WorksComputer Science
HiGenTrans: Hierarchical LLM Knowledge Transfer for Generative Recommendation
Abstract
Generative recommendation promises to simplify conventional multi-stage recommender systems by generating item Semantic IDs (SIDs) autoregressively. However, two obstacles limit its effective- ness in industrial settings: interaction-only supervision underuses the semantic knowledge encoded in pretrained language models, while beam-search likelihood ranks token sequences rather than directly estimating item relevance. To address these limitations, we propose HiGenTrans, a three-level framework that introduces LLM knowledge at the input, representation, and decision levels. First, Global Semantic Priming (GSP) adds an LLM-derived user representation to the beginning-of-sequence embedding, provid- ing a semantic prior before SID generation. Second, Hierarchi- cal Representation Transfer (HRT) aligns intermediate LLM representations with depth-matched GenRec layers through mean- centered projections and gated residual connections. Finally, Multi- Signal Decision Calibration (MDC) uses a lightweight mixture- of-experts calibrator to combine semantic relevance, generation confidence, and candidate-set context when reranking beam-search outputs. All semantic features are computed only from the observ- able interaction prefix, preventing target leakage, and can be cached offline. Experiments on three public benchmarks show that HiGen- Trans improves Recall@10 by up to 12.9% and NDCG@10 by up to 33.7% over the strongest baselines. These results demonstrate the value of transferring LLM knowledge at complementary stages of the generative recommendation pipeline.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected].
Conference acronym ’XX, Woodstock, NY
© 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-xxxx-xxxx-x/YYYY/MM https://doi.org/XXXXXXX.XXXXXXX
Abstract
CCS Concepts
• Information systems → Recommender systems.
Keywords
Generative Recommendation, Large Language Models, Knowledge Transfer, Sequential Recommendation.
ACM Reference Format
Zhuohang Jiang, Yuxin Chen, Haohao Qu, Zefei Luo, Yunlong Wang, Xu Yuan, Haojie Ding, Boheng Zhang, Fan Yan, Hongyang Wang, Jun Wang, Zhongbo Sun, Wenqi Fan, Qing Li, Wenwu Ou, and Peng Jiang. 2018. HiGenTrans: Hierarchical LLM Knowledge Transfer for Generative Recommendation. In *Proceedings of Make sure to enter the correct conference title from your rights confirmation email (Conference acronym ’XX). ACM, New York, NY, USA, 10 pages. https://doi.org/XXXXXXX.XXXXXXX
Introduction
Recommender systems are central to Web applications such as e-commerce, job search, and social media, where they help users navigate rapidly growing item spaces [6, 7, 18]. Conventional industrial systems commonly organize retrieval, ranking, and reranking as separately optimized stages Conventional industrial systems commonly organize multi-stage discriminative architectures [2, 27], where candidate items are progressively filtered through retrieval [40], ranking [35], and reranking [14] stages. Although this paradigm has achieved widespread success in large-scale industrial recommender systems [24], different stages often rely on relatively independent modeling objectives and optimization strategies. Such a multi-stage architecture increases system design, training, and maintenance complexity, while making it challenging to jointly optimize user preference modeling and final recommendation decisions.
In recent years, generative recommendation (GenRec) [5, 26, 40] has emerged as a promising alternative paradigm. Unlike traditional discriminative approaches, GenRec reformulates recommendation as an autoregressive sequence generation task, typically encoding items into discrete Semantic IDs (SIDs) [11, 23, 25] and employing Transformer-based models to generate target items token by token based on users’ historical interaction sequences. Representative studies, such as TIGER [28] and OneRec [5], have demonstrated the potential of generative modeling for sequential recommendation, opening up new directions for building more unified and scalable recommendation models.
Despite the progress of GenRec in advancing recommendation paradigms, as illustrated in Figure 1, they also encounter several limitations. The first limitation arises from how GenRec uses semantic knowledge [39]. Interaction supervision is effective for learning collaborative regularities, but it can be sparse and does not directly expose the textual relations among items or the high-level intent underlying a history. Pretrained LLMs provide complementary semantic and contextual knowledge [4, 16, 17, 22, 31, 38]; accordingly, recent LLM-enhanced recommenders derive interest descriptions or semantic features from item metadata [9, 12, 20, 42]. Most fusion strategies, however, compress this information into one or a few input embeddings, such as prompts, prefixes, or beginning-of-sequence (BOS) tokens [1]. Because representations at different Transformer depths encode complementary abstractions [3, 13, 15, 30], a single input-level injection leaves much of the LLM hierarchy unused and requires the recommendation backbone to preserve one semantic vector through all subsequent layers. The second limitation concerns the decision rule. Beam search orders SID sequences by autoregressive likelihood, which is affected by tokenization, prefix probabilities, and decoding dynamics. This sequence-level likelihood need not equal the relevance of the decoded item to the user [32]. A relevant item can therefore appear in the candidate set but occupy a poor rank. Correcting this mismatch requires candidate-level evidence beyond the raw generation score, including semantic relevance, token-level confidence, and each candidate’s position relative to the rest of the beam. These two limitations suggest that LLM knowledge should enter GenRec not only once at the input, but also within its representation hierarchy and at its final decision stage.
Figure 1. Motivation and overview of HiGenTrans. Existing GenRec and LLM4Rec methods suffer from shallow semantic fusion and misalignment between generation likelihood and item relevance. HiGenTrans addresses these limitations through Global Semantic Prior (GSP), Hierarchical Representation Transfer (HRT), and Multi-signal Decision Calibration (MDC), enabling richer semantic integration and more accurate recommendation ranking.
We therefore propose HiGenTrans, a Hierarchical LLM-to-GenRec Knowledge-Transfer framework, which integrates three interactive modules, namely Global Semantic Priming (GSP) for prior knowledge encoding, Hierarchical Representation Transfer (HRT) for layer-wise representation adaptation, and Multi-Signal Decision Calibration (MDC) for efficient feature mixing. Specifically, the GSP module projects a frozen LLM representation of the observable user-history prefix into the BOS embedding, giving the generator a global semantic prior. Moreover, HRT then mean-centers, projects, and injects intermediate LLM representations into depth-matched GenRec layers through learned residual gates. GSP and HRT preserve the collaborative backbone while exposing it to semantic information both before and during SID generation. Finally, MDC applies a lightweight mixture-of-experts calibrator to semantic, generative, and candidate-context features and reranks only the beam-generated items. Thus, GSP initializes what the model should attend to, HRT enriches how the history is represented, and MDC calibrates which generated candidates should be recommended. All LLM representations are computed from the prefix available at the corresponding prediction step, which prevents target leakage, and the frozen encoder permits offline feature caching.
Considering the stringent storage and inference efficiency requirements of industrial recommender systems, we develop an efficient deployment strategy for large-scale deployment. Since HRT relies on semantic representations extracted from multiple intermediate LLM layers, directly storing and transmitting these high-dimensional features would introduce substantial system overhead. To address this issue, we introduce a vector quantization-based feature compression mechanism that quantizes the intermediate LLM representations at different layers into compact Semantic IDs and reconstructs the corresponding features through a shared codebook. By performing LLM feature extraction and quantization encoding entirely offline, the online inference stage only requires lightweight feature reconstruction without invoking the LLM.
We conduct extensive experiments on three public recommendation datasets, namely Beauty, Sports, and Toys, to comprehensively evaluate the effectiveness of HiGenTrans. Experimental results demonstrate that HiGenTrans consistently outperforms existing baseline methods across all evaluation metrics on all three datasets, achieving substantial performance improvements. Specifically, HiGenTrans improves Recall@10 by 2.69%, 5.69%, and 12.94% on the Beauty, Sports, and Toys datasets, respectively. The ablations show complementary gains from semantic transfer (GSP and HRT together) and decision calibration (MDC), while experiments in two industrial domains support the applicability of hierarchical semantic fusion in production settings.
Our main contributions are summarized as follows:
We propose HiGenTrans, a three-level framework that transfers LLM knowledge at the input, internal representation, and decision stages of generative recommendation.
We introduce Global Semantic Priming (GSP) and Hierarchical Representation Transfer (HRT), which combine prefix-safe global initialization with mean-centered, gated fusion of depth-matched LLM representations.
We develop Multi-Signal Decision Calibration (MDC), a lightweight candidate-level calibrator that uses a mixture of experts to integrate semantic, generative, and contextual evidence.
Experiments on three public benchmarks and two industrial domains demonstrate consistent improvements, while ablation and segment-level analyses characterize the contributions of semantic transfer and decision calibration.
Related Works
Generative Recommendation
Generative recommendation (GenRec) casts item retrieval as conditional sequence generation, replacing a catalog-wide scoring operation with autoregressive decoding over discrete item representations [28]. TIGER [28] pioneered this paradigm by representing items with hierarchical Semantic IDs (SIDs), allowing recommendation to be performed through discrete token generation. Subsequent work has mainly advanced GenRec along three directions. First, tokenization methods improve the correspondence between discrete codes and recommendation structure: LETTER [33] jointly models semantic, collaborative, and code-diversity objectives, TokenRec [25] learns ID tokenization for LLM-based generation, and qualification-aware or adaptive collision handling makes SIDs more suitable for large catalogs [11, 23]. Second, model architectures strengthen the interaction between behavioral and semantic signals [36]. EAGER [34] uses two streams for behavior-semantic collaboration, RPG [10] generates long SIDs in parallel, and GRLM [43] uses structured term identifiers to expose the recommendation capability of LLM-based generators. Third, industrial systems seek unified or scalable generation, as illustrated by OneRec [5], One-Mall [40], and OneLive [32]. These studies improve what is generated or how a generator is trained, but they do not fully address where external semantic knowledge should enter the recommendation pipeline. Most semantic information is encoded into item tokens, fused at the input, or distilled into the generator. In contrast,
HiGenTrans keeps the SID-generation backbone while transferring prefix-conditioned LLM knowledge at three distinct locations: GSP supplies a global prior before decoding, HRT injects depth-specific features inside the backbone, and MDC calibrates the ordering of the resulting candidate set. This design separates semantic augmentation of candidate generation from candidate-level decision calibration.
LLM-Enhanced Recommendation
LLMs have been incorporated into recommender systems as semantic encoders, reasoning modules, or recommendation backbones [19, 26, 44]. One line of work adapts an LLM to collaborative signals. CLLM4Rec [46] extends the LLM vocabulary with user and item tokens, while CLLM [41] connects collaborative representations or learned identifiers to language models. Another line transfers language-model knowledge to comparatively lightweight recommenders. LLM-ESR [21] enhances ID-based sequential recommendation with LLM-derived semantics, IDLE-Adapter [37] aligns representations and output distributions, and distribution-matching approaches explicitly bridge collaborative and language spaces [42]. IMFuse [45] further shows that intermediate LLM layers contain complementary information that can be aggregated for sequential recommendation. Our setting differs in both transfer direction and granularity. Rather than turning the LLM itself into the online recommender, HiGenTrans uses a frozen LLM as an offline semantic feature provider for a generative recommendation backbone. Moreover, it does not collapse all intermediate layers into one representation. Motivated by evidence that Transformer layers encode different levels of abstraction [3, 13, 30], HRT preserves the ordering of selected LLM layers and maps them monotonically to GenRec depth. Mean-centered projections address the mismatch between the two representation spaces, and learned residual gates let recommendation supervision determine the strength of each transfer. The resulting model therefore complements input-only, final-layer, and layer-aggregation approaches with explicit depth-to-depth transfer.
Methodology
Framework Overview
Existing generative recommendation models predict item sequences autoregressively using Semantic IDs, but often underutilize the hierarchical semantic knowledge of pretrained LLMs and rely on generative likelihoods that may not align with item-level ranking relevance. To address these limitations, we propose HiGenTrans, a framework that transfers LLM semantic knowledge into generative recommendation at multiple levels, as illustrated in Figure 2.
Figure 2. Overview of HiGenTrans, a three-level semantic knowledge transfer framework for generative recommendation. Global Semantic Priming (GSP) incorporates global user semantics into the BOS token. Hierarchical Representation Transfer (HRT) injects multi-level LLM representations into the generative backbone. Multi-Signal Decision Calibration (MDC) employs MoE-based multi-signal calibration to refine candidate rankings.
Notation. Let denote the observable interaction history of user up to step , and let be the target item. Each item is represented by a sequence of Semantic ID (SID) tokens , where . We use to denote the collection of next-item training instances. A generative recommender forecasts the next-item probability as
Our framework enhances this generative process through three complementary components. Global Semantic Priming (GSP) incorporates global user semantics to guide the generation process. Hierarchical Representation Transfer (HRT) injects intermediate LLM representations into depth-matched layers of the generative backbone, enabling hierarchical semantic knowledge transfer. Multi-Signal Decision Calibration (MDC) integrates semantic, generative, and contextual signals to refine candidate rankings. The pretrained LLM remains frozen and is used only for prefix-safe feature precomputation, while the generative backbone and fusion parameters are jointly optimized. Finally, MDC is trained on candidates produced by the trained generator to improve item-level ranking.
Global Semantic Priming
GenRec processes discrete interaction tokens, which preserve collaborative structure but may omit the high-level intent expressed by item titles and categories. The Global Semantic Priming (GSP) module supplies this information before autoregressive SID generation begins. For each training instance , we serialize only the observable prefix with available item metadata. A frozen semantic encoder produces a global prefix representation
where denotes the frozen encoder together with its fixed sequence-pooling operation, and is its output dimension. We project this vector into the -dimensional GenRec space and add it to the beginning-of-sequence representation:
Thus, the BOS token becomes a semantic prior conditioned on the same history available to the recommender. Computing from rather than the complete sequence is essential: using or later interactions would leak the target into the input. Because the semantic encoder is frozen, all prefix representations can be computed and cached before GenRec training.
Hierarchical Representation Transfer
GSP provides the generator with a global semantic prior,but a single input-level representation cannot fully preserve the complementary abstractions encoded at different LLM depths [3, 13, 30]. Moreover, its influence may weaken as it propagates through the generative backbone. We therefore propose Hierarchical Representation Transfer (HRT), which transfers depth-specific semantic knowledge directly into the backbone’s internal representations. HRT contains three components: (i) multi-level semantic extraction, which obtains prefix-conditioned features from representative LLM layers; (ii) representation adaption, which mean-centers, projects, and normalizes each feature into the GenRec space; and (iii) gated layer-wise fusion, which injects the adapted features into depth-matched backbone layers through residual connections. The resulting transfer parameters and the GenRec backbone are then trained jointly using the autoregressive recommendation objective. The following subsections describe these components and the joint optimization procedure in detail.
Multi-level semantic extraction
Let be the selected LLM layers. Generally, pretrained LLMs encode semantic information at different levels of abstraction across their Transformer layers [13, 30]. To exploit this hierarchical knowledge, we extract hidden representations from a set of representative layers of the frozen semantic encoder. Specifically, for each observable prefix, we extract
where applies the same fixed pooling operation to the token representations returned by LLM layer . In our implementation, we use Qwen3-Embedding-8B as the frozen encoder and choose four layers distributed across its depth. An order-preserving injective mapping associates each selected LLM layer with a GenRec layer of similar relative depth, where is the number of backbone layers. Thus, implies , preventing multiple semantic features from being assigned ambiguously to the same backbone layer.
Representation adaptation
The LLM and GenRec feature spaces differ in both dimension and distribution [42]. Moreover, LLM embeddings can be anisotropic, so a direction shared by most histories may dominate direct residual injection.
To mitigate these issues, we introduce a lightweight representation adaptation mechanism that combines mean centering, linear projection, and normalization. For each selected layer, we estimate the training-set mean using only training prefixes:
The centered feature is projected and normalized as
where denotes the mean semantic representation estimated from the training data at layer . Mean subtraction is performed in FP32 to avoid cancellation errors under mixed-precision training. Centering removes shared components, the learned matrix aligns the two feature spaces, and RMS normalization controls the scale of the injected vector.
Gated layer-wise fusion
Let be the hidden state at SID decoding step immediately before GenRec layer . HRT applies the same prefix-conditioned semantic feature at every decoding step through
where is a trainable layer-specific scalar gate. The residual path preserves the backbone’s collaborative representation, while depth-matched injection lets earlier semantic features undergo further transformation and supplies later features closer to the output.
Joint generative optimization
Let collect the global and layer-wise semantic features used by GSP and HRT. The GenRec backbone and the transfer parameters are jointly optimized with
Here, collects the trainable GenRec, projection, and fusion parameters; the LLM remains frozen. We use no auxiliary representation-alignment loss, so recommendation supervision determines how the transferred features are adapted.
Multi-Signal Decision Calibration
Although GSP and HRT improve candidate generation, beam-search scores measure the conditional likelihood of SID sequences rather than item-level recommendation relevance [5, 28]. Consequently, candidates with similar generation scores may differ substantially in semantic relevance and relative quality within the beam. We therefore propose Multi-Signal Decision Calibration (MDC), which calibrates the generator’s outputs at the decision level. MDC contains three components: (i) multi-signal candidate representation, which combines LLM semantic evidence, GenRec confidence statistics, and candidate-set context; (ii) nonlinear signal calibration, which uses an input-dependent mixture of experts to estimate candidate relevance; and (iii) calibrated candidate ranking, which reorders and truncates the beam according to the calibrated scores. The following subsections present these components and the corresponding training objective.
Multi-signal candidate representation
For user prefix , beam search generates SID sequences that are mapped to an item-level candidate set , where is the beam size and denotes a decoded item candidate. For every , we construct
The feature extractor combines three signal families: (i) LLM semantic signals, which measure consistency between a candidate and the user’s observed history; (ii) GenRec signals, including sequence likelihood, beam rank, and token-level confidence statistics; and (iii) candidate-context signals, which describe a candidate relative to the score distribution of . We retain raw features and append candidate-set-wise standardized versions, thereby preserving both absolute confidence and within-beam differences.
Nonlinear Signal Calibration
To capture nonlinear interactions among heterogeneous signals, MDC employs MLP experts with an input-dependent gating network. Given the candidate feature vector , the calibrated relevance score is computed as
where denotes the -th expert, and the Softmax operation produces input-dependent gating weights for expert aggregation.
With the generator frozen, MDC is optimized over generated candidates using binary cross-entropy:
where indicates whether a candidate matches the ground-truth item, and denotes the candidate training instances. This two-stage optimization enables refined candidate ranking without modifying the underlying generator.
Calibrated candidate ranking
At inference time, the recommendation list contains the candidates with the highest calibrated scores:
MDC can change only the ordering and truncation of ; it cannot recover a relevant item that the generator failed to include. Together, HiGenTrans establishes a unified framework for transferring LLM knowledge into generative recommendation at three complementary levels. Global Semantic Priming introduces global semantic priors through input initialization, Hierarchical Representation Transfer enhances internal representation learning through hierarchical semantic injection, and Multi-Signal Decision Calibration refines final recommendation decisions through MoE-based multi-signal calibration. These components enable semantic knowledge to contribute throughout the recommendation pipeline, from input initialization and candidate generation to final ranking.
Industry Deployment
Deploying HiGenTrans at industrial scale introduces additional storage and bandwidth costs, as HRT utilizes semantic embeddings from multiple LLM layers rather than only the final layer. To address this challenge, we adopt the SID-based compression scheme illustrated in Figure 3. Specifically, hierarchical semantic embeddings are quantized into compact discrete codes and reconstructed through layer-specific quantizers and shared codebooks:
Figure 3. SID-based compression for hierarchical transfer. Multi-level LLM embeddings are compressed into SIDs and reconstructed for stage-wise semantic transfer.
where and denote the quantizer and decoder at layer , respectively. Since codebooks are shared across user interaction prefixes, each cached instance stores only a few discrete indices, significantly reducing storage and transmission overhead.
LLM feature extraction and SID assignment are performed offline, while reconstructed embeddings are used during training to mitigate quantization errors. During online inference, lightweight codebook lookups recover semantic features for HRT without LLM execution, enabling efficient and scalable industrial deployment.
Experiments
We evaluate HiGenTrans by addressing four research questions:
RQ1: How does HiGenTrans compare with representative sequential and generative recommenders?
RQ2: How much does each transfer module contribute?
RQ3: Are the gains robust to reduced training data and consistent across user and item segments?
RQ4: Does hierarchical transfer remain effective in industrial recommendation scenarios?
Experimental Settings
Datasets and Protocol
We conduct experiments on three widely used categories from the Amazon review benchmark: Beauty, Sports and Outdoors (Sports), and Toys and Games (Toys). As summarized in Table 1, Beauty contains 22,363 users, 12,101 items, and 198,502 interactions; Sports contains 35,598 users, 18,357 items, and 296,337 interactions; and Toys contains 19,412 users, 11,924 items, and 167,597 interactions. We order each user’s interactions chronologically and adopt a leave-one-out protocol: the last interaction is used for testing, the penultimate interaction for validation, and all preceding interactions for training. We construct next-item prediction instances from the training prefixes and retain at most the 50 most recent interactions. Both the LLM input and the GenRec input are truncated at the same prediction boundary; neither contains the validation/test target or any later interactions. This protocol ensures that every semantic representation is computed from information observable before the target interaction.
| Dataset | #Users | #Items | #Interactions | Density (%) | Avg. Len. |
| Beauty | 22,363 | 12,101 | 198,502 | 0.073 | 8.88 |
| Sports | 35,598 | 18,357 | 296,337 | 0.045 | 8.32 |
| Toys | 19,412 | 11,924 | 167,597 | 0.072 | 8.63 |
Table 1. Statistics of the datasets used in our experiments.
Baselines
We compare HiGenTrans with representative methods from three recommendation paradigms. GRU4Rec [8] is a recurrent sequential recommendation model. SASRec [18], HSTU [39], and BERT4Rec [29] are attention-based models that learn sequential dependencies through self-attention, hierarchical transduction, and bidirectional masked-item prediction, respectively. The generative baselines include TIGER [28], which autoregressively generates hierarchical SIDs, RPG [10], which generates long SIDs in parallel, and GRLM [43], our generative backbone.
Implementation and Evaluation
We build HiGenTrans on the GRLM generative backbone and use Qwen2-Embedding-8B as the frozen semantic encoder. Each item is represented by a three-level SID obtained through residual K-means quantization. Item embeddings are projected to 256 dimensions and quantized with three codebooks of size 8, 192, yielding the discrete tokens used for autoregressive generation. All semantic features are extracted only from the observable history prefix and can be cached offline. We optimize the trainable modules with AdamW (, ) in bfloat16, using a batch size of 32, a maximum sequence length of 1,280, and gradient clipping at 1.0. The learning rate follows a cosine schedule from to after 50 warmup steps. Hyperparameters are selected on the validation set. We report Recall@1/5/10 (R@1/5/10) and NDCG@5/10 (N@5/10). For generative methods, decoded SIDs are mapped to item IDs; when multiple items share an SID, all corresponding items are included in evaluation.
Overall Performance (RQ1)
Table 2 reports performance on the three public datasets. HiGenTrans achieves the best result for every dataset and metric. On Beauty, it obtains 0.0917 R@10 and 0.0566 N@10, improving over the strongest baselines by 2.7% and 6.8%, respectively. On Sports, it reaches 0.0594 R@10 and 0.0342 N@10, corresponding to gains of 5.7% and 5.9%. The largest improvements occur on Toys, where HiGenTrans achieves 0.1047 R@10 and 0.0674 N@10, exceeding the best competing results by 12.9% and 33.7%.
| Paradigm | Model | Beauty | Sports | Toys | ||||||||||||
| R@1 | R@5 | R@10 | N@5 | N@10 | R@1 | R@5 | R@10 | N@5 | N@10 | R@1 | R@5 | R@10 | N@5 | N@10 | ||
| Recurrent | GRU4Rec | .0125 | .0343 | .0512 | .0235 | .0290 | .0043 | .0160 | .0262 | .0102 | .0134 | .0091 | .0239 | .0358 | .0167 | .0205 |
| Attention-based | SASRec | .0077 | .0566 | .0880 | .0327 | .0428 | .0046 | .0307 | .0492 | .0180 | .0239 | .0059 | .0620 | .0927 | .0352 | .0450 |
| HSTU | .0118 | .0563 | .0868 | .0346 | .0445 | .0076 | .0260 | .0427 | .0167 | .0220 | .0094 | .0630 | .0919 | .0370 | .0464 | |
| BERT4Rec | .0050 | .0230 | .0413 | .0138 | .0197 | .0033 | .0147 | .0259 | .0090 | .0126 | .0043 | .0241 | .0414 | .0139 | .0194 | |
| Generative Rec. | TIGER | .0113 | .0391 | .0626 | .0253 | .0329 | .0077 | .0250 | .0409 | .0164 | .0216 | .0128 | .0360 | .0565 | .0245 | .0311 |
| RPG | .0196 | .0506 | .0732 | .0354 | .0427 | .0104 | .0280 | .0420 | .0194 | .0239 | .0199 | .0557 | .0817 | .0382 | .0465 | |
| GRLM | .0251 | .0634 | .0893 | .0447 | .0530 | .0142 | .0388 | .0562 | .0267 | .0323 | .0241 | .0603 | .0842 | .0428 | .0504 | |
| Ours | HiGenTrans | .0292 | .0677 | .0917 | .0489 | .0566 | .0147 | .0422 | .0594 | .0287 | .0342 | .0372 | .0818 | .1047 | .0600 | .0674 |
Table 2. Overall performance comparison on three Amazon datasets. Best results are highlighted in bold, and second-best results are underlined.
The gains are particularly strong at the top of the ranking. On Beauty, HiGenTrans improves GRLM from 0.0251 to 0.0292 in R@1 (16.3%) and from 0.0447 to 0.0489 in N@5 (9.4%). On Sports, R@5 rises from 0.0388 to 0.0422 (8.8%), while N@5 increases from 0.0267 to 0.0287 (7.5%). On Toys, the strongest competing R@1 is 0.0241, whereas HiGenTrans reaches 0.0372, a 54.4% relative improvement. It also raises the best R@5 from 0.0630 to 0.0818 (29.8%) and the best N@5 from 0.0428 to 0.0600 (40.2%). These improvements across recall and position-sensitive NDCG indicate that the proposed transfer modules help both retrieve relevant items and place them nearer the top of the recommendation list.
The comparison also reveals that no single baseline paradigm dominates all settings. GRLM is the strongest baseline on Beauty and Sports, whereas attention-based SASRec and HSTU remain competitive on Toys. HiGenTrans improves over both groups, suggesting that LLM-derived semantics complement rather than replace the collaborative patterns learned from interaction sequences. The relative N@10 gain is larger than the R@10 gain on every dataset (6.8% versus 2.7% on Beauty, 5.9% versus 5.7% on Sports, and 33.7% versus 12.9% on Toys). Since NDCG is more sensitive to rank positions, this gap indicates that the framework improves ordering in addition to retrieval. The component analysis below attributes part of this effect directly to MDC.
Component Analysis (RQ2)
Table 3 isolates the three transfer levels on Toys by adding each module separately to the same LLM4GenRec base. Every module improves all reported metrics. GSP increases R@10 from 0.0803 to 0.0958 and N@10 from 0.0480 to 0.0569, absolute gains of 1.55 and 0.89 percentage points, respectively. Thus, a single prefix-level semantic prior already compensates for information absent from discrete interaction IDs. HRT reaches 0.0967 R@10 and 0.0577 N@10, slightly outperforming GSP at the deeper cutoffs. Because the two variants use the same semantic source and differ mainly in transfer placement, the result supports preserving depth-specific features instead of relying only on input propagation. MDC is the strongest individual component, reaching 0.1010 R@10 and 0.0628 N@10. Its larger NDCG gain is consistent with its objective: it can not introduce a missing beam candidate, but it can promote an already generated target into the evaluated prefix of the list.
| Method | R@1 | R@5 | N@10 | R@10 |
| Base LLM4GenRec | .0229 | .0581 | .0480 | .0803 |
| +Global Semantic Priming | .0269 | .0696 | .0569 | .0958 |
| +Hierarchical Transfer | .0269 | .0707 | .0577 | .0967 |
| +Decision Calibration | .0319 | .0753 | .0628 | .1010 |
| Full Model | .0372 | .0818 | .0674 | .1047 |
Table 3. Ablation study of HiGenTrans on the Toys dataset.
Combining all three modules yields the best results: 0.0372 R@1, 0.0818 R@5, 0.1047 R@10, and 0.0674 N@10. Relative to the base model, the full framework improves these metrics by 62.4%, 40.8%, 30.4%, and 40.4%, respectively. It also outperforms MDC, the strongest single-module variant, by 16.6% in R@1 and 7.3% in N@10. This improvement over every single-module model is important: the generator-side modules and the reranker are not substitutes. GSP and HRT change the representations used to form the beam, while MDC uses evidence unavailable to raw beam scores to determine which of those candidates occupy the top positions.
Stage-wise Discriminability of LLM Representations (RQ2)
To investigate how semantic discriminability evolves across different stages of the LLM, we evaluate the candidate-ranking capability of representations extracted from four stages (Stage 1–4). Specifically, we use cosine similarity between candidate items and user historical representations as the ranking score, and measure their ability to distinguish ground-truth items from negative candidates using AUC. The evaluation considers users whose ground-truth items appear among the 20 generated candidates and have available LLM representations. As shown in Table 4, deeper stages generally exhibit stronger discriminative ability across all three datasets. Compared with Stage 1, Stage 3 improves AUC by 0.0168, 0.0041, and 0.0179 on Toys, Beauty, and Sports, respectively, suggesting that deeper LLM representations capture more recommendation-relevant semantic information. However, Stage 4 does not consistently outperform Stage 3, indicating that semantic discriminability tends to saturate at deeper stages. These findings suggest that deeper stages provide stronger standalone ranking signals, motivating hierarchical semantic transfer across multiple representation stages without requiring dataset-specific stage selection.
| Dataset | S1 | S2 | S3 | S4 | ||||
| Toys | 0.5788 | 0.5863 | 0.5956 | 0.5953 | +0.75 | +0.93 | -0.03 | +1.65 |
| Beauty | 0.5434 | 0.5414 | 0.5475 | 0.5530 | -0.20 | +0.61 | +0.55 | +0.96 |
| Sports | 0.5302 | 0.5335 | 0.5481 | 0.5440 | +0.33 | +1.46 | -0.41 | +1.38 |
Table 4. Stage-wise discriminative ability of LLM representations measured by candidate-ranking AUC. indicates absolute AUC changes in percentage points.
Data-Efficiency Analysis (RQ3)
Table 5 evaluates HiGenTrans after subsampling the training users on Beauty. Reducing the training set from 100% to 95%, 90%, and 80% changes R@10 from 0.0917 to 0.0908, 0.0906, and 0.0903, respectively. Even at the 80% setting, the absolute decrease is only 0.0014 (1.5% relative to the full-data result), while the improvement over the correspondingly trained base model remains 7.24%. The relative gain varies by only 1.62 percentage points across the four ratios. This stability is consistent with the role of a frozen semantic encoder: when interaction supervision decreases, the external semantic features remain available.
| Users | Hit@10 | Gain (%) | Range (%) | (pp) |
| 100% | 0.0917 | +8.86 | — | — |
| 95% | 0.0908 | +7.84 | 7.17–8.44 | −0.09 ↓ |
| 90% | 0.0906 | +7.59 | 7.48–7.70 | −0.11 ↓ |
| 80% | 0.0903 | +7.24 | 6.58–8.01 | −0.14 ↓ |
Table 5. Robustness to reduced training data on Beauty.
Segment-Level Benefits (RQ3)
Impact of Item Popularity. We evaluate HiGenTrans across different item-popularity groups on Toys, as shown in Table 6. Our method consistently outperforms the baseline across all groups, demonstrating its effectiveness for both head and long-tail items. Notably, the largest improvements occur for items with popularity 3–5, achieving relative gains of 30.4% in Hit@10 and 64.1% in NDCG@10. Even for extremely long-tail items (popularity ≤ 2), HiGenTrans achieves gains of 21.3% and 42.8%, respectively. These results highlight the effectiveness of semantic knowledge transfer in alleviating item sparsity.
| Popularity | Hit@10 Base / Ours | Hit@10 Gain | NDCG@10 Base / Ours | NDCG@10 Gain |
| All | .0932 / .1047 | +12.3% | .0544 / .0674 | +23.9% |
| ≤ 2 | .0330 / .0400 | +21.3% | .0239 / .0341 | +42.8% |
| 3–5 | .0591 / .0770 | +30.4% | .0327 / .0536 | +64.1% |
| 6–15 | .0964 / .1097 | +13.7% | .0575 / .0690 | +20.2% |
| ≥ 16 | .1341 / .1412 | +5.2% | .0768 / .0867 | +13.0% |
Table 6. Performance across item popularity groups on Toys. Gains denote relative improvements (%) over the baseline. Darker green indicates larger gains.
Impact of User History Length. We further investigate performance across different user history lengths in Table 7. HiGenTrans consistently improves recommendation accuracy across all groups, with relative Hit@10 gains ranging from 10.7% to 16.3%. Users with longer histories (≥ 10) achieve the largest Hit@10 improvement of 16.3%, while users with 6–9 interactions obtain the highest NDCG@10 gain of 32.4%. These findings demonstrate that our method effectively leverages semantic knowledge across varying levels of user interaction sparsity.
| History | Hit@10 Base / Ours | Hit@10 Gain | NDCG@10 Base / Ours | NDCG@10 Gain |
| All | .0932 / .1047 | +12.3% | .0544 / .0674 | +23.9% |
| 3 | .1023 / .1138 | +11.3% | .0612 / .0736 | +20.3% |
| 4–5 | .1001 / .1108 | +10.7% | .0582 / .0719 | +23.6% |
| 6–9 | .0739 / .0849 | +14.9% | .0419 / .0554 | +32.4% |
| ≥ 10 | .0841 / .0978 | +16.3% | .0479 / .0598 | +25.0% |
Table 7. Performance across user history lengths on Toys. Gains denote relative improvements (%) over the baseline. Darker green indicates larger gains.
Candidate Ranking Correction (RQ3)
To evaluate the effectiveness of MDC, we analyze the ranking transitions of ground-truth items within the top-20 candidates, as shown in Figure 4. MDC successfully promotes 53.4%, 55.1%, and 43.5% of items originally ranked 11–20 into the top-10 on Beauty, Toys, and Sports, respectively. Meanwhile, 83–86% of top-5 items retain their positions within the top-5. These results demonstrate that MDC effectively corrects suboptimal rankings while preserving high-quality predictions.
Figure 4. Ground-truth ranking transitions before and after MDC. Each cell shows the transition percentage and user count. Dashed lines indicate the top-10 boundary.
Industrial Evaluation (RQ4)
We further evaluate hierarchical transfer in two real-world cross-domain e-commerce scenarios. As shown in Table 8, HiGenTrans consistently improves SID prediction accuracy across all hierarchical levels in both domains. Notably, SID1 achieves substantial absolute gains of 48.18 and 61.98 percentage points, while SID2 and SID3 show smaller but consistent improvements. This trend suggests that LLM semantic knowledge is particularly beneficial for coarse-grained predictions, while fine-grained predictions rely more on collaborative signals. Overall, these results validate the effectiveness of hierarchical transfer in industrial recommendation systems.
| Domain 1 | Domain 2 | |||||
| Method | SID | SID | SID | SID | SID | SID |
| Baseline | 0.1360 | 0.4789 | 0.8038 | 0.2149 | 0.4698 | 0.6083 |
| Ours | 0.6178 | 0.6155 | 0.8226 | 0.8347 | 0.5797 | 0.6383 |
| +0.4818 | +0.1366 | +0.0188 | +0.6198 | +0.1099 | +0.0300 |
Table 8. SID prediction performance across different hierarchical levels in two industrial cross-domain recommendation scenarios. denotes absolute improvements, with darker green indicating larger gains.
Conclusion
In this work, we proposed HiGenTrans, a hierarchical LLM-to-GenRec knowledge transfer framework that enhances generative recommendation at the input, representation, and decision levels. Specifically, GSP introduces global semantic priors, HRT transfers multi-level LLM representations into the generative backbone, and MDC integrates complementary semantic, generative, and contextual signals to refine candidate rankings. To support efficient deployment, we further incorporate offline semantic feature compression, avoiding online LLM inference. Extensive experiments on three public benchmarks demonstrate consistent improvements, achieving up to 12.94% relative gains in Recall@10. Overall, our findings highlight the potential of hierarchical semantic knowledge transfer and decision-level calibration for building accurate and efficient generative recommender systems.
References
- [1]Ben Chen, Siyuan Wang, Yufei Ma, Zihan Liang, Xuxin Zhang, Yue Lv, Ying Yang, Huangyu Dai, Lingtao Mao, Tong Zhao, et al. 2026. OneSearch-V2: The latent Reasoning Enhanced Self-distillation Generative Search Framework. arXiv preprint arXiv:2603.24422 (2026).arxiv.org/abs/2603.24422
- [2]Wei Cheng, Yucheng Lu, Boyang Xia, Jiangxia Cao, Kuan Xu, Mingxing Wen, Wei Jiang, Jiaming Zhang, Zhaojie Liu, Liyin Hong, et al. 2026. ChorusCVR: Chorus supervision for entire space post-click conversion rate modeling. In Proceedings of the Nineteenth ACM International Conference on Web Search and Data Mining. 1115–1119.arxiv.org/abs/2502.08277
- [3]Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Youn Kim, James R Glass, and Pengcheng He. 2024. Dola: Decoding by contrasting layers improves factuality in large language models. In International Conference on Learning Representations, Vol. 2024. 54158–54183.arxiv.org/abs/2309.03883
- [4]Xinnan Dai, Haohao Qu, Yifei Shen, Bohang Zhang, Qihao Wen, Wenqi Fan, Dongsheng Li, Jiliang Tang, and Caihua Shan. 2025. How do large language models understand graph patterns? A benchmark for graph pattern comprehension. In International Conference on Learning Representations, Vol. 2025. 58405–58435.arxiv.org/abs/2410.05298
- [5]Jiaxian Deng, Shiyao Wang, Kuo Cai, Leijian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment. arXiv preprint arXiv:2502.18965 (2025).
- [6]Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin. 2019. Graph neural networks for social recommendation. In The world wide web conference. 417–426.arxiv.org/abs/1902.07243
- [7]Wenqi Fan, Yao Ma, Qing Li, Jianping Wang, Guoyong Cai, Jiliang Tang, and Dawei Yin. 2020. A graph neural network framework for social recommendations. IEEE Transactions on Knowledge and Data Engineering 34, 5 (2020), 2033–2047.DOI
- [8]Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2015. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939 (2015).
- [9]Minjie Hong, Yan Xia, Zehan Wang, Jieming Zhu, Ye Wang, Sihang Li, Xiaoda Yang, Quanyu Dai, Zhenhua Dong, Zhimeng Zhang, et al. 2025. Eager-llm: Enhancing large language models as recommenders through exogenous behavior-semantic integration. In Proceedings of the ACM on Web Conference 2025. 2754–2762.
- [10]Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating long semantic ids in parallel for recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 956–966.arxiv.org/abs/2506.05781
- [11]Zheng Hu, Yuxin Chen, Yongsen Pan, Xu Yuan, Yuting Yin, Daoyuan Wang, Boyang Xia, Zefei Luo, Hongyang Wang, Songhao Ni, et al. 2026. Stop treating collisions equally: Qualification-aware semantic id learning for recommendation at industrial scale. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 7401–7411.
- [12]Zheng Hu, Yongsen Pan, Zetao Li, Jiaming Huang, Satoshi Nakagawa, Jiawen Deng, Shimin Cai, and Fuji Ren. 2026. Retrieval-enhanced, Adaptively Collaborative, and Temporal-aware user behavior comprehension for LLM-based sequential recommendation. Information Processing & Management 63, 1 (2026), 104354.DOI
- [13]Ganesh Jawahar, Benoît Sagot, and Djamé Sed̄dah. 2019. What does BERT learn about the structure of language? In Proceedings of the 57th annual meeting of the association for computational linguistics. 3651–3657.DOI
- [14]Zhuohang Jiang, Yuxin Chen, Yongsen Pan, Zheng Hu, Wenqi Fan, Qing Li, Hongyang Wang, Jun Wang, and Wenwu Ou. 2026. A/B Agent: A Self-Evolving Agent for Strategy Iteration in Industrial A/B Testing. arXiv preprint arXiv:2608.04625 (2026).arxiv.org/abs/2608.04625
- [15]Zhuohang Jiang, Yuxin Chen, Shijie Wang, Haohao Qu, Jindong Zhou, Wenqi Fan, Qing Li, Dongxu Liang, and Jun Wang. 2026. Atomic intent reasoning: Bringing llm semantics to industrial cross-domain recommendations. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 7501–7510.DOI
- [16]Zhuohang Jiang, Pangjing Wu, Xu Yuan, Wenqi Fan, and Qing Li. 2025. Qa-dragon: Query-aware dynamic rag system for knowledge-intensive visual question answering. arXiv preprint arXiv:2508.05197 (2025).arxiv.org/abs/2508.05197
- [17]Zhuohang Jiang, Xu Yuan, Haohao Qu, Shanru Lin, Kanglong Liu, Wenqi Fan, and Li Qing. 2026. Superglasses: Benchmarking vision language models as intelligent agents in a smart glasses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2165–2175.arxiv.org/abs/2602.22683
- [18]Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM). IEEE, 197–206.
- [19]Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large language models meet collaborative filtering: An efficient all-round llm-based recommender system. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 1395–1406.arxiv.org/abs/2404.11343
- [20]Sunkyung Lee, Minjin Choi, Eunseong Choi, Hye-young Kim, and Jongwuk Lee. 2025. Gram: Generative recommendation via semantic-aware multi-granular late fusion. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 33294–33312.arxiv.org/abs/2506.01673
- [21]Qidong Liu, Xian Wu, Yejing Wang, Zijian Zhang, Feng Tian, Yefeng Zheng, and Xiangyu Zhao. 2024. LLM-esr: Large language models enhancement for long-tailed sequential recommendation. Advances in Neural Information Processing Systems 37 (2024), 26701–26727.arxiv.org/abs/2405.20646
- [22]Liangbo Ning, Ziran Liang, Zhuohang Jiang, Haohao Qu, Yujuan Ding, Wenqi Fan, Xiao-yong Wei, Shanru Lin, Hui Liu, Philip S Yu, et al. 2025. A survey of webagents: Towards next-generation ai agents for web automation with large foundation models. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 6140–6150.arxiv.org/abs/2503.23350
- [23]Yongsen Pan, Yuxin Chen, Zheng He, Xu Yuan, Daoyuan Wang, Yuting Yin, Songhao Ni, Hongyang Wang, Jun Wang, Fuji Ren, et al. 2026. Beyond Static Collision Handling: Adaptive Semantic ID Learning for Multimodal Recommendation at Industrial Scale. arXiv preprint arXiv:2604.23522 (2026).
- [24]Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lei Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 2685–2692.
- [25]Haohao Qu, Wenqi Fan, Zihui Zhao, and Qing Li. 2025. Tokenrec: Learning to tokenize id for llm-based generative recommendations. IEEE Transactions on Knowledge and Data Engineering 37, 10 (2025), 6216–6231.
- [26]Haohao Qu, Shanru Lin, Yujuan Ding, Yiqi Wang, and Wenqi Fan. 2026. Diffusion sequence recommendation with continuous tokens. In Proceedings of the ACM Web Conference 2026. 7259–7270.
- [27]Haohao Qu, Yifeng Zhang, Liangbo Ning, Wenqi Fan, and Qing Li. 2024. Ssd4rec: a structured state space duality model for efficient sequential recommendation. arXiv preprint arXiv:2409.01192 (2024).
- [28]Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al. 2023. Recommender systems with generative retrieval. Advances in Neural Information Processing Systems 36 (2023), 10299–10315.arxiv.org/abs/2305.05065
- [29]Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management. 1441–1450.
- [30]Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th annual meeting of the association for computational linguistics. 4593–4601.arxiv.org/abs/1905.05950
- [31]Shijie Wang, Wenqi Fan, Yue Feng, Lin Shanru, Xinyu Ma, Shuaiqiang Wang, and Dawei Yin. 2025. Knowledge graph retrieval-augmented generation for llm-based recommendation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 27152–27168.
- [32]Chen Sheng, Yusheng Huang, Ruochen Yang, Shuang Wen, Pengbo Xu, Jiangxia Cao, Yueyang Liu, Kuo Cai, Chengcheng Guo, Shiyao Wang, et al. 2026. Onelive: Dynamically unified generative framework for live-streaming recommendation. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 8155–8166.
- [33]Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua. 2024. Learnable item tokenization for generative recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2400–2409.arxiv.org/abs/2405.07314
- [34]Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, et al. 2024. Eager: Two-stream generative recommendation with behavior-semantic collaboration. In Proceedings of the 30th
- [35]Boyang Xia, Zhou Yu, Zhiliang Zhu, Hanxiao Sun, Biyun Han, Jun Wang, Runnan Liu, and Wenwu Ou. 2026. HarmonRank: Ranking-aligned Multi-objective Ensemble for Live-streaming E-commerce Recommendation. arXiv e-prints (2026), arXiv–2601.
- [36]Longtao Xiao, Haozhao Wang, Cheng Wang, Linfei Ji, Yifan Wang, Jieming Zhu, Zhenhua Dong, Rui Zhang, and Ruixuan Li. 2025. Progressive collaborative and semantic knowledge fusion for generative recommendation. arXiv preprint arXiv:2502.06269 (2025).
- [37]Xiaohan Yu, Li Zhang, Xin Zhao, and Yue Wang. 2024. Break the ID-Language Barrier: An Adaption Framework for LLM-based Sequential Recommendation. arXiv preprint arXiv:2411.18262 (2024).
- [38]Xu Yuan, Yi Wang, Zhuohang Jiang, Haohao Qu, Yujuan Ding, Shanru Lin, Guoliang Xing, Hongxia Yang, Jiannong Cao, Qing Li, et al. 2026. AI Smart Glasses for Wearable Intelligence: From Egocentric Sensing to Agentic Personalization. arXiv preprint arXiv:2609.17993 (2026).
- [39]Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jayuan He, Yinghai Lu, and Yu Shi. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 58484–58509. https://proceedings.mlr.press/v235/zhai24a.html
- [40]Kun Zhang, Jingming Zhang, Wei Cheng, Yansong Cheng, Jiaqi Zhang, Hao Lu, Xu Zhang, Haixiang Gan, Jiangxia Cao, Tenglong Wang, et al. 2026. OneMall: One Model, More Scenarios–End-to-End Generative Recommender Family at Kuaishou E-Commerce. arXiv e-prints (2026), arXiv–2601.
- [41]Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He. 2025. Collm: Integrating collaborative embeddings into large language models for recommendation. IEEE Transactions on Knowledge and Data Engineering 37, 5 (2025), 2329–2340.
- [42]Yi Zhang, Yiwen Zhang, Yu Wang, Tong Chen, and Hongzhi Yin. 2025. Towards distribution matching between collaborative and language spaces for generative recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2006–2016.
- [43]Zhiyang Zhang, Junda She, Kuo Cai, Bo Chen, Shiyao Wang, Xinchen Luo, Qiang Luo, Ruiming Tang, Han Li, Kun Gai, et al. 2026. Unleashing the Native Recommendation Potential: LLM-Based Generative Recommendation via Structured Term Identifiers. In Findings of the Association for Computational Linguistics: ACL 2026. 19659–19673.
- [44]Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, and Qing Li. 2024. Recommender Systems in the Era of Large Language Models (LLMs). IEEE Transactions on Knowledge and Data Engineering 36, 11 (2024), 6889–6907. doi:10.1109/TKDE.2024.3392335
- [45]Yuheng Zheng, Yu Cui, Bin Wu, Jian Zhang, Ye Feng, Can Wang, and Jiawei Chen. 2026. IMFuse: Instance-Aware Multi-Layer Fusion for LLM-Enhanced Sequential Recommendation. arXiv preprint arXiv:2607.27002 (2026).
- [46]Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024. Collaborative large language model for recommender systems. In Proceedings of the ACM Web Conference 2024. 3162–3172.



