., achieving a better scaling law behavior compared with conventional recommenders [15].

Practical Concerns for LLM-style Recommenders

Technically, these LLM-style generative models typically consist of: a) an item encoder, e.g., the item tokenizer, that transforms the item information into a token sequence in a unified representation space; b) a model with scalable sequential recommender architecture that captures the sequential patterns in the unified token space; c) a decoder, i.e., an item retriever that outputs items based on the generated tokens from the recommender; and d) a trainer that either applies generative, discriminative, or hybrid learning strategies. Though promising, as this field has matured from its early stage, designers have frequently encountered critical challenges in practice:

  • Continual Learning under Dynamic Content Pool (RQ1): In content-intensive scenarios, e.g., video sharing platforms, composers frequently upload new materials and consumers continuously interact with the platform on a daily basis. On the one hand, this results in a highly dynamic item pool that forces the recommender system to design robust and long-term consistent solutions. Some studies [17, 18, 19] show that generative recommenders with generalized item representation methods (e.g., semantic representations) are more effective at handling incoming new items, but others notice that item IDs can become more expressive after these items receive sufficient interactions [20, 21, 22]. On the other hand, samples are chronologically fed into the training process in practice, following a continual learning environment [23] that directly avoids the pitfall of “learning the future, evaluating the past” which is ubiquitous in conventional benchmarks [24, 25]. This means that we need a more rigorous evaluation-while-training protocol for recommender comparisons, but it also brings new problems, including catastrophic forgetting [26], robustness under distributional shifts [27], and restricted computational resources [28].

  • Controllable Combinatorial Diversity in Token Space (RQ2): The item generation involves a token generation phase and a token-to-item retrieval phase, where both phases could be significantly biased even if the tokens are well-separated in the semantic space. In other words, the challenge of the accuracy-diversity trade-off [29, 30, 31], which is widely studied with respect to items, also exists at the token level. In some extreme cases, this problem could be even worse since each item is typically represented by multiple tokens, where token-wise biases are inevitably aggregated. Consequently, solving this trade-off is expected to be much more challenging at the token level since the diversity-boosting strategy has to consider the combinatorial effect (e.g., using diversified

beam search). On the other hand, diversity demands are often induced from explicit business requirements that have predefined categories. This terminology is not necessarily the same as the tokenization method in LLM-style generative recommenders; Thus, more effort is needed to align the diversity control protocols.

  • Multi-task Balancing and Reward Design (RQ3): Different from discriminative recommenders that can directly formulate the multi-task problem with multi-head/multi-expert predictive models [32, 33, 34, 35, 36], generative sequential recommenders are not directly designed to optimize and control different metrics. During training, some studies show that multi-task SFT may help improve the generalizability of the model [37, 38, 39], but the computational cost becomes a bottleneck due to the duplicated runs for each task during inference. Other solutions adopt an additional RL-based preference alignment phase [10, 38, 40] that merges all objectives into a reward model for generator training guidance or inference-time reranking. Yet, there is no evidence that the current design of the reward is optimal for multi-task settings, neither is there a conclusion on whether training with multi-task learning or training with a single reward is a better strategy.

Contributions:

To address these challenges and further investigate the effectiveness of modern recommendation solutions in a more realistic environment, we introduce KuaiRecV2, a benchmark that consists of deliberately constructed industrial-level datasets, rigorously augmented evaluation protocols, and a standardized leaderboard. The data were collected from the social video-sharing platform Kuaishou3. Correspondingly, we summarize our contributions as follows:

  • Large-scale industrial datasets: We first provide the KuaiRecV2-FullMonth dataset that consists of 124k users’ interaction records over 30 days, to answer the large-scale continual learning challenge. The interactions on the last day are held out for leaderboard evaluation. We further select the first two days of the interactions as KuaiRecV2-Small dataset, to recover standard offline evaluation and enable quick model validity verifications. For all datasets, we provide multi-modal features on items (including multi-modal embeddings and semantic IDs). We summarize the organization of the datasets in Figure 1 and the main statistics in Table 1.

Figure 1. Data Organization of KuaiRecV2.

  • The long-term recommendation pipeline: With the FullMonth dataset, we formulate the long-term recommendation problem, where the solution models are required to handle the dynamic first-in, first-out video candidate pool and evolving user interests. This environment also addresses the importance of model stability under continual learning strategies, which favor a generalized tokenizer and a balanced recommendation policy between cold-start items and hot items. One counterintuitive result we observe is that the RL-based alignment without restriction may potentially downgrade the model performance in the long term.

  • Augmented evaluation: The evaluation protocol augments the conventional learning-to-rank metrics with multi-level diversity estimation and multi-task evaluation. We empirically show that existing generative recommenders may suffer from challenges such as catastrophic forgetting, unstable long-term performance, diversity bottlenecks, and multi-task trade-offs.

Datasets and Recommendation Pipeline

Data Preparation

Data Collection: Our KuaiRecV2 datasets are collected from Kuaishou, a video-sharing platform serving hundreds of millions of users every day with a billion-scale candidate item set. We uniformly sample users from the full user population and collect their interaction logs over 30 days. We then apply user-side n-core filtering [41] with n=100n=100, retaining users with at least 100 interactions in the collection window to ensure sufficiently long histories for sequential-model evaluation. For each day, we retain interactions involving items created within the preceding seven days, forming a dynamic candidate item pool. As summarized in Figure 1, we provide two benchmark datasets: KuaiRecV2-FullMonth and KuaiRecV2-Small. The FullMonth dataset also includes a hold-out test set for the online leaderboard4. Table 1 summarizes their key statistics, and Appendix F provides additional analysis.

subset#user#item#recordssparsity#item/user#user/item
FullMonth124,05524,898,648484,045,14599.9843%3901.8519.44
Small124,0542,452,98738,955,94999.9871%314.0215.88

caption: Statistics of KuaiRecV2 datasets.

:::

Figure 2. Daily Statistics of KuaiRecV2.

Data Structure: Each user-item interaction record contains labels for multiple tasks, including viewing behaviors (e.g., short view and complete view), positive interactions (e.g., like and comment), and a negative interaction label (hate). For each retained item, we provide a 64-dimensional multimodal embedding using the model in [42], an eight-level semantic ID generated by a fixed RQ-VAE-style tokenizer [8], and MLLM-inferred tag IDs [43] for semantic matching analysis. All these models were fixed during the generation of item meta information, and none was trained on the released 30-day interaction logs or held-out interactions, to avoid information leakage. The tokenizer card reports its training corpus, time window, and model version, and separately discloses any item-content overlap with external pretraining data. User IDs, item IDs, and tags are independently reindexed within the released datasets, and user profile features are omitted for privacy protection. Detailed data descriptions and sample examples are provided in Appendix § 19.

The Long-term Recommendation Pipeline

In addition to the KuaiRecV2-FullMonth dataset, we provide a recommendation pipeline designed for a more realistic continual learning environment. In such an environment, the recommendation models are required to adapt to the time-varying item candidate pool and dynamic interaction sequences. Specifically, we summarize the following key features:

  • Consecutive training on a daily basis: On the first day, the pipeline trains a randomly initialized generative recommender or a pretrained model using the data from Day 1, i.e., D1\mathcal{D}_1. Denote the resulting model as θ1\theta_1. Starting from t=2t = 2, for each subsequent day, the model parameters θt−1\theta_{t-1} learned from day t−1t - 1 are used to initialize the training on the newly arrived data Dt\mathcal{D}_t, generating a new model θt\theta_t. In this way, the model is continually updated as the data stream evolves over time.

  • Cross-day history context: To preserve temporal consistency across days, for each user, when constructing the first few interaction instances of a new day, we concatenate the last several interactions from the previous day as historical context. This design avoids breaking user sequences at day boundaries, providing a more accurate representation of real-world streaming recommendation scenarios. For better regularization of the history lengths, we provide hyperparameter controls on their minimum Lmin⁡L_{\min} and maximum Lmax⁡L_{\max}.

  • Dynamic candidate pool: As illustrated in Figure 1, we use a sliding-window strategy that only includes items created within the past 7 days for each day’s interactions. For any Day tt, the candidate pool will remove the candidate items created at Day t−7t - 7 and admits any items created at Day tt. As we will show in § 4.1, one may observe a downgrade of model performance whenever a new day’s data comes in with unseen items, even if they are using the shared token space. This addresses the need for stable model training and inference under a continual learning environment. In practice, the item removals and increments may happen at any time of the day, inducing smoother shifts in the distribution. Yet, the challenge of the model stability under candidate pool changes in long-term recommendation scenarios still exists.

Benchmarks

Though the pipeline in section 2.2 leaves specific model and training settings to researchers and future studies, we provide implementations of several key baselines and basic training strategies, in addition to the pipeline and the augmented evaluation protocol.

Task Formulation and Augmented Evaluation Metrics

Evaluation Protocol: In the long-term recommendation pipeline, we implement a daily leave-one-out evaluation, i.e., for each DtD_t and any given user’s interaction logs in DtD_t, all interactions before the last one (of the user) are used for training, while the final one is used for validation. Nevertheless, we encourage readers to customize their evaluation protocols for more advanced settings. Note that for the FullMonth data, we engage in continual learning through {D1,D2,…,D30}\{D_1,D_2,\ldots,D_{30}\}, where DiD_i represents the data of the ii-th day. For each day tt, we first compute every metric on the NtN_t validation instances of that day. When a single FullMonth score is required, we use the instance-weighted aggregation MFullMonth=(∑tNtMt)/(∑Nt)M_{FullMonth} = (\sum_t N_t M_t)/(\sum N_t), where MtM_t is the corresponding daily metric. We retain the daily scores as the primary report because the trajectory itself exposes adaptation and forgetting. For backward evaluation, checkpoint θk\theta_k is evaluated on the Day 1 instances, and the resulting score is reported as the Day-k-to-Day-1 retention performance.

Problem Formulation: Given a user request that consists of a user ID and the interaction history H\mathcal{H}, the recommendation model θ\theta is responsible for retrieving the item candidates that best align with user preferences. Formally, in addition to the user history, each training sample includes a target item ii and multi-behavior feedback Yi=(yi,1,…,yi,k)Y_i = (y_{i,1},\ldots,y_{i,k}) for kk labels on this item. Then the goal of the optimization includes three major metric families:

  • Retrieval accuracy metrics: Following standard evaluation for retrievers, include ranking metrics such as NDCG and HitRate (HR) for the main behavior “click”. Note that under a continual learning environment, we aim to learn a model that exhibits stability on these metrics across training days or even gradually increases these metrics with the long-term recommendation pipeline.

  • Multi-level diversity metrics: We first provide the conventional item-level diversity metrics such as Item Coverage (IC) and GINI coefficient. To better observe the combinatorial effects on the diversity, we include several token-level diversity metrics: TopK Layer Diversity reflects diversity in top-KK predictions, Average Path Diversity measures structural diversity of recommendation paths, Average Layer Unique Tokens quantifies token variety, and Average Layer Entropy captures prediction uncertainty. Detailed explanations are provided in the appendix D.

  • Multi-task metrics: For the kk labels in YY space, we set corresponding rate-prediction tasks as B=(b1,…,bk)B = (b_1,\ldots,b_k) (e.g., like-rate prediction for behavior “like”) and observe the ranking metrics (i.e., NDCG and HR) for each of the task. For behavior-wise evaluation, we further partition the test instances according to the behavior type of the held-out interaction. We additionally allow the training data to be filtered based on the choice of task, so that the model can perform multi-task SFT. Nevertheless, we enforce a single-retrieval evaluate-all strategy during evaluation, where only one model can provide recommendations for the estimation of all task metrics.

Implemented Baselines

Models

For reproduction purposes and to identify the challenges of generative sequential recommenders (mentioned in section 1), we implemented several key baselines under the sequential recommendation formulation: SASRec [44], GRURec [45], TIGER [46], LLaDA-Rec [47], MiniOneRec [48]. More detailed information about these models can be seen in Appendix § 18.1.

Training Strategies

Task-aware sequence modeling: Given the set of tasks BB that correspond to kk types of interactions, while our default setting uses “click” sequences for model training and evaluation, we remind readers that constructing user interaction sequences based on different behavioral types in BB may better suit the learning of the corresponding task [10]. The comparison for this part is in the appendix § 20.1.

RepresentationScaleTasks
FORGE [25] (2025.09)Extracted SID & MMU embeddings14B interactions, 100K users, 250M itemsSID quality assessment
MUSE [49] (2025.12)User/Item feature & MMU embeddings99M interactions, 8M users, 35M itemsUser Interest Modeling
KuaiRec [24] (2022.02)User/Item feature4M interactions, 1,411 users, 3,327 itemsFully-observed Dataset Evaluation
KuaiRecV2Extracted SID & MMU embeddings484M interactions, 124K users, 24M itemsLong-term Recommendation & Diversity & Multi-task Retrieval

Table 2. Comparisons between related benchmarks.

SFT vs. RL: In addition to standard SFT on the next-item prediction dataset, we extend the recommendation model with reinforcement learning (RL) to optimize recommendation lists toward user feedback-oriented objectives. Given a user history H\mathcal{H}, the policy θ\theta generates a ranked list of items X\mathcal{X}. To align with user preferences, we design reward functions, formatted as RB=∑b∈BαbRbR_B = \sum_{b \in B} \alpha_b R_b, based on behavior types in BB, and different interaction behaviors are given different levels of significance α\alpha in user preference estimation. This enables the RL policy to favor recommendations that are more likely to induce desired user responses. Note that this reward function is implemented merely as training signals but not evaluation metrics, and we evaluate the RL-optimized model under the same protocol as in section 3.1.

Related Benchmarks

Table 2 compares KuaiRecV2 with related benchmarks. Existing datasets such as FORGE [25] and MUSE [49] are mainly designed for SID quality assessment and user interest modeling, respectively, while KuaiRec [24] is an earlier version of the benchmark that focuses on fully-observed dataset evaluation in a much smaller-scale setting. In contrast, KuaiRecV2 is built for continual learning, combinatorial diversity, and multi-task retrieval. The substantially large-scale dataset enables a better scaling law observation for existing methods. It also adopts extracted SID and Multi-Media Understanding (MMU) embeddings to support diversified generative recommendation modeling and evaluation. Full related work is provided in Appendix § 14.

Benchmark Results, Key Observations, and Open Questions

Evaluating Continual Adaptation to Dynamic Item Poolsol (RQ1)

We evaluate generative models on KuaiRecV2-FullMonth with evolving item pools using the long-term recommendation pipeline (section 2.2). Figure 3 presents the loss curve of MiniOneRec [48] during training over the first five days on the FullMonth dataset. It indicates that item replacement under a dynamic item pool requires the model to relearn and adapt continuously. More detailed results on the full 30-day continual learning are provided in Appendix § 20.

Training loss during 5-day continual learning

Figure 3. Training loss during 5-day continual learning. Orange squares indicate the initial loss at each day boundary (when item pools change), and green dots show the final loss after training. The inset shows the rapid initial convergence on Day 1.

The existence of distributional shift and catastrophic forgetting in long-term recommendation.

We evaluate backward transfer by testing each checkpoint on the Day 1 validation instances after training through Day kk. We compare MiniOneRec with two continual-learning baselines: EWC LoRA, which applies elastic weight consolidation to LoRA parameters during daily updates, and Replay, which retains 5% of previously observed training instances in a replay buffer. Table 3 reports Day-1 HR@50. EWC LoRA improves retention through Day 5, while Replay performs best from Day 2 onward, reaching 3.20 after Day 7. Despite these gains, all methods show substantial forgetting, highlighting the need to evaluate both adaptation to the current item pool and retention of earlier preferences.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7
MiniOneRec39.0917.896.6251.030.410.540.19
MiniOneRec + EWC LoRA38.6722.937.181.130.600.530.20
MiniOneRec + Replay (5%)39.0925.2410.265.595.453.973.20

Table 3. Backward transfer (HR@50 on Day 1), checkpoints trained Days 1–7.

Item cold-start and temporal dynamics of user interests: New items enter the catalog daily, but they lack historical interactions, forcing models to balance exploring new items with exploiting known user preferences—a trade-off that is particularly acute in continual learning. While using semantic representations like SID and MMU embeddings may mitigate the issue, we still observe daily loss spikes in Figure 3, indicating insufficient adaptation. Systematically, the dynamic item pool also contributes to user preference shifts, inducing different patterns at different timescales—short-term interests may shift daily, while long-term preferences remain stable. Current approaches treat all temporal dynamics uniformly, failing to capture the multi-scale nature of user behavior evolution. Intuitively, we expect a recommender system that can reduce catastrophic forgetting of long-term preferences rather than short-term preferences.

These interconnected challenges highlight why continual recommendation is fundamentally more difficult than static settings and underscore the need for algorithms that can efficiently handle distribution shifts, item cold-start scenarios, and multi-scale temporal dynamics.

Token-level and Item-level Diversity Metrics (RQ2)

We evaluate all implemented models on KuaiRecV2-Small. We summarize their ranking performance in Table 4 and their item-level diversity metrics in Table 5. Note that due to the excessive number of items, we only calculate the Gini coefficient for items that appear once or more. More detailed results are provided in the appendix G.4. It can be seen from these results that the diversity performance of model recommendations is poor, with a high Gini coefficient value and low item coverage. On average, there are only 12 categories at the first level across fifty groups of SID, while users’ interests in real-world scenarios are far more diverse than this number.

ModelNDCG@1NDCG@5NDCG@10NDCG@20NDCG@50
SASRec0.28 (0.03)0.68 (0.08)0.97 (0.11)1.31 (0.09)1.92 (0.11)
GRURec0.31 (0.04)0.80 (0.05)1.20 (0.08)1.71 (0.10)2.82 (0.16)
LLaDA-Rec0.78 (0.08)2.25 (0.10)3.10 (0.11)4.21 (0.08)5.60 (0.11)
TIGER-Tiny1.26 (0.09)2.96 (0.12)4.01 (0.17)5.26 (0.31)6.98 (0.55)
TIGER-Small2.12 (0.07)5.34 (0.09)7.37 (0.17)9.81 (0.17)12.90 (0.26)
MiniOneRec3.53 (0.04)8.72 (0.07)12.01 (0.15)15.33 (0.18)19.22 (0.82)
ModelHR@1HR@5HR@10HR@20HR@50
SASRec0.28 (0.03)1.10 (0.08)1.96 (0.11)3.31 (0.08)6.61 (0.10)
GRURec0.31 (0.04)1.32 (0.06)2.71 (0.12)5.05 (0.48)10.99 (0.62)
LLaDA-Rec0.78 (0.08)3.71 (0.17)6.47 (0.20)11.02 (0.29)17.31 (0.42)
TIGER-Tiny1.26 (0.09)4.72 (0.16)8.01 (0.70)13.17 (1.29)20.52 (3.27)
TIGER-Small2.12 (0.07)8.71 (0.12)15.58 (0.61)25.07 (0.72)41.07 (0.81)
MiniOneRec3.53 (0.04)14.05 (0.12)23.59 (0.33)38.45 (0.41)62.63 (1.10)

Table 4. Performance comparison in terms of NDCG and HR at different cutoffs. Values are means over three random seeds (42, 2026, and 999), with standard deviations in parentheses. All values are reported in 10−310^{-3} for readability.

ModelsGini (×10−2\times10^{-2})Coverage (×10−4\times10^{-4})
@1@5@10@20@50@1@5@10@20@50
LLaDA-Rec91.4392.3393.1093.6092.411.264.459.5623.5398.86
TIGER-Tiny93.6996.0796.4296.4795.870.834.238.6918.4248.73
TIGER-Small86.6990.1091.5092.5193.101.835.6210.2020.0554.38
MiniOneRec88.1089.5089.6089.5190.691.283.255.038.1318.80

Table 5. Diversity metrics of different models at various cutoffs. All models exhibit limited diversity, as reflected by high Gini values and low coverage scores across cutoffs.

Combinatorial diversity in the token level: Table 6 summarizes the token-level diversity metrics obtained from the MiniOneRec model on training day 21. We can see that most of the diversity is achieved by the first few tokens of an item, indicating that the latter tokens are extremely biased, with little room for diversity control. With both the item-level and token-level diversity metrics available, recommender systems are now entering a more sophisticated but flexible state on diversity control. For example, one can choose whether the given user demands a more accurate recommendation with items in the same first-level token but differing in latter positions, or a more diverse recommendation with items from different first-tokens.

MetricLayer
12345678
Avg Path Diversity12.4731.3442.8445.9346.6146.8347.5048.69
Avg Layer Unique Tokens12.4725.6138.6442.8143.5343.8444.8646.04
Avg Layer Entropy2.744.115.095.325.355.375.425.47
TopK Layer Diversity(Top1)1.001.001.001.001.001.001.00
TopK Layer Diversity(Top5)3.664.554.744.774.774.794.78
TopK Layer Diversity(Top10)5.998.799.459.599.609.659.65
TopK Layer Diversity(Top20)8.9315.7918.5319.0619.1319.2119.27
TopK Layer Diversity(Top50)12.4725.6138.6442.8143.5343.8444.86

Table 6. Diversity metrics across layers for streaming training day 21 (click behavior).

Alignment with existing terminologies: As we have explained in section 1, diversity demands are often induced from explicit business requirements that have predefined categories. This terminology is not necessarily the same as the tokenization method in LLM-style generative recommenders; Thus, more efforts are needed to align the diversity control protocols. Fortunately, although SID is not a manually defined classification and cannot directly match actual business terms, the SID token-level diversity metric can still reflect the variety of item semantics in the language space, partially optimizing business-defined diversities with implicit control.

Estimating Multi-task Performance (RQ3)

We extend the streaming training framework with reinforcement learning to leverage multiple behavior signals. While the SFT stage uses only click sequences, we design RL reward functions based on diverse behaviors (Appendix § 18.3). We evaluate the effectiveness of multi-behavior RL by comparing three training strategies starting from Day 20: (1) RL-only, where we apply only RL without SFT; (2) SFT-only, where the model continues to train with SFT and no RL is applied; and (3) SFT and RL alternating, where SFT and RL are performed alternatively in each daily training step. Figure 4 shows the performance comparison across these strategies. Days 1–19 are trained with SFT only. Days 20–30 are trained with alternating SFT and RL. Blue circles denote post-RL results, red diamonds denote pre-RL results, and green markers denote RL-only results.

Figure 4. HR@50 across training days. We compare the results of alternating SFT and RL with those of RL-only.

Multi-behavior RL improves performance: On Day 20, the alternating SFT and RL strategy achieves noticeable improvements over the SFT-only baseline, indicating that multi-behavior rewards successfully guide the model toward better performance and user preference alignment. By first adapting to new data through SFT and then optimizing the objective via RL, this hybrid training scheme better handles the non-stationary nature of streaming recommendation.

Data shifting challenges in multi-task recommendation. Compared to SFT+RL and SFT-only methods, the RL-only strategy reveals a critical limitation: although RL improves performance on Day 20, it fails to maintain performance when the item pool changes on Day 21. This sharp degradation suggests that RL alone is insufficient to adapt to new data distributions or large-scale continual learning. Following these findings, we show evidence that RL-based solutions for multi-task recommendation might be fragile to distribution shifts. In our empirical study, multi-tasking RL is effective only when combined with SFT.

Task relations and effective advancement. We further provide Table 7 that presents the performance metrics of MiniOneRec (SFT-only) across different behaviors, evaluated at the 21st training day. Semantically related behaviors tend to exhibit similar recommendation patterns. For example, long view, like, and profile enter achieve comparable NDCG and HR values across cutoffs, suggesting that the model captures shared preference structures among positive feedback signals. In contrast, behavior-specific signals can differ substantially. For instance, hate remains near zero at all cutoffs, indicating a sharply different retrieval pattern (e.g., anomaly detection) from other behaviors. More detailed results are provided in Appendix G.3. In general, task metrics may also exhibit trade-off performances in practice (e.g., when the model is optimized without careful designs [35]). In such cases, we note them as defective advancement, and only the new method improves all metrics can be regarded as effective advancement.

BehaviorNDCG (×10−3\times10^{-3})HR (×10−3\times10^{-3})
@1@5@10@20@50@1@5@10@20@50
click2.877.309.8512.7816.032.8711.7719.7031.3947.60
long view3.188.0210.7813.9717.433.1812.8921.4934.2051.50
like2.487.059.4812.4716.452.4812.0319.4731.5150.99
follow0.007.309.6610.5310.530.0015.3222.9826.8126.81
forward2.645.446.328.309.822.647.9110.5518.4626.38
comment3.745.207.299.3812.723.746.5413.0821.4938.31
search0.003.2610.7114.6316.180.007.9631.8747.8055.77
profile enter2.715.207.7010.2413.092.718.1215.9126.0740.29
hate0.000.000.000.000.000.000.000.000.000.00

Table 7. Behavior metrics at different cutoffs. Related behaviors show similar ranking performance, while unrelated behaviors, such as hate, differ sharply.

Scalability Comparison on Large-Scale Item Pools

We first evaluate different recommendation architectures on KuaiRecV2-Small. This setting allows us to assess how different architectures scale to large item pools before introducing temporal dynamics. Table 4 shows the detailed results.

Generative sequential models substantially outperform traditional sequential models. All three generative models (MiniOneRec, LLaDA-Rec, TIGER) achieve higher NDCG compared to traditional models (SASRec, GRU4Rec). This gap reflects the scalability limitations of traditional ranking-based models on large item pools. Generative models, designed for large output vocabularies through token-based generation, scale more efficiently.

Model capacity matters for generative models. MiniOneRec (0.5B) and TIGER-Small (30M) substantially outperform smaller models (TIGER-Tiny: 4.9M, LLaDA-Rec: 5.3M), indicating that larger capacity enables better performance on large-scale item pools. Notably, even the smallest generative model substantially outperforms the best traditional model, indicating that architectural choice is more critical than model size alone.

Through our benchmark datasets, we believe that it becomes easier to validate the scaling behavior of existing and future generative recommender solutions, in both static offline and continual learning environments.

Conclusion and Discussion

This paper introduces KuaiRecV2, a benchmark for large-scale diversified and multi-task recommendation under a continual learning environment. The benchmark covers 124 thousand users, 24 million items, as well as 0.48B interactions across a long-term span of 30 days, providing ultra-long interaction sequences. We present the continual learning design choices within our long-term recommendation pipeline, which implements a rigorous evaluation system and an online leaderboard. On this basis, KuaiRecV2 narrows the gap between offline experiments and online experiments by a comprehensive evaluation that consists of long-term recommendation accuracy, multi-level diversity, and multi-task balancing. We present limitations and future work in Appendix § 15.

Acknowledgments

This research is supported by the National Natural Science Foundation of China (No.U25A20445 and No.62625608).

NeurIPS Paper Checklist

Claims

Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope?

Answer: [Yes]

Justification: The main claims in the abstract and introduction are consistent with the paper’s theoretical contributions, experimental results, and scope.

Guidelines:

  • The answer [N/A] means that the abstract and introduction do not include the claims made in the paper.

  • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers.

  • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.

  • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.

Limitations

Question: Does the paper discuss the limitations of the work performed by the authors?

Answer: [Yes]

Justification: The limitations of this work are explicitly discussed in Appendix B of the main paper.

Guidelines:

  • The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper.

  • The authors are encouraged to create a separate “Limitations” section in their paper.

  • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.

  • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.

  • The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.

  • The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.

  • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.

  • While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.

Theory assumptions and proofs

Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [N/A]

Justification: the paper does not include theoretical results.

Guidelines:

  • The answer [N/A] means that the paper does not include theoretical results.

  • All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.

  • All assumptions should be clearly stated or referenced in the statement of any theorems.

  • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.

  • Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.

  • Theorems and Lemmas that the proof relies upon should be properly referenced.

Experimental result reproducibility

Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?

Answer: [Yes]

Justification: We fully disclose all necessary details for reproducing main results, and provide code and data to support faithful reproduction, as detailed in the supplementary material.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.

  • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.

  • Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.

  • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.

Open access to data and code

Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?

Answer: [Yes]

Justification: We provided code and data.

Guidelines:

  • The answer [N/A] means that paper does not include experiments requiring code.

  • Please see the NeurIPS code and data submission guidelines (https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.

  • While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).

  • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.

  • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.

  • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.

  • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).

  • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.

Experimental setting/details

Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results?

Answer: [Yes]

Justification: We discuss experimental details in Appendix.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.

  • The full details can be provided either with the code, in appendix, or as supplemental material.

Experiment statistical significance

Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?

Answer: [No]

Justification: Statistical significance analysis is not provided in this work. Experimental results are consistent across repeated runs to ensure reliability.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.

  • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).

  • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)

  • The assumptions made should be given (e.g., Normally distributed errors).

  • It should be clear whether the error bar is the standard deviation or the standard error of the mean.

  • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified.

  • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates).

  • If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.

8. Experiments compute resources

Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?

Answer: [Yes]

Justification: We have specified the computational resource requirements in the appendix.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.

  • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.

  • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper).

9. Code of ethics

Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?

Answer: [Yes]

Justification: We strive to minimize the exposure of any personally identifiable information.

Guidelines:

  • The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics.

  • If the authors answer [No], they should explain the special circumstances that require a deviation from the Code of Ethics.

  • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).

10. Broader impacts

Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?

Answer: [Yes]

Justification: We present a corresponding discussion in Appendix C.

Guidelines:

  • The answer [N/A] means that there is no societal impact of the work performed.

  • If the authors answer [N/A] or [No], they should explain why their work has no societal impact or why the paper does not address societal impact.

  • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.

  • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.

  • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.

  • If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).

Safeguards

Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)?

Answer: [N/A]

Justification: The paper poses no such risks.

Guidelines:

  • The answer [N/A] means that the paper poses no such risks.

  • Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.

  • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.

  • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.

Licenses for existing assets

Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?

Answer: [Yes]

Justification: All existing assets (datasets, code, models) are properly credited, and their licenses and terms of use are explicitly stated and respected in the paper.

Guidelines:

  • The answer [N/A] means that the paper does not use existing assets.

  • The authors should cite the original paper that produced the code package or dataset.

  • The authors should state which version of the asset is used and, if possible, include a URL.

  • The name of the license (e.g., CC-BY 4.0) should be included for each asset.

  • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.

  • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.

  • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.

  • If this information is not available online, the authors are encouraged to reach out to the asset’s creators.

13. New assets

Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?

Answer: [Yes]

Justification: All new assets are well-documented, and the detailed information is provided in both the main paper and the appendix.

Guidelines:

  • The answer [N/A] means that the paper does not release new assets.

  • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.

  • The paper should discuss whether and how consent was obtained from people whose asset is used.

  • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.

14. Crowdsourcing and research with human subjects

Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?

Answer: [N/A]

Justification: The paper does not involve crowdsourcing nor research with human subjects.

Guidelines:

  • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.

  • Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.

  • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.

15. Institutional review board (IRB) approvals or equivalent for research with human subjects

Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?

Answer: [N/A]

Justification: The paper does not involve crowdsourcing nor research with human subjects.

Guidelines:

  • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.

  • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.

  • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.

  • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

Declaration of LLM usage

Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required.

Answer: [Yes]

Justification: We clearly describe the usage of LLMs as an important component in the core methodology of this work, as stated in the main paper.

Guidelines:

  • The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.

  • Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.

Related Work

Recommendation Diversity

Besides ranking accuracy metrics, diversity metrics are equally critical in practical recommendation tasks. The mainstream applications of diversity indicators are dominated by list-level metrics, including Maximum Marginal Relevance[50], Intra-list Diversity[51], Normalized Discounted Cumulative Gain[52], and Determinantal Point Process Score[53, 54]. Other works also reveal that system-level diversity metrics such as item coverage[55], entropy[56], GINI-index[57], and list variance[58] are also important for maintaining a content-productive environment. In this work, we extend the aforementioned system-level diversity indicators to the token level. For generative recommendation models, token-level diversity metrics can reflect the diversity of generated results in a more granular manner.

Generative Recommendation

The pioneering attempts that studied the generative recommendation formulation date back to a decade ago [59, 60, 61] when deep generative models [62, 63, 64] were proposed in machine learning and later integrated in recommender systems. The main difference between generative and discriminative methodologies is whether the recommender is modeling the item generation probability (i.e., P(i∣u)P(i|u)) or the expected feedback from the user (i.e., P(y∣i,u)P(y|i,u)). While the generative modeling of P(i∣u)P(i|u) is widely adopted in retrieval tasks [8, 9, 10, 11, 12], the discriminative modeling of P(y∣i,u)P(y|i,u) is more favorable in ranking tasks [13, 14, 16]. Despite this divergent viewpoint, both methodologies are functionally feasible for both retrieval and ranking tasks, and some recent works [13] also notice that recommenders are fundamentally energy-based models that are simultaneously generative and discriminative (i.e., the generation probability should be proportional or monotonic to the user’s positive feedback).

Technically, the generative modeling strategies in recommendation generally consist of two focuses: 1) The first focus is autoregressive generative solutions that aim to model user interaction sequences as next item/list prediction tasks. Following the insights from the development of large language models(LLMs), new architectures [8, 10, 65] are proposed for both retrieval and ranking tasks to reformulate the user interaction sequences with different tokenization techniques. For list-wise generation, additional designs are required to capture the candidate context information and item correlations in the output list [66, 67, 15]. 2) The second focus is non-autoregressive generative solutions [68, 69, 70] that aim to model the distribution of user preferences across all items.

Except for these LLM-style generative sequential recommenders, there are other directions that incorporate generative thinking. VAE-based models [62] are capable of generating the entire item space [60] and list space [71], recent efforts have found that diffusion-based models [72] and generative flow networks may have higher potential in generating item embeddings [73], item scores [74], discrete item selection space [47, 75], and list permutation space [13, 76, 77].

In this work, our main focus is to study the behavior of modern generative recommendation models on a large scale dataset, especially when the item space is too large to apply standard ID-based retrieval. The proposed KuaiRecV2 benchmark is intentionally made agnostic to the recommendation models.

Continual Learning in Recommender Systems

For online systems that receive persistent user engagements and data updates, the notion of continual learning [78, 23] exists as a machine learning problem that addresses the challenges related to system dynamics, such as catastrophic forgetting [26, 79], robustness under distributional shifts [27], and restricted computational resources [28]. Existing continual recommendation methods can be broadly grouped into replay-based[80], regularization-based[81], and architecture-based approaches[82]. Our work provides the data and fundamental framework for continual recommendation.

Limitations and Future Directions

Limitations

While KuaiRecV2 moves the field towards a more online-alike evaluation of recommender systems, the data used in industrial recommendation systems can still be far more complex than ours (e.g., excluding user features for privacy concerns), making it impossible for us to fully simulate real-world online recommendation systems. Generally, there is still a major gap between offline and online evaluation of recommender systems. Besides, we do not consider cross-domain scenarios that allow users to interaction with different recommendation domains that provide a separate set of content, since KuaiRecV2 does not address the generalizability or transferability of recommendation models.

Future Directions

Scaling law under continual learning. Unlike static benchmarks, streaming recommendation benchmarks feature dynamic iteration of item sets, continuous changes in user interests, and the coexistence of multiple behavioral optimization objectives. Learning large models in such an environment usually comes with higher training costs. Additionally, in scenarios requiring continuous training, it is essential to clarify the trade-off between short-term performance improvements and long-term stability. It remains unclear how model scale, update frequency, and memory budget jointly affect the long-term adaptive capability and forgetting degree of models.

Token-to-item decoder for better diversity and multi-task performances. There still exists a semantic gap between current token-level generation and item-level output. Such token-to-item decoders can better align token-level generation with item-level semantics, and explicitly integrate diversity and multi-task optimization objectives in the decoding process. This helps alleviate the misalignment between generated content and final recommendation results, and meanwhile improves behavioral coverage and ranking quality.

Dynamic user pool. In real-world systems, users constantly join, leave, and shift their interests over time. Modeling such user-side dynamics together with evolving item pools would provide a more realistic setting for studying adaptation, robustness, and long-term personalization.

Low-activity and cold-start users. In this benchmark, we set user sequence lengths with n≥100n \ge100. In the continual learning setting, this means that the day-averaged lower bound is around 3 interactions per user per day. The resulting data accounts for more than 93% of the original log dataset before filtering. Yet, for extremely low-activity and cold-start users, though they are less representative in the industrial environment, they still account for users in millions in large-scale systems. We considered it a different problem setting where continual learning on these users may require data collection across a much longer range, and these user groups require more careful modeling and evaluation methods.

Broader Imapacts.

Our work focuses on new-generation recommendation models and has both positive and negative implications for society. On the positive side, this publication will boost the research field forward to establish more robust, task-balanced, fairness-aware, and diversity-aware generative sequential recommendation models, and benefit the study of large models in general. The benchmark also potentially improves user engagement and satisfaction on web services. On the negative side, malicious use of this dataset may also exist, including attacks against recommender systems. Further complementary research in this direction may also be necessary, especially for large-scale datasets like ours.

Detailed Diversity Metrics

  • Gini Coefficient: Measures the statistical inequality of token distribution across all generated SIDs, computed as the Gini coefficient of token frequencies.

  • TopK Layer Diversity: Measures the diversity of tokens at each layer when evaluating top-K recommendations. For each layer ii in the SID, we compute the proportion of unique tokens among the top-K predictions.

  • Average Path Diversity: Measures the average number of distinct paths (complete SIDs) from the root to layer ii, reflecting the diversity of recommendation paths up to that layer.

  • Average Layer Unique Tokens: When generating KK candidate SIDs, measures the average number of unique tokens at each layer across all KK candidates.

  • Average Layer Entropy: Measures the entropy of token distribution at each layer when generating KK candidate SIDs, quantifying the uncertainty and diversity of predictions at each hierarchical level.

Detailed Experiments Settings

Models

The models used in our experiments are as follows:

  • SASRec[44] utilizes a self-attentive model to capture item correlations.

  • GRURec[45] uses a Gated Recurring Unit (GRU) to enhance long-term memory in recommendation tasks.

  • TIGER[46] proposes a novel learned Semantic ID representation for each item and achieved excellent performance in sequential recommendation tasks. In this paper, we use two sizes of the TIGER model.

  • LLaDA-Rec[47] is a diffusion-based generative recommendation model that captures user preferences through iterative denoising for recommendation generation.

  • MiniOneRec[48] is a lightweight open-sourced generative recommendation model that formulates recommendation as an autoregressive next-item generation task based on user interaction sequences.

Experiments Settings

SASRec and GRURec. The optimizer is AdamW, the learning rate is denoted by lr, weight decay by l2_decay, the learning rate scheduler is get_cosine_schedule_with_warmup, the warm-up steps are warmup_steps, and the loss function is BCEWithLogitsLoss. A set of basic hyperparameters is shared: epoch=10, batch_size=512, hidden_factor=256, lr=0.001, l2_decay=0.05, warmup_steps=1000, dropout_rate=0.5, max_k=200, patience=5, num_negatives=1000, num_workers=4, item_num=2452987.

TIGER. Both the Small and Tiny versions of the model adopt the official parameter sizes. We use GPUs and enable bf16 mixed precision. The core hyperparameters include: learning rate lr=1e-4, epochs per round epochs_per_round=1, training batch size train_batch_size=256, and evaluation batch size eval_batch_size=32.

LLaDA-Rec. The model adopt the official parameter sizes. We use GPUs and enable bf16 mixed precision. The core hyperparameters include: learning rate lr=1e-3, epochs per round epochs_per_round=1, training batch size train_batch_size=256, and evaluation batch size eval_batch_size=32.

MiniOneRec. The SFT experiment adopts Qwen2.5-0.5B as the base model, performs daily continued fine-tuning in streaming scenarios, and conducts GPU parallel training via accelerate launch. The training hyperparameters include num epochs = 1 and learning rate = 1e-4. The inference batch size is set to batch size = 4, and the beam search width is set to num beams = 50. For the overall experiment, only the click behavior click is retained as the behavior type for training and evaluation.

The RL experiment adopts SFT-Qwen2.5-0.5B as the base model and conducts continued training under the streaming reinforcement learning setting. The hyperparameters in the training phase are as follows: num train epochs=1, learning rate=1e-6, train_batch_size=256, rl eval batch size=32, gradient_accumulation_steps=1, temperature=1.0, num generations=8, beta=0.2. The adopted reward type is reward_type=exact_prefix_behavior, which is introduced in Appendix § 18.3. Alternative options include exact, prefix, semantic, and exact_prefix_semantic. The weight and ratio parameters related to sample construction are wexact=5w_{\mathrm{exact}}=5, wprefix=1w_{\mathrm{prefix}}=1, wsem=0.2w_{\mathrm{sem}}=0.2, subtree sample ratio=0.020.02, wpos=10w_{\mathrm{pos}}=10, winter=10w_{\mathrm{inter}}=10, wneg=1w_{\mathrm{neg}}=1. Among them, the positive behavior is long_view, interactive behaviors include like, follow, forward, comment, search and is_profile_enter, and the negative behavior is hate. The training also supports several toggle options, including add_gt, beam_search, test_during_training, dynamic_sampling, sync_ref_model, dapo and gspo. The evaluation phase is executed automatically by default, adopting distributed inference with eval_num_gpus=8, along with eval_batch_size=16, eval_num_beams=50, eval_max_new_tokens=32, eval_k=50, and the random seed is set to seed=42.

RL Reward Design

While the SFT stage uses only click sequences, we design RL reward functions based on diverse behaviors: long-view, like, forward, comment, search, profile-enter, hate. More details about experiments setting and these behaviors can be found in appendix E. As described in § 3.2.2, we design the interaction reward as

RB=∑b∈BαbRb.R_B = \sum_{b \in B} \alpha_b R_b.

Specifically, we regard long-view interactions as positive behaviors, while likes, forwards, comments, searches, and profile entries are defined as general interactive behaviors, and hate interactions are treated as negative behaviors.

We also introduce additional reward designs to engage fine-grained optimization, including the SID prefix matching reward

Rpre=1L∑i=1LI[prefixigen=prefixitarget].R_{\mathrm{pre}} = \frac{1}{L}\sum_{i=1}^{L}\mathbb{I}[\mathrm{prefix}^{\mathrm{gen}}_i = \mathrm{prefix}^{\mathrm{target}}_i].

where LL is the total length of the SID. Moreover, we adopt a semantic reward

Rsem=cos⁡(egen,etarget),R_{\mathrm{sem}} = \cos(e_{\mathrm{gen}}, e_{\mathrm{target}}),

determined by the cosine similarity between the generated SID and the target SID. The final reward is

R=λBRB+λpreRpre+λsemRsem.R = \lambda_B R_B + \lambda_{\mathrm{pre}} R_{\mathrm{pre}} + \lambda_{\mathrm{sem}} R_{\mathrm{sem}}.

Data Specification

Data Structure

Each interaction sample contains the following fields:

  • user_id: an integer identifier of the user.

  • item_id: an integer identifier of the item, which is aligned with the item metadata file.

  • timestamp: the interaction time in Unix epoch milliseconds, e. g., 1755009906758.

  • play-time and duration: continuous feedback signals measured in milliseconds.

  • long-view, like, follow, forward, comment, search, and profile-enter: binary positive labels indicating whether the corresponding behavior occurred.

  • hate: a binary negative label indicating whether the user explicitly expressed negative feedback.

The item metadata file contains the following fields:

  • sid_list: the SID sequence representation of the item.

  • emb_vec: a 64-dimensional float vector, with each dimension ranging from −16.0-16.0 to 16.016.0.

  • item_id: the item identifier, which corresponds to the item_id in the interaction file.

  • tag_list: the item tags. For privacy protection, the original tags are anonymized and encoded as integers.

Data Analysis

As show in Table 8 and Table 9, the tag distribution of KuaiRecV2 is extremely long-tailed. Although the dataset contains 11.07M unique tags and 161.52M total tag occurrences, most tags are very sparse. The median frequency is only 1, and both the 25th and 50th percentiles are 1, which means that more than half of the tags appear only once. Even at the 90th percentile, a tag appears only 5 times, showing that the vast majority of tags are low-frequency. In contrast, only 0.97% of tags appear at least 100 times, and only 8 tags appear at least 500,000 times. This highly skewed distribution indicates that KuaiRecV2 provides rich fine-grained semantic signals, while also posing substantial challenges for modeling long-tail and sparse tag information.

25%50%75%90%95%99%
1.001.002.005.0012.0096.00

Table 8. Frequency quantiles of tags in KuaiRecV2.

Frequency ThresholdNumber of TagsProportion
≥1\ge111,068,844100.00%
≥5\ge51,234,84711.16%
≥10\ge10677,3516.12%
≥50\ge50186,9481.69%
≥100\ge100107,3870.97%
≥500\ge50027,8840.25%
≥1,000\ge1{,}00015,1230.14%
≥5,000\ge5{,}0003,3510.03%
≥10,000\ge10{,}0001,6190.01%
≥50,000\ge50{,}0002930.00%
≥100,000\ge100{,}0001200.00%
≥500,000\ge500{,}00080.00%
≥1,000,000\ge1{,}000{,}00020.00%

Table 9. Number and proportion of tags exceeding different frequency thresholds in KuaiRecV2.

SID is a hierarchical discrete representation with an 8-level structure and a codebook size of 512 at each level. The item metadata therefore, provides both discrete semantic codes and continuous embeddings, enabling richer modeling of item semantics and user-item interactions. The distribution of SID is show in Figure 5. Our dataset contains a total of 24,401,875 items and 24,205,418 unique SID sequences, with an SID collision rate of 0.81%.

SID distribution across 8 positions

Figure 5. The distribution of SID, the SID distribution of higher-level codebooks is more uneven.

Detailed results

Comparison of results from different training data

We use MiniOneRec to train on different behavior sequences and predict the long view target on day 21. The results are shown in the table 10, and training on a specific sequence can improve the prediction performance of that sequence.

DataBehaviorNDCG (×10−3\times10^{-3})HR (×10−3\times10^{-3})
@1@5@10@20@50@1@5@10@20@50
clicklong view3.188.0210.7813.9717.433.1812.8921.4934.2051.50
long viewlong view3.508.9011.8915.3319.763.5014.3723.7237.4059.60

Table 10. Comparison of training and predicting long view results using click and long view sequences.

Detailed Analysis of Token-Level Diversity

In this section, we take 50 responses generated by one model for a single sequence as an example. We use the same color to mark identical tokens at this level. We find that when the prefix tokens are the same or similar, the model is more likely to generate the same next token, and may even produce nearly identical SID combinations. Therefore, the token level diversity metric can reflect in more detail that the diversity of the recommended items generated by the model is poor.

1 <a_181><b_183><c_348><d_31><e_26><f_140><g_297><h_67>
2 <a_181><b_183><c_348><d_243><e_448><f_319><g_35><h_383>
3 <a_181><b_183><c_387><d_168><e_251><f_140><g_105><h_336>
4 <a_10><b_85><c_493><d_67><e_1><f_47><g_371><h_428>
5 <a_200><b_115><c_76><d_349><e_269><f_334><g_478><h_119>
6 <a_10><b_158><c_341><d_15><e_20><f_269><g_498><h_498>
7 <a_181><b_198><c_383><d_250><e_258><f_421><g_484><h_429>
8 <a_181><b_183><c_394><d_431><e_251><f_376><g_431><h_375>
9 <a_181><b_224><c_382><d_451><e_368><f_342><g_304><h_116>
10 <a_10><b_400><c_214><d_34><e_56><f_362><g_331><h_42>
11 <a_181><b_183><c_348><d_420><e_7><f_6><g_410><h_69>
12 <a_181><b_183><c_348><d_443><e_13><f_14><g_477><h_75>
13 <a_129><b_394><c_107><d_368><e_254><f_49><g_260><h_54>
14 <a_500><b_228><c_191><d_132><e_96><f_18><g_191><h_455>
15 <a_181><b_198><c_350><d_439><e_25><f_20><g_503><h_376>
16 <a_315><b_179><c_146><d_310><e_96><f_19><g_190><h_181>
17 <a_181><b_480><c_334><d_483><e_132><f_187><g_240><h_471>
18 <a_181><b_198><c_414><d_142><e_343><f_491><g_5><h_39>
19 <a_453><b_291><c_182><d_483><e_396><f_448><g_369><h_120>
20 <a_500><b_343><c_340><d_247><e_481><f_166><g_351><h_424>
21 <a_163><b_168><c_209><d_106><e_60><f_231><g_263><h_50>
22 <a_405><b_291><c_382><d_1><e_212><f_360><g_96><h_276>
23 <a_181><b_183><c_348><d_403><e_311><f_267><g_270><h_494>
24 <a_453><b_291><c_471><d_318><e_98><f_111><g_288><h_315>
25 <a_203><b_291><c_302><d_449><e_437><f_285><g_469><h_286>
26 <a_181><b_226><c_257><d_407><e_287><f_112><g_9><h_193>
27 <a_181><b_88><c_192><d_361><e_15><f_200><g_103><h_62>
28 <a_10><b_158><c_279><d_207><e_478><f_300><g_112><h_391>
29 <a_181><b_183><c_218><d_106><e_419><f_238><g_452><h_384>
30 <a_500><b_228><c_505><d_312><e_196><f_18><g_10><h_414>
31 <a_181><b_183><c_348><d_131><e_225><f_140><g_480><h_169>
32 <a_181><b_224><c_199><d_115><e_228><f_57><g_498><h_331>
33 <a_453><b_291><c_81><d_290><e_230><f_130><g_460><h_276>
34 <a_129><b_291><c_458><d_251><e_363><f_54><g_121><h_450>
35 <a_453><b_291><c_334><d_429><e_355><f_421><g_327><h_315>
36 <a_181><b_205><c_429><d_144><e_395><f_488><g_83><h_393>
37 <a_10><b_291><c_410><d_382><e_493><f_277><g_274><h_473>
38 <a_181><b_224><c_500><d_503><e_507><f_437><g_440><h_450>
39 <a_181><b_183><c_348><d_302><e_379><f_485><g_470><h_383>
40 <a_181><b_183><c_387><d_168><e_251><f_140><g_105><h_138>
41 <a_181><b_183><c_148><d_178><e_53><f_351><g_578><h_80>
42 <a_181><b_183><c_387><d_168><e_251><f_140><g_105><h_439>
43 <a_181><b_183><c_387><d_168><e_251><f_140><g_105><h_124>
44 <a_181><b_183><c_348><d_131><e_446><f_212><g_480><h_227>
45 <a_181><b_183><c_387><d_168><e_251><f_140><g_105><h_253>
46 <a_181><b_183><c_387><d_168><e_251><f_140><g_105><h_254>
47 <a_181><b_183><c_387><d_168><e_251><f_140><g_105><h_48>
48 <a_181><b_183><c_387><d_168><e_251><f_140><g_105><h_128>
49 <a_181><b_183><c_394><d_431><e_251><f_376><g_431><h_138>
50 <a_200><b_115><c_76><d_349><e_269><f_334><g_478><h_48>

Detailed Behavior Metrics Across Streaming Training Days

This section presents comprehensive behavior-specific metrics of MiniOneRec as the example for continual learning of large generative recommenders. We show results in Tables for day 1, day 21, and day 30. For each training day, we report NDCG (Normalized Discounted Cumulative Gain) and HR (Hit Rate) at different cutoff values (@1, @5, @10, @20, @50) for various user behaviors. The behaviors include overall performance as well as specific interactions such as long view, like, follow, forward, comment, search, profile enter, and hate. These detailed results demonstrate how recommendation quality evolves over time for different types of user engagement. All NDCG and HR values are scaled by 10−310^{-3} for readability. The variation in metrics across behaviors reveals the model’s ability to capture diverse user intents, while the progression across rounds illustrates the adaptation dynamics in the streaming recommendation scenario.

BehaviorNDCG (×10−3\times10^{-3})HR (×10−3\times10^{-3})
@1@5@10@20@50@1@5@10@20@50
overall0.342.143.605.638.440.344.138.6816.7330.92
long view0.352.273.856.049.160.354.419.3418.0333.82
like0.662.283.455.298.660.663.977.6114.8932.09
follow0.002.552.555.127.600.006.606.6016.5029.70
forward0.002.462.464.285.290.004.934.9312.3217.24
comment0.942.804.927.7211.210.944.6911.2522.4940.30
hate0.000.000.000.000.000.000.000.000.000.00
search0.000.000.971.743.640.000.003.226.4316.08
profile enter0.323.044.126.228.420.326.129.3417.7128.65

Table 11. Behavior metrics at different cutoffs for streaming training day 1.

BehaviorNDCG (×10−3\times10^{-3})HR (×10−3\times10^{-3})
@1@5@10@20@50@1@5@10@20@50
overall2.877.309.8512.7816.032.8711.7819.7031.4047.60
long view3.188.0210.7813.9717.433.1812.9021.5034.2151.50
like2.487.059.4812.4716.452.4812.0419.4831.5250.99
follow0.007.309.6610.5310.530.0015.3322.9926.8226.82
forward2.645.446.328.309.822.647.9210.5518.4726.39
comment3.745.207.299.3812.723.746.5413.0821.5038.32
hate0.000.000.000.000.000.000.000.000.000.00
search0.003.2610.7114.6316.180.007.9731.8747.8155.78
profile enter2.715.207.7010.2413.092.718.1315.9226.0840.30

Table 12. Behavior metrics at different cutoffs for streaming training day 21.

BehaviorNDCG (×10−3\times10^{-3}) @1NDCG (×10−3\times10^{-3}) @5NDCG (×10−3\times10^{-3}) @10NDCG (×10−3\times10^{-3}) @20NDCG (×10−3\times10^{-3}) @50HR (×10−3\times10^{-3}) @1HR (×10−3\times10^{-3}) @5HR (×10−3\times10^{-3}) @10HR (×10−3\times10^{-3}) @20HR (×10−3\times10^{-3}) @50
overall2.245.918.0910.6713.692.249.7016.4726.7341.83
long view2.456.388.8011.5514.802.4510.4017.9328.8745.11
like0.743.695.457.889.770.746.6412.1721.7631.35
follow0.001.791.792.883.630.004.154.158.3012.45
forward2.754.137.559.6910.822.755.5116.5324.7930.30
comment0.984.286.497.4810.550.987.8214.6618.5734.21
hate0.0019.1219.1219.1225.290.0030.3030.3030.3060.61
search12.2414.2915.5120.3921.2812.2416.3320.4140.8244.90
profile enter2.285.767.469.3611.682.289.4414.6422.1333.52

Table 13. Behavior metrics at different cutoffs for streaming training day 30.

Layer-wise Diversity Metrics Across Streaming Rounds

This section provides layer-wise diversity analysis of MiniOneRec for the click behavior on streaming training day 1, 21, 30. Details of these metrics are provided in Appendix D.

MetricLayer 1Layer 2Layer 3Layer 4Layer 5Layer 6Layer 7Layer 8
Avg Path Diversity11.0232.5846.2349.0949.5849.8949.9450.00
Avg Layer Unique Tokens11.0226.7641.8344.9646.0846.9646.7747.23
Avg Layer Entropy2.474.215.265.435.485.525.515.53
TopK Layer Diversity (Top1)1.001.001.001.001.001.001.001.00
TopK Layer Diversity (Top5)3.324.674.934.964.984.984.974.98
TopK Layer Diversity (Top10)5.238.849.719.829.899.909.879.90
TopK Layer Diversity (Top20)7.5915.7418.9119.2919.4719.5519.4719.59
TopK Layer Diversity (Top50)11.0226.7641.8344.9646.0846.9646.7747.23

Table 14. Diversity metrics across layers for streaming training day 1 (click behavior).

MetricLayer 1Layer 2Layer 3Layer 4Layer 5Layer 6Layer 7Layer 8
Avg Path Diversity12.9034.8546.5049.2649.7349.8749.9550.00
Avg Layer Unique Tokens12.9028.6241.8145.4045.9946.5047.2947.27
Avg Layer Entropy2.754.355.265.455.475.505.535.53
TopK Layer Diversity (Top1)1.001.001.001.001.001.001.001.00
TopK Layer Diversity (Top5)3.594.764.894.964.964.974.984.99
TopK Layer Diversity (Top10)5.899.139.659.869.859.889.909.92
TopK Layer Diversity (Top20)8.9316.6318.8519.4119.4119.5019.5919.63
TopK Layer Diversity (Top50)12.9028.6241.8145.4045.9946.5047.2947.27

Table 15. Diversity metrics across layers for streaming training day 21 (click behavior).

MetricLayer12345678
Avg Path Diversity12.4834.5545.9448.7749.6849.8649.9250.00
Avg Layer Unique Tokens12.4828.5741.1244.9645.6346.6046.9647.02
Avg Layer Entropy2.774.405.225.435.465.505.525.52
TopK Layer Diversity(Top1)1.001.001.001.001.001.001.001.00
TopK Layer Diversity(Top5)3.554.774.914.964.924.984.984.98
TopK Layer Diversity(Top10)5.739.149.639.829.769.899.909.91
TopK Layer Diversity(Top20)8.6216.6118.7119.3019.2519.5119.5619.59
TopK Layer Diversity(Top50)12.4828.5741.1244.9645.6346.6046.9647.02

Table 16. Diversity metrics across layers for streaming training day 30 (click behavior).

References

  1. [1]Charu C. Aggarwal. Recommender Systems - The Textbook. Springer, 2016.DOI
  2. [2]Francesco Ricci, Lior Rokach, and Bracha Shapira. Recommender systems: Techniques, applications, and challenges. In Francesco Ricci, Lior Rokach, and Bracha Shapira, editors, Recommender Systems Handbook, pages 1–35. Springer US, 2022.DOI
  3. [3]Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, and Qing Li. Recommender systems in the era of large language models (llms). IEEE Trans. Knowl. Data Eng., 36(11):6889–6907, 2024.DOI
  4. [4]Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, and Weinan Zhang. How can recommender systems benefit from large language models: A survey. ACM Trans. Inf. Syst., 43(2):28:1–28:47, 2025.arxiv.org/abs/2306.05817
  5. [5]Qing Yu, Xiaobei Wang, Shuchang Liu, chengfeng, Xiaoyu Yang, Xueliang Wang, Chang Meng, Shanshan Wu, Hailang Yang, Bin Wen, Huihui Xiao, Xiang Li, Fan Yang, Xiaoqiang Feng, Lantao Hu, Han Li, Kun Gai, and Lixin Zou. Who you are matters: Bridging interests and social roles via LLM-enhanced logic recommendation. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026.DOI
  6. [6]Qijiong Liu, Jieming Zhu, Lu Fan, Kun Wang, Hengchang Hu, Wei Guo, Yong Liu, and Xiao-Ming Wu. Can LLMs outshine conventional recommenders? a comparative evaluation. In The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2026.
  7. [7]Yashar Deldjoo, Zhankui He, Julian J. McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. A review of modern recommender systems using generative models (gen-recsys). In Ricardo Baeza-Yates and Francesco Bonchi, editors, Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024, pages 6448–6458. ACM, 2024.arxiv.org/abs/2404.00579
  8. [8]Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikai Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. Recommender systems with generative retrieval. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023.arxiv.org/abs/2305.05065
  9. [9]Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations. In Ruslan Salakhutdinov, Zico Kolter, Katherine A. Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024, Proceedings of Machine Learning Research, pages 58484–58509. PMLR / OpenReview.net, 2024.arxiv.org/abs/2402.17152
  10. [10]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, Jinhui 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. Onerec technical report. CoRR, abs/2506.13695, 2025.arxiv.org/abs/2506.13695
  11. [11]Xiaoxi Li, Jiajie Jin, Yujia Zhou, Yuyao Zhang, Peitian Zhang, Yutao Zhu, and Zhicheng Dou. From matching to generation: A survey on generative information retrieval. ACM Trans. Inf. Syst., 43(3):83:1–83:62, 2025.arxiv.org/abs/2404.14851
  12. [12]Yupeng Hou, Jiacheng Li, Xiangjun Fu, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. Bridging language and items for retrieval and recommendation: Benchmarking llms as semantic encoders, 2026.
  13. [13]Shuchang Liu, Qingpeng Cai, Zhankui He, Bowen Sun, Julian J. McAuley, Dong Zheng, Peng Jiang, and Kun Gai. Generative flow network for listwise recommendation. In Ambuj K. Singh, Yizhou Sun, Leman Akoglu, Dimitrios Gunopulos, Xifeng Yan, Ravi Kumar, Fatma Ozcan, and Jieping Ye, editors, Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023, Long Beach, CA, USA, August 6-10, 2023, pages 1524–1534. ACM, 2023.DOI
  14. [14]Yanhua Huang, Yuqi Chen, Xiong Cao, Rui Yang, Mingliang Qi, Yinghao Zhu, Qingchang Han, Yaowei Liu, Zhaoyu Liu, Xuefeng Yao, Yuting Jia, Leilei Ma, Yingqi Zhang, Taoyu Zhu, Liujie Zhang, Lei Chen, Weihang Chen, Min Zhu, Ruiwen Xu, and Lei Zhang. Towards large-scale generative ranking. CoRR, abs/2505.04180, 2025.arxiv.org/abs/2505.04180
  15. [15]Kaike Zhang, Xiaobei Wang, Shuchang Liu, Hailan Yang, Xiang Li, Lantao Hu, Han Li, Qi Cao, Fei Sun, and Kun Gai. Goalrank: Group-relative optimization for a large ranking model. In The Fourteenth International Conference on Learning Representations, 2026.arxiv.org/abs/2509.22046
  16. [16]Pengyue Jia, Xiaobei Wang, Yingyi Zhang, Shuchang Liu, Yupeng Hou, Hailan Yang, Xu Gao, Xiaopeng Li, Yeijing Wang, Julian McAuley, et al. From local indices to global identifiers: Generative reranking for recommender systems via global action space. arXiv preprint arXiv:2604.25291, 2026.arxiv.org/abs/2604.25291
  17. [17]Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. Recommendation as language processing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5). In Jennifer Golbeck, F. Maxwell Harper, Vanessa Murdock, Michael D. Ekstrand, Bracha Shapira, Justin Basilico, Keld T. Lundgaard, and Even Oldridge, editors, RecSys ’22: Sixteenth ACM Conference on Recommender Systems, Seattle, WA, USA, September 18 - 23, 2022, pages 299–315. ACM, 2022.arxiv.org/abs/2203.13366
  18. [18]Wenj​​ie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua. Learnable item tokenization for generative recommendation. In Edoardo Serra and Francesca Spezzano, editors, Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM 2024, Boise, ID, USA, October 21-25, 2024, pages 2400–2409. ACM, 2024.arxiv.org/abs/2405.07314
  19. [19]Xiaopeng Li, Bo Chen, Junda She, Shiteng Cao, You Wang, Qinlin Jia, Haiying He, Zheli Zhou, Zhao Liu, Ji Liu, et al. A survey of generative recommendation from a tri-decoupled perspective: Tokenization, architecture, and optimization. 2025.DOI
  20. [20]Hai Zhu, Yantao Yu, Lei Shen, Bing Wang, and Xiaoyi Zeng. Gatesid: Adaptive gating for semantic-collaborative alignment in cold-start recommendation. CoRR, abs/2603.22916, 2026.
  21. [21]Jiawei Feng, Xiaoyu Kong, Leheng Sheng, Bin Wu, Chao Yi, Feifang Yang, Xiang-Rong Sheng, Han Zhu, Xiang Wang, Jiancan Wu, and Xiangnan He. Fine-grained semantics integration for large language model-based recommendation. CoRR, abs/2602.22632, 2026.arxiv.org/abs/2602.22632
  22. [22]Jiaju Chen, Wang Wenjie, Chongming Gao, Peng Wu, Jianxiong Wei, and Qingsong Hua. Treatment effect estimation for user interest exploration on recommender systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1861–1871, 2024.DOI
  23. [23]Seyed Amir Bidaki, Amir Mohammadkhah, Kiyan Rezaee, Faeze Hassani, Sadegh Eskandari, Maziar Salahi, and Mohammad M. Ghassemi. Online continual learning: A systematic literature review of approaches, challenges, and benchmarks. CoRR, abs/2501.04897, 2025.arxiv.org/abs/2501.04897
  24. [24]Chongming Gao, Shijun Li, Wenqiang Lei, Jiawei Chen, Biao Li, Peng Jiang, Xiangnan He, Jiaxin Mao, and Tat-Seng Chua. Quairec: A fully-observed dataset and insights for evaluating recommender systems. In Mohammad Al Hasan and Li Xiong, editors, Proceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022, pages 540–550. ACM, 2022.DOI
  25. [25]Kairui Fu, Tao Zhang, Shuwen Xiao, Ziyang Wang, Xinming Zhang, Chenchi Zhang, Yuliang Yan, Junjun Zheng, Yu Li, Zhihong Chen, Jian Wu, Xiangheng Kong, Shengyu Zhang, Kun Kuang, Yu-Gang Jiang, and Bo Zheng. FORGE: forming semantic identifiers for generative retrieval in industrial datasets. CoRR, abs/2509.20904, 2025.arxiv.org/abs/2509.20904
  26. [26]Everton Lima Aleixo, Juan Gabriel Colonna, Marco Cristo, and Everlandio Fernandes. Catastrophic forgetting in deep learning: A comprehensive taxonomy. J. Braz. Comput. Soc., 30(1), 2024.arxiv.org/abs/2312.10549
  27. [27]Ke Wan, Yi Liang, and Susik Yoon. Online drift detection with maximum concept discrepancy. In Ricardo Baeza-Yates and Francesco Bonchi, editors, Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024, pages 2924–2935. ACM, 2024.arxiv.org/abs/2407.05375
  28. [28]Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 11816–11825, 2019.arxiv.org/abs/1903.08671
  29. [29]Jaime G. Carbonell and Jade Goldstein. The use of mmr, diversity-based reranking for re-ordering documents and producing summaries. In W. Bruce Croft, Alistair Moffat, C. J. van Rijsbergen, Ross Wilkinson, and Justin Zobel, editors, SIGIR ’98: Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, August 24-28 1998, Melbourne, Australia, pages 335–336. ACM, 1998.
  30. [30]Cai-Nicolas Ziegler, Sean M. McNee, Joseph A. Konstan, and Georg Lausen. Improving recommendation lists through topic diversification. In Allan Ellis and Tatsuy​​a Hagino, editors, Proceedings of the 14th international conference on World Wide Web, WWW 2005, Chiba, Japan, May 10-14, 2005, pages 22–32. ACM, 2005.DOI
  31. [31]Yuying Zhao, Yu Wang, Yunchao Liu, Xueqi Cheng, Charu C. Aggarwal, and Tyler Derr. Fairness and diversity in recommender systems: A survey. ACM Trans. Intell. Syst. Technol., 16(1):2:1–2:28, 2025.arxiv.org/abs/2307.04644
  32. [32]Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H. Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Yike Guo and Faisal Farooq, editors, Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD 2018, London, UK, August 19-23, 2018, pages 1930–1939. ACM, 2018.DOI
  33. [33]Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. Progressive layered extraction (PLE): A novel multi-task learning (MTL) model for personalized recommendations. In Rodrigo L. T. Santos, Leandro Balby Marinho, Elizabeth M. Daly, Li Chen, Kim Falk, Noam Koenigstein, and Edlenc Silva de Moura, editors, RecSys 2020: Fourteenth ACM Conference on Recommender Systems, Virtual Event, Brazil, September 22-26, 2020, pages 269–278. ACM, 2020.
  34. [34]Yuhao Wang, Ha Tsz Lam, Yi Wong, Ziru Liu, Xiangyu Zhao, Yichao Wang, Bo Chen, Huifeng Guo, and Ruiming Tang. Multi-task deep recommender systems: A survey. CoRR, abs/2302.03525, 2023.arxiv.org/abs/2302.03525
  35. [35]Zijian Zhang, Shuchang Liu, Jiaao Yu, Qingpeng Cai, Xiangyu Zhao, Chunxu Zhang, Ziru Liu, Qidong Liu, Hongwei Zhao, Lantao Hu, Peng Jiang, and Kun Gai. M³oe: Multi-domain multi-task mixture-of-experts recommendation framework. In Grace Hui Yang, Honning Wang, Sam Han, Claudia Hauff, Guido Zuccon, and Yi Zhang, editors, Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, Washington DC, USA, July 14-18, 2024, pages 893–902. ACM, 2024.DOI
  36. [36]Jiaj u Chen, Chongming Gao, Shuai Yuan, Shuchang Liu, Qingpeng Cai, and Peng Jiang. Dlcrec: A novel approach for managing diversity in llm-based recommender systems. In Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining, pages 857–865, 2025.arxiv.org/abs/2408.12470
  37. [37]Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, Yifei Hu, Qigen Hu, Xinchen Luo, Lejian Ren, Zixing Zhang, Qianqian Wang, Kuo Cai, Yunfan Wu, Hongtao Cheng, Zexuan Cheng, Lu Ren, Huanjie Wang, Yi Su, Ruiming Tang, Kun Gai, and Guorui Zhou. Onerec-think: In-text reasoning for generative recommendation. CoRR, abs/2510.11639, 2025.arxiv.org/abs/2510.11639
  38. [38]Yang Yu, Lei Kou, Huaikuan Yi, Bin Chen, Yayu Cao, Lei Shen, Chao Zhang, Bing Wang, and Xiaoyi Zeng. A semantic-grounded instruction-driven generative multi-task recommender at aliexpress. CoRR, abs/2602.22913, 2026.arxiv.org/abs/2602.22913
  39. [39]Xiaofeng Liu, Wenliang Li, Tao Yu, Guanliang Song, Zhen Chen, Jing Zhou, Dongyue Wang, Xiwei Zhao, and Sulong Xu. From modularity to unity: Towards industrial-scale generative recommendation. In Hakim Hacid, Yoelle Maarek, Francesco Bonchi, Ido Guy, and Emine Yilmaz, editors, Proceedings of the ACM Web Conference 2026, WWW 2026, Dubai, United Arab Emirates, originally scheduled for April 13-17, 2026, rescheduled for June 29 - July 3, 2026, pages 8183–8190. ACM, 2026.
  40. [40]Hongxun Ding, Keqin Bao, Jizhi Zhang, Yi Fang, Wenxin Xu, Fuli Feng, and Xiangnan He. Towards sample-efficient and stable reinforcement learning for llm-based recommendation. CoRR, abs/2602.00632, 2026.arxiv.org/abs/2602.00632
  41. [41]Ruining He and Julian J. McAuley. VBPR: visual bayesian personalized ranking from implicit feedback. In Dale Schuurmans and Michael P. Wellman, editors, Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, February 12-17, 2016, Phoenix, Arizona, USA, pages 144–150. AAAI Press, 2016.arxiv.org/abs/1510.01784
  42. [42]Da Li, Yuxiao Luo, Keping Bi, Jiafeng Guo, Wei Yuan, Biao Yang, Yan Wang, Fan Yang, Tingting Gao, and Guorui Zhou. Compressing then matching: An efficient pre-training paradigm for multimodal embedding. In Maria Liakata, Viviane P. Moreira, Jiajun Zhang, and David Jurgens, editors, Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3707–3718, San Diego, California, United States, July 2026. Association for Computational Linguistics.
  43. [43]Jinke Wu, Yinghao Wu, Shuchang Liu, Yandong Bai, Qi Zhu, Huihui Xiao, and Fan Yang. GUIDER: generative user interest discovery via explicit reasoning with large language models. In Won Kim, Jae-Gil Lee, Kyuseok Shim, James Caverlee, David F. Gleich, Xiangliang Zhang, Qi He, Faisal Farooq, Rómer Rosales, Sayan Ranu, Haixun Wang, Yan Liu, Carl Yang, Leman Akoglu, Nitesh V. Chawla, and Vagelis Papalexakis, editors, Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V2, KDD 2026, Jeju Island, Korea, August 9-13, 2026, pages 8246–8257. ACM, 2026.DOI
  44. [44]Wang-Cheng Kang and Julian J. McAuley. Self-attentive sequential recommendation. In IEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018, pages 197–206. IEEE Computer Society, 2018.
  45. [45]Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. Session-based recommendations with recurrent neural networks. In Yoshua Bengio and Yann LeCun, editors, 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016.arxiv.org/abs/1511.06939
  46. [46]Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiomoorthy. Recommender systems with generative retrieval. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023.arxiv.org/abs/2305.05065
  47. [47]Teng Shi, Chenglei Shen, Weijie Yu, Shen Nie, Chongxuan Li, Xiao Zhang, Ming He, Yan Han, and Jun Xu. Llada-rec: Discrete diffusion for parallel semantic ID generation in generative recommendation. CoRR, abs/2511.06254, 2025.arxiv.org/abs/2511.06254
  48. [48]Xiaoyu Kong, Leheng Sheng, Junfei Tan, Yuxin Chen, Jiancan Wu, An Zhang, Xiang Wang, and Xiangnan He. Minionrec: An open-source framework for scaling generative recommendation. CoRR, abs/2510.24431, 2025.arxiv.org/abs/2510.24431
  49. [49]Bin Wu, Feifan Yang, Zhangming Chan, Yu-Ran Gu, Jiawei Feng, Chao Yi, Xiang-Rong Sheng, Han Zhu, Jian Xu, Mang Ye, and Bo Zheng. MUSE: A simple yet effective multimodal search-based framework for lifelong user interest modeling. CoRR, abs/2512.07216, 2025.arxiv.org/abs/2512.07216
  50. [50]Jaime Carbonell and Jade Goldstein. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’98, page 335–336, New York, NY, USA, 1998. Association for Computing Machinery.
  51. [51]Cai-Nicolas Ziegler, Sean M. McNee, Joseph A. Konstan, and Georg Lausen. Improving recommendation lists through topic diversification. In Proceedings of the 14th International Conference on World Wide Web, WWW ’05, page 22–32, New York, NY, USA, 2005. Association for Computing Machinery.DOI
  52. [52]Charles L.A. Clarke, Maheedhar Kolla, Gordon V. Cormack, Olga Vechtomova, Azin Ashkan, Stefan Büttcher, and Ian MacKinnon. Novelty and diversity in information retrieval evaluation. In Proceedings of the 31st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’08, page 659–666, New York, NY, USA, 2008. Association for Computing Machinery.DOI
  53. [53]Laming Chen, Guoxin Zhang, and Hanning Zhou. Fast greedy map inference for determinantal point processes to improve recommendation diversity. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS’18, page 5627–5638, Red Hook, NY, USA, 2018. Curran Associates Inc.arxiv.org/abs/1709.05135
  54. [54]Mike Gartrell, Ulrich Paquet, and Noam Koenigstein. Bayesian low-rank determinantal point processes. In Proceedings of the 10th ACM Conference on Recommender Systems, RecSys ’16, page 349–356, New York, NY, USA, 2016. Association for Computing Machinery.DOI
  55. [55]Nathaniel Good, J. Ben Schafer, Joseph A. Konstan, Al Borchers, Badrul Sarwar, Jon Herlocker, and John Riedl. Combining collaborative filtering with personal agents for better recommendations. In Proceedings of the Sixteenth National Conference on Artificial Intelligence and the Eleventh Innovative Applications of Artificial Intelligence Conference Innovative Applications of Artificial Intelligence, AAAI ’99/IAAI ’99, page 439–446, USA, 1999. American Association for Artificial Intelligence.
  56. [56]Lijing Qin and Xiaoyan Zhu. Promoting diversity in recommendation by entropy regularizer. In Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence, IJCAI ’13, page 2698–2704. AAAI Press, 2013.
  57. [57]Yi Ren, Xiao Han, Xu Zhao, Shenzheng Zhang, and Yan Zhang. Slate-aware ranking for recommendation. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, WSDM ’23, page 499–507, New York, NY, USA, 2023. Association for Computing Machinery.arxiv.org/abs/2302.12427
  58. [58]Shuchang Liu, Fei Sun, Yingqiang Ge, Changhua Pei, and Yongfeng Zhang. Variation control and evaluation for generative slate recommendations. In Proceedings of the Web Conference 2021, WWW ’21, page 436–448, New York, NY, USA, 2021. Association for Computing Machinery.DOI
  59. [59]Jun Wang, Lantao Yu, Weinan Zhang, Yu Gong, Yinghui Xu, Benyou Wang, Peng Zhang, and Dell Zhang. IRGAN: A minimax game for unifying generative and discriminative information retrieval models. In Noriko Kando, Tetsuya Sakai, Hideo Joho, Hang Li, Arjen P. de Vries, and Ryen W. White, editors, Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, Shinjuku, Tokyo, Japan, August 7-11, 2017, pages 515–524. ACM, 2017.
  60. [60]Dawen Liang, Rahul G. Krishnan, Matthew D. Hoffman, and Tony Jebara. Variational autoencoders for collaborative filtering. In Pierre-Antoine Champin, Fabien Gandon, Mounia Lalmas, and Panagiotis G. Ipeirotis, editors, Proceedings of the 2018 World Wide Web Conference on World Wide Web, WWW 2018, Lyon, France, April 23-27, 2018, pages 689–698. ACM, 2018.arxiv.org/abs/1802.05814
  61. [61]Irwan Bello, Sayali Kulkarni, Sagar Jain, Craig Boutilier, Ed Huai-hsin Chi, Elad Eban, Xiyang Luo, Alan Mackey, and Ofer Meshi. Seq2slate: Re-ranking and slate optimization with rnns. CoRR, abs/1810.02019, 2018.arxiv.org/abs/1810.02019
  62. [62]Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In Yoshua Bengio and Yann LeCun, editors, 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014.arxiv.org/abs/1312.6114
  63. [63]Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Zoubin Ghahramani, Max Welling, Corinna Cortes, Neil D. Lawrence, and Kilian Q. Weinberger, editors, Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 2672–2680, 2014.
  64. [64]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998–6008, 2017.arxiv.org/abs/1706.03762
  65. [65]Hao Guo, Erpeng Xue, Lei Huang, Shichao Wang, Xiaolei Wang, Lei Wang, Jinpeng Wang, Zeshun Li, and Sheng Chen. Twin-flow generative ranking network for recommendation. In Meeyoung Cha, Chanyoung Park, Noseong Park, Carl Yang, Senjuti Basu Roy, Jessie Li, Jaap Kamps, Kijung Shin, Bryan Hooi, and Lifang He, editors, Proceedings of the 34th ACM International Conference on Information and Knowledge Management, CIKM 2025, Seoul, Republic of Korea, November 10-14, 2025, pages 771–780. ACM, 2025.
  66. [66]Hongyu Yao, Zijin Hong, Hao Chen, Zhiqing Li, Qijie Shen, Zuobin Ying, Qihua Feng, Huan Gong, and Feiran Huang. Macro graph of experts for billion-scale multi-task recommendation. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1, KDD ’26, page 2506–2517, New York, NY, USA, 2026. Association for Computing Machinery.arxiv.org/abs/2506.10520
  67. [67]Jingtong Gao, Bo Chen, Xiangyu Zhao, Weiwen Liu, Xiangyang Li, Yichao Wang, Wanyu Wang, Huifeng Guo, and Ruiming Tang. Llm4rerank: Llm-based auto-reranking framework for recommendations. In Guodong Long, Michale Blumestein, Yi Chang, Liane Lewin-Eytan, Zi Helen Huang, and Elad Yom-Tov, editors, Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia, 28 April 2025- 2 May 2025, pages 228–239. ACM, 2025.arxiv.org/abs/2406.12433
  68. [68]Shuo Liu, An Zhang, Guoqing Hu, Hong Qian, and Tat-Seng Chua. Preference diffusion for recommendation. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025.arxiv.org/abs/2410.13117
  69. [69]Guoqing Hu, Zhengyi Yang, Zhibo Cai, An Zhang, and Xiang Wang. Generate and instantiate what you prefer: Text-guided diffusion for sequential recommendation. CoRR, abs/2410.13428, 2024.arxiv.org/abs/2410.13428
  70. [70]Chao Yi, Feifan Yang, Jiawei Feng, Sishuo Chen, Zhangming Chan, Xiang-Rong Sheng, and Han Zhu. Native multimodal representation learning for click-through rate prediction in e-commerce scenarios. CoRR, abs/2608.024091, 2026.
  71. [71]Shuchang Liu, Fei Sun, Yingqiang Ge, Changhua Pei, and Yongfeng Zhang. Variation control and evaluation for generative slate recommendations. In Jure Leskovec, Marko Grobelnik, Marc Najork, Jie Tang, and Leila Zia, editors, WWW ’21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021, pages 436–448. ACM / IW3C2, 2021.DOI
  72. [72]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020.arxiv.org/abs/2006.11239
  73. [73]Zihao Li, Aixin Sun, and Chenliang Li. Diffurec: A diffusion model for sequential recommendation. ACM Trans. Inf. Syst., 42(3):66:1–66:28, 2024.arxiv.org/abs/2304.00686
  74. [74]Wenyu Mao, Shuchang Liu, Hailan Yang, Xiaobei Wang, Xiaoyu Yang, Xu Gao, Xiang Li, Lantao Hu, Han Li, Kun Gai, An Zhang, and Xiang Wang. Denoising neural reranker for recommender systems. In The Fourteenth International Conference on Learning Representations, 2026.arxiv.org/abs/2509.18736
  75. [75]Guoqing Hu, An Zhang, Shuchang Liu, Wenyu Mao, Jiancan Wu, Xun Yang, Xiang Li, Lantao Hu, Han Li, Kun Gai, and Xiang Wang. Fading to grow: Growing preference ratios via preference fading discrete diffusion for recommendation. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026.arxiv.org/abs/2509.26063
  76. [76]Xiao Lin, Xiaokai Chen, Chenyang Wang, Hantao Shu, Linfeng Song, Biao Li, and Peng Jiang. Discrete conditional diffusion for reranking in recommendation. In Tat-Seng Chua, Chong-Wah Ngo, Roy Ka-Wei Lee, Ravi Kumar, and Hady W. Lauw, editors, Companion Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, Singapore, May 13-17, 2024, pages 161–169. ACM, 2024.arxiv.org/abs/2308.06982
  77. [77]Jiaju Chen, Chongming Gao, Chenxiao Fan, Haoyan Liu, Qipeng Cai, Peng Jiang, and Xiangnan He. Position-aware drafting for inference acceleration in llm-based generate list-wise recommendation. arXiv preprint arXiv:2604.27747, 2026.arxiv.org/abs/2604.27747
  78. [78]Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Trans. Pattern Anal. Mach. Intell., 46(8):5362–5383, 2024.DOI
  79. [79]Xu Huang, Jiancan Wu, Hongzheng Qin, Jiawei Feng, Can Ye, Wencong Zeng, Fengze Li, and Xiang Wang. Generative recommenders forget the long tail: Restoring memorization in retrieval and ranking, 2026.arxiv.org/abs/2610.29298
  80. [80]Hongyi Wan, Shiyuan Ren, Wei Huang, Miao Zhang, Xiang Deng, Yixin Bao, and Lijiang Nie. Understanding the forgetting of (Replay-based) continual learning via feature learning: Angle matters. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors, Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 61956–62019. PMLR, 13–19 Jul 2025.
  81. [81]Gyuseok Lee, Hyunsik Yoo, Junyoung Hwang, SeongKu Kang, and Hwanjo Yu. Leveraging historical and current interests for continual sequential recommendation. CoRR, abs/2506.07466, 2025.
  82. [82]Hyunsik Yoo, Ting-Wei Li, SeongKu Kang, Zhining Liu, Charlie Xu, Qilin Qi, and Hanghang Tong. Continual low-rank adapters for llm-based generative recommender systems. CoRR, abs/2510.25093, 2025.arxiv.org/abs/2510.25093

Paper details

vlnc:W_ge4jjmjq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents