Paper deep dive
Tlow: Flow-based Item Tokenizer for Recommendation
Nian Li, Chonggang Song, Jingtao Ding, Lingling Yi, Yong Li, Qingmin Liao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/29/2026, 4:08:20 AM
Summary
The paper introduces Tlow, a flow-based item tokenizer for recommendation systems that transforms semantic embeddings into a latent space conforming to a standard normal distribution. This approach addresses limitations of existing tokenizers like RQ-VAE and OPQ by ensuring dimensional independence and distributional simplicity, leading to more semantically clear token IDs. Tlow incorporates a codebook guidance mechanism to align token embeddings with codebook spaces. Experiments on Amazon Reviews datasets and an online A/B test on WeChat demonstrate significant improvements in recommendation performance, including a 10.32% increase in user CTR globally.
Entities (9)
Relation Signals (8)
Tlow → evaluatedon → Amazon Reviews
confidence 95% · Offline experiments on four public datasets... Amazon Reviews datasets
Tlow → improves → CTR
confidence 95% · The retrieval model based on token IDs improves user CTR by 10.32%
Tlow → testedon → WeChat
confidence 95% · Online experiments... on China's largest social media platform WeChat validate Tlow's powerful distribution transformation capability.
Tlow → uses → Flow-based Model
confidence 95% · we propose a flow-based item Tokenizer (Tlow)
Codebook Guidance → partof → Tlow
confidence 92% · Additionally, we introduce a novel codebook guidance... in this work
Nian Li → affiliatedwith → Tsinghua University
confidence 90% · Nian Li... Affiliation: Tsinghua University, Shenzhen, China
Chonggang Song → affiliatedwith → Tencent Inc.
confidence 90% · Chonggang Song... Affiliation: Tencent Inc., Shenzhen, China
RQ-VAE → suffersfrom → low decoding efficiency
confidence 90% · RQ-VAE, suffers from low decoding efficiency due to the inherent dependencies among its codebooks.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Item tokenizer encodes semantic embeddings into token IDs to replace the randomly assigned item IDs used in traditional recommendation models, fundamentally addressing the problems of excessive parameters and cold starts. However, the most common tokenizer, RQ-VAE, suffers from low decoding efficiency due to the inherent dependencies among its codebooks. Meanwhile, efficient independent tokenizers such as optimized product quantization (OPQ) still struggle with dimensional correlations and distribution complexity of semantic embeddings. In this work, we propose a f\underline{low}-based item \underline{T}okenizer (Tlow) to transform raw semantic embeddings into a latent space where embeddings conform to a unified standard normal distribution, achieving dual advantages of dimensional independence and distributional simplicity. Independent tokenization performed on these latent embeddings yields semantically clear token IDs. Additionally, we introduce a novel codebook guidance to align the codebook space with the token embedding space, further aiding the learning of more semantically distinct token embeddings. Offline experiments on four public datasets demonstrate that Tlow's tokenization and codebook guidance significantly improve recommendation performance. The improvement on cross-domain and multi-modal recommendations also proves the effectiveness of item tokenization in a simplified embedding space. Online experiments for a multi-modal retrieval task on China's largest social media platform WeChat validate Tlow's powerful distribution transformation capability. The retrieval model based on token IDs improves user CTR by 10.32\% globally and by 11.64\% for new items. Our codes are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.24176v1
- Canonical: https://arxiv.org/abs/2608.24176v1
Trouble viewing inline? Open PDF directly →
Full Text
50,986 characters extracted from source content.
Expand or collapse full text
Tlow: Flow-based Item Tokenizer for Recommendation Conference: Proceedings of the 35th ACM International Conference on Information and Knowledge Management; November 07–11, 2026; Rome, ItalyProceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM ’26), November 07–11, 2026, Rome, ItalyDOI: 10.1145/3799682.3840087ISBN: 979-8-4007-2539-5/2026/11CCS: Information systems Retrieval models and ranking Nian Li Note: Both authors contribute equally to this work. Affiliation: Tsinghua University, Shenzhen, China , Chonggang Song Affiliation: Tencent Inc., Shenzhen, China , Jingtao Ding Affiliation: Tsinghua University, Beijing, China , Lingling Yi Affiliation: Tencent Inc., Shenzhen, China , Yong Li Affiliation: Tsinghua University, Beijing, China and Qingmin Liao Affiliation: Shenzhen International Graduate School, Tsinghua University, Shenzhen, China © c Abstract. Item tokenizer encodes semantic embeddings into token IDs to replace the randomly assigned item IDs used in traditional recommendation models, fundamentally addressing the problems of excessive parameters and cold starts. However, the most common tokenizer, RQ-VAE, suffers from low decoding efficiency due to the inherent dependencies among its codebooks. Meanwhile, efficient independent tokenizers such as optimized product quantization (OPQ) still struggle with dimensional correlations and distribution complexity of semantic embeddings. In this work, we propose a flow-based item Tokenizer (Tlow) to transform raw semantic embeddings into a latent space where embeddings conform to a unified standard normal distribution, achieving dual advantages of dimensional independence and distributional simplicity. Independent tokenization performed on these latent embeddings yields semantically clear token IDs. Additionally, we introduce a novel codebook guidance to align the codebook space with the token embedding space, further aiding the learning of more semantically distinct token embeddings. Offline experiments on four public datasets demonstrate that Tlow’s tokenization and codebook guidance significantly improve recommendation performance. The improvement on cross-domain and multi-modal recommendations also proves the effectiveness of item tokenization in a simplified embedding space. Online experiments for a multi-modal retrieval task on China’s largest social media platform WeChat validate Tlow’s powerful distribution transformation capability. The retrieval model based on token IDs improves user CTR by 10.32% globally and by 11.64% for new items. Our codes are available at https://github.com/wjjln/Tlow. Keywords: Recommender System; Item Tokenizer; Flow-based Model †c-license: by 1. Introduction Traditional recommender systems often assign a random ID embedding to each item, with learning embeddings as the model’s core objective. However, embeddings learned by ID-based models (GNNs (Wang et al., 2019), RNNs (Hidasi et al., 2015), or self-attention (Kang and McAuley, 2018; Zhou et al., 2018)) are completely independent of one another, so the number of parameters grows linearly with the number of items, limiting model capacity in practical online recommender systems. Moreover, the cold-start issue is another significant drawback, i.e., the model struggles to recommend new items with few interaction records. Inspired by the training paradigm of large language models, some studies now tokenize items based on their semantic embeddings to generate a sequence of discrete token IDs (also called semantic IDs) (Rajput et al., 2023; Hou et al., 2025; Hou et al., 2023), creating a globally shared token vocabulary. Sequential recommendation models then learn token embeddings by decoding the token IDs of the next item. This approach relies on token embeddings to model atomic information in the semantic space of all items, effectively constraining the model’s parameter size while naturally resolving the cold-start issue. A representative work, TIGER (Rajput et al., 2023), uses residual-quantized variational auto-encoder (RQ-VAE) (Zeghidour et al., 2021) as its tokenizer, applying hierarchical encoding to semantic embeddings. This creates strong correlations between codebooks, so decoding the next token ID relies on previously decoded IDs, causing an efficiency bottleneck: the number of inference steps must equal the number of codebooks. Moreover, the codebook collapse and conflicts common in RQ-VAE (Kuai et al., 2024) make its industrial application require substantial experience and expertise (Zhou et al., 2025). Pursuing independent tokenization to enable efficient parallel decoding is thus a natural next step, and this decoding-efficiency advantage over RQ-VAE-based sequential decoding has been empirically validated in (Hou et al., 2025). Existing methods using product quantization (PQ) (Jegou et al., 2010) directly partition semantic embeddings and encode each part independently (Hou et al., 2023), but still face two critical challenges: • Dimension correlations. The correlations between the dimensions of the semantic embeddings violate the dimensional-independence assumption required for independent tokenization. Although optimized product quantization (OPQ) (Ge et al., 2013) alleviates this by decomposing the raw space into subspaces via an orthogonal transformation, dimensions within each subspace remain correlated (Hou et al., 2025). When dimensions are correlated, embeddings lie on a complex, skewed manifold, but independent quantization imposes a grid-like structure over it, leading to poor quantization (Zhang et al., 2022). • Complex embedding distribution. The unknown, complex distribution harms the semantic representational ability of embeddings (Li et al., 2020) and causes large quantization errors when tokenized directly. Specifically, these embeddings are often highly anisotropic, occupying a non-uniform cone rather than being evenly spread (Jegou et al., 2010), making standard quantization codebooks an inefficient fit, especially in cross-domain or multi-modal scenarios where embeddings from diverse sources form distinct, separated clusters. In this work, we propose Tlow, a flow-based item tokenizer that transforms raw semantic embeddings into latent embeddings conforming to a standard normal distribution. Compared to the raw, correlated, and anisotropic embedding space, the transformed latent space better satisfies the dimensional-independence assumption underlying PQ-based independent quantization (Jegou et al., 2010; Ge et al., 2013), mitigating the quantization error induced by correlated and skewed distributions. The resulting latent embeddings combine dimensional independence with distributional simplicity, facilitating more accurate tokenization and more semantically distinct token embeddings. Specifically, Tlow processes semantic embeddings with a multi-scale architecture (Kingma and Dhariwal, 2018), built from basic units called “A Step of Flow”, each comprising three layers: Activation Normalization (ActNorm), Invertible Linear, and Affine Coupling. We apply PQ on the latent embeddings to generate token IDs for each item. Figure 1 compares 16-bit (i.e., 16-part) OPQ and Tlow tokenization on the “CDs and Vinyl” dataset from Amazon Reviews (McAuley et al., 2015), showing albums sharing the same first-position token ID. Although these albums are all related to the “Pop” genre, OPQ groups in many albums of different genres (e.g., “Metal”, “British Invasion”, “Jazz”), whereas Tlow’s tokenization yields distinct, clear semantics, with most albums genuinely “Pop”. Tlow also exhibits significantly lower genre diversity than OPQ, as indicated by fewer categories and lower entropy. We then use user behavioral data to train the sequential recommendation model and the globally shared token embeddings, introducing a novel codebook guidance that aligns the token embedding space with the codebook space so token embeddings capture distinct semantics and atomic information. We conduct extensive experiments on four Amazon Reviews datasets from different commodity categories. Tlow significantly improves recommendation performance given the same sequential model, validating the effectiveness of tokenization within a standard normal distribution space; ablation studies further confirm the importance of the latent embeddings’ semantic information and the codebook guidance. Performance gains in cross-domain and multi-modal recommendations further support Tlow’s effectiveness in simplifying complex semantic embedding distributions, although the improvement is not uniformly largest across every domain and modality metric. On China’s largest social media platform WeChat, we conduct online A/B testing for a multi-modal picture11 1 A typical user post consists of a few pictures and a brief caption. retrieval task, where replacing item IDs with Tlow’s token IDs yields significant gains in conversion metrics, especially for newly published pictures. Figure 1. A Comparison of semantic clarity of token IDs generated by OPQ and Tlow tokenization. 2. Tlow In this work, we formulate the task as sequential recommendations. Specifically, based on each user’s historical interaction sequence i1,i2,⋯,iT\i_1,i_2,·s,i_T\, the recommendation model aims to predict the next interaction item iT+1i_T+1, where T is the sequence length. Traditional ID-based models assign a random ID for each item and learn ID embeddings through sequential models like Transformer decoder. However, ID-based models suffer from inherent problems of cold-start issues, excessive parameters, and so on. Item tokenizer leverages the semantic embedding x to transform each item into a sequence of token IDs, where x is obtained from pre-trained models with item features (e.g., title, cover image) as the input. In this way, tokenization-based sequential models have the objective of decoding token IDs of the next item, similar to the training paradigm of large language models. This approach overcomes the aforementioned fundamental problems by learning a shared set of token embeddings. 2.1. Overview As stated in the introduction, dimensional correlations and complex distribution of semantic embeddings lead to challenges of independent tokenization. In this work, we propose Tlow with a multi-scale architecture (Kingma and Dhariwal, 2018) to transform semantic embedding ∈ℝdsx ^d_s into latent embedding ∈ℝdsz ^d_s conforming to a standard normal distribution, where dsd_s is the embedding dimension. We can perform more accurate independent tokenization on z, generating token IDs with clearer semantics. Like all flow-based models, Tlow’s log-likelihood objective is also to minimize: (1) ℒf=−1||∑∈logp(),L_f=- 1|X| _x p_ θ(x), where X denotes the set of semantic embeddings of all the items and θ are learnable model parameters. The probability density function of Tlow can be further written as: (2) logp()=logp()+log|det(d/d)|, p_ θ(x)= p_ θ(z)+ | (dz/dx)|, where d/ddz/dx is the Jacobian matrix of the transformation from x to z, and det(⋅) (·) denotes the matrix determinant. Tlow contains multiple blocks, each consisting of a multi-step flow where each step includes three transformation layers, i.e., ActNorm, Invertible Linear, and Affine Coupling layer. Figure 2 illustrates the model architecture of Tlow. Figure 2. The illustration of Tlow Architecture with N=3N=3 blocks and M=2M=2 flows, where the dimension of semantic embedding is ds=4d_s=4 (left). An example of “A Step of Flow” processing embedding with the dimension d=4d=4 (right). 2.2. A Step of Flow A step of flow consists of the following three transformation layers: • ActNorm Layer. For the input 0∈ℝdx_0 ^d, this layer performs activation normalization for stable training as follows, (3) 1=⊙(0+),x_1=s (x_0+t), where ⊙ denotes the element-wise product and ,∈ℝds,t ^d are learnable scale and bias parameters, respectively. The log-determinant of this layer is: (4) log|det(d1/d0)|=∑j=1dlog|j|. | (dx_1/dx_0)|= _j=1^d |s_j|. • Invertible Linear Layer. For the input 1x_1, this layer learns an invertible linear transformation to allow the dimensions of the embedding to influence one another: (5) 2=1=()1.x_2=Wx_1=(PLU)x_1. The linear matrix ∈ℝd×dW ^d× d is decomposed into the product of a permutation matrix P, a lower triangular matrix L, and an upper triangular matrix U to facilitate the computation of the determinant. We keep the diagonal elements of L as ones, while making the diagonal of U a set of learnable parameters ∈ℝdw ^d. In this way, the log-determinant of this layer is: (6) log|det(d2/d1)| | (dx_2/dx_1)| =log|det()| = | (W)| =log|det()|+log|det()|+log|det()| = | (P)|+ | (L)|+ | (U)| =∑j=1dlog|j|. = _j=1^d |w_j|. • Affine Coupling Layer. For the input 2x_2, this layer first divides it into two halves 2ax_2^a and 2bx_2^b, and performs the transformation as follows: (7) 3=[2a,b⊙(2b+b)], _3=[x_2^a,s^b (x_2^b+t^b)], [b,b]=g(2a), [s^b,t^b]=g(x_2^a), where g:ℝd2→ℝdg:R d2 ^d is a MLP to generate scale and bias parameters based on 2ax_2^a to transform 2bx_2^b. The log-determinant of this layer is: (8) log|det(d3/d2)|=∑j=1d/2log|jb|. | (dx_3/dx_2)|= _j=1^d/2 |s_j^b|. Combining these three layers together, the log-determinant of a step of flow is: (9) Δ=log|det(d1/d0)|+log|det(d2/d1)|+log|det(d3/d2)|. = | (dx_1/dx_0)|+ | (dx_2/dx_1)|+ | (dx_3/dx_2)|. 2.3. Flow-based Tokenizer Tlow contains N blocks and each block consists of M steps of flow. For the n-th block, the output n∈ℝds2n−1z_n d_s2^n-1 is divided into two halves na∈ℝds2nz_n^a d_s2^n and nb∈ℝds2nz_n^b d_s2^n, which serve as the input of the next block and partial output (i.e., partial z) of Tlow, respectively. We estimate the log-likelihood of nbz_n^b as follows: (10) logp(nb)=log(nb,n,n2), p(z_n^b)= (z_n^b; μ_n, σ_n^2), [n,n2]=h(na), [ μ_n, σ_n^2]=h(z_n^a), where h:ℝds2n→ℝds2n−1h:R d_s2^n d_s2^n-1 is an MLP module to predict the mean n μ_n and variance n2 σ_n^2 of nbz_n^b based on naz_n^a. Combining the outputs of all the blocks, we obtain final output of Tlow =[1b,2b,⋯,Nb]∈ℝdsz=[z_1^b,z_2^b,·s,z_N^b] ^d_s. Note that Nb=Nz_N^b=z_N since we do not need to divide the output of the last block. Moreover, the input of the first block is x. We then formulate: (11) logp()=∑n=1Nlogp(nb). p_ θ(z)= _n=1^N p(z_n^b). In addition, the log-determinant of Tlow is calculated as follows: (12) log|det(d/d)|=∑n=1N∑m=1MΔn,m, | (dz/dx)|= _n=1^N _m=1^M _n,m, where Δn,m _n,m denotes the log-determinant of the m-th step of flow in the n-th block. After training Tlow with the log-likelihood objective ℒfL_f, the semantic embedding x is transformed into a latent embedding z conforming to a standard normal distribution, where the dimensions are independent of each other. For item tokenization, we perform the PQ on z to obtain C token IDs. Specifically, z is divided into C parts [1,2,⋯,C][z_1,z_2,·s,z_C], each of which is independently encoded using K-means to obtain a corresponding codebook k=k1,k2,⋯,kSC_k=\c_k^1,c_k^2,·s,c_k^S\. Here, kj∈ℝdsCc_k^j d_sC is the j-th quantized codeword and S denotes the codebook size. For the k-th tokenization, the token ID is defined as: (13) ck=argminj∈1,2,⋯,S‖k−kj‖2.c_k= _j∈\1,2,·s,S\\|z_k-c_k^j\|^2. In this way, Tlow tokenizes each item into discrete IDs [c1,c2,⋯,cC][c_1,c_2,·s,c_C]. 2.4. Token IDs for Recommendation Given token IDs of all the items, a sequential model (e.g., Transformer decoder) aims to generate the IDs of the next item, with a set of token embeddings learned. Specifically, there is a one-to-one correspondence between the token embeddings ∈ℝC×S×dmE ^C× S× d_m and codebooks ∈ℝC×S×dsCC ^C× S× d_sC, where dmd_m denotes the dimension of token embeddings. Traditional ID embedding can be replaced with the aggregated token embeddings ∈ℝdme ^d_m for each item: (14) =1C∑k=1Ck,ck.e= 1C _k=1^CE_k,c_k. The sequential model further encodes the user’s historical embedding sequence [i1,i2,⋯] [e_i_1,e_i_2,·s ] and outputs a final hidden state ∈ℝdmh ^d_m. Following (Hou et al., 2025), we predict the log-likelihood of interaction with the target item iti_t with token IDs [c1t,c2t,⋯,cCt] [c_1^t,c_2^t,·s,c_C^t ] as follows: (15) logp(it|)=∑k=1Clogp(ckt|)=∑k=1Clogexp(k,ckt⊤gk()/τ)∑j=1Sexp(k,j⊤gk()/τ), p(i_t|h)= _k=1^C p (c_k^t|h )= _k=1^C (E_k,c_k^t g_k(h)/τ ) _j=1^S (E_k,j g_k(h)/τ ), where τ is the temperature hyperparameter. The k-th project head gk:ℝdm→ℝdmg_k:R^d_m ^d_m decodes the k-th token ID of the next item. This decoding approach owes to the independence between different parts of the latent embedding z. Furthermore, the training loss for recommendation can be formulated as: (16) ℒrec=−∑(,it)logp(it|).L_rec=- _(h,i_t) p(i_t|h). 2.5. Learning Token Embeddings with Codebook Guidance Since Tlow tokenizes items based on latent embeddings, the quantized codebooks imply clearer semantics contained in the whole set of items. In other words, each codeword in a codebook encodes atomic information derived from the decomposition of items’ semantics, as the illustration in Figure 1. Therefore, learning token embeddings with the codebook guidance will help enhance recommendations. Unlike independent alignment for each item or behavior sequence used in existing works (Wang et al., 2024a; Liu et al., 2024), we propose to directly align two latent spaces spanned by token embeddings and codebooks. The token space ∈ℝC×S×S ^C× S× S and codebook space ∈ℝC×S×S ^C× S× S are defined based on cosine similarities between token embeddings and codewords as follows: (17) k,i,j=k,i⊤k,j‖k,i‖⋅‖k,j‖,k,i,j=k,i⊤k,j‖k,i‖⋅‖k,j‖. _k,i,j= E_k,i E_k,j\|E_k,i\|·\|E_k,j\|, _k,i,j= C_k,i C_k,j\|C_k,i\|·\|C_k,j\|. Furthermore, we implement the codebook guidance as a simple but effective MSE loss: (18) ℒsim=MSE(|−|).L_sim=MSE (| - | ). Finally, the sequential model and token embeddings are learned by training the combined loss ℒ=ℒrec+λℒsimL=L_rec+ _sim, where λ is the degree of codebook guidance. 3. Experiments We conduct experiments under three scenarios to evaluate the effectiveness of our proposed Tlow, including general, cross-domain, and multi-modal sequential recommendation. 3.1. Experimental Setup 3.1.1. Datasets We use four commodity categories in Amazon Reviews dataset (McAuley et al., 2015) for experiments, including “Sports and Outdoors (Sports), “Beauty”, “Toys and Games (Toys)”, and “CDs and Vinyl (CDs)”. Following existing works (Hou et al., 2025; Hou et al., 2023; Rajput et al., 2023), users’ reviews are regarded as item interactions and arranged chronologically as historical item sequences. We also use a processed dataset “Cloth-Sports” from a state-of-the-art (SOTA) work (Liu et al., 2025) for cross-domain sequential recommendation. This dataset is collected from “Clothing Shoes and Jewelry” and “Sports and Outdoors” categories in Amazon dataset. Most of the users in this dataset overlap across two categories. The statistics of used datasets are shown in Table 1, where Avg. Length means the length of historical sequence averaged over all the users. The last item in a historical sequence is used for testing and the second-to-last item for validation, which is the widely adopted leave-one-out evaluation protocol in existing works (Kang and McAuley, 2018; Hou et al., 2025; Zhao et al., 2022). Table 1. Statistics of used datasets. Datasets #Users #Items #Interactions Avg. Length Sports 35,598 18,357 260,739 8.32 Beauty 22,363 12,101 176,139 8.87 Toys 19,412 11,924 148,185 8.63 CDs 75,258 64,443 1,022,334 14.58 Cloth 9,933 3,278 97,741 10.71 Sports 4,263 1,021 11,879 3.1.2. Baselines We compare Tlow with two types of baselines following existing works (Hou et al., 2025; Liu et al., 2024; Rajput et al., 2023). Traditional ID-based models include Caser (Tang and Wang, 2018), GRU4Rec (Hidasi et al., 2015), HGN (Ma et al., 2019), BERT4Rec (Sun et al., 2019), SASRec (Kang and McAuley, 2018), FDSA (Zhang et al., 2019), and S3-Rec (Zhou et al., 2020), which adopt different neural architectures (e.g., CNNs, RNNs, and Transformers) to model user behavior sequences with randomly assigned item IDs. Tokenization-based models include VQRec (Hou et al., 2023), TIGER (Rajput et al., 2023), ETEGRec (Liu et al., 2024), RecJPQ (Petrov and Macdonald, 2024), HSTU (Zhai et al., 2024), and RPG (Hou et al., 2025), which leverage various quantization techniques (e.g., PQ, RQ-VAE, and OPQ) to tokenize item semantic embeddings for generative recommendation. 3.1.3. Evaluation Metrics We use widely adopted ranking metrics Recall@k (R@k) and NDCG@k (N@k) for performance evaluation, where k is set as 55 and 1010 following (Hou et al., 2025; Rajput et al., 2023). All the items are included for full ranking to avoid sampling bias when calculating evaluation metrics (Krichene and Rendle, 2020). 3.1.4. Implementation Details We implement our Tlow with Pytorch. All the experimental setups for training and evaluation follow RPG (Hou et al., 2025), thus we use baseline results from the original paper and rerun RPG. For a fair comparison, all the models use the extracted item semantic embeddings by the text encoder sentence-t5-base (Ni et al., 2021), leading to ds=768d_s=768. We set the number of blocks and the steps of flow in Tlow as N=4N=4 and M=4M=4. In fact, the recommendation performance is not sensitive to the number of blocks and flows, because Tlow can easily transform the semantic embeddings into latent embeddings conforming to a standard normal distribution. The number of codebooks C∈16,32,64,96,128C∈\16,32,64,96,128\ and the codebook size S is fixed as 256256 following RPG. The temperature τ∈0.03,0.05,0.07τ∈\0.03,0.05,0.07\ when decoding token IDs. For the hyperparameters of the sequential model, we keep them consistent with those of RPG to ensure a similar number of model parameters: a 2-layer GPT-2 with the token embedding dimension of dm=448d_m=448. Since Tlow decodes all C token IDs of the next item independently within a single forward pass (Eq. 15), in the same manner as RPG, it retains the same decoding parallelism as RPG and thus the same decoding-efficiency advantage over RQ-VAE-based sequential decoding (e.g., TIGER) that has already been empirically validated in (Hou et al., 2025); we therefore do not repeat separate latency/throughput benchmarking in this work. 3.2. Overall Performance The overall performance comparison is shown in Table 2. It is obvious that tokenization-based models using item semantics significantly outperform purely ID-based models. These results prove the effectiveness of item tokenization for more accurate modeling of user behaviors. Moreover, Tlow performs best on all the metrics across four datasets, demonstrating better tokenization in the transformed latent space through a flow-based model. It is worth noting that when using the same semantic embeddings generated by sentence-t5-base, RPG with its independent tokenizer does not outperform TIGER consistently. This also demonstrates the crucial role of Tlow’s tokenization on latent embeddings conforming to a standard normal distribution. Table 2. Performance comparison on general recommendation. Underline denotes the best baseline and bold denotes better performance than the best baseline at a significance level of p<0.01p<0.01 under paired t-test. Model Sports and Outdoors Beauty Toys and Games CDs and Vinyl R@5 N@5 R@10 N@10 R@5 N@5 R@10 N@10 R@5 N@5 R@10 N@10 R@5 N@5 R@10 N@10 ID Caser 0.0116 0.0072 0.0194 0.0097 0.0205 0.0131 0.0347 0.0176 0.0166 0.0107 0.0270 0.0141 0.0116 0.0073 0.0205 0.0101 GRU4Rec 0.0129 0.0086 0.0204 0.0110 0.0164 0.0099 0.0283 0.0137 0.0097 0.0059 0.0176 0.0084 0.0195 0.0120 0.0353 0.0171 HGN 0.0189 0.0120 0.0313 0.0159 0.0325 0.0206 0.0512 0.0266 0.0321 0.0221 0.0497 0.0277 0.0259 0.0153 0.0467 0.0220 BERT4Rec 0.0115 0.0075 0.0191 0.0099 0.0203 0.0124 0.0347 0.0170 0.0116 0.0071 0.0203 0.0099 0.0326 0.0201 0.0547 0.0271 SASRec 0.0233 0.0154 0.0350 0.0192 0.0387 0.0249 0.0605 0.0318 0.0463 0.0306 0.0675 0.0374 0.0351 0.0177 0.0619 0.0263 FDSA 0.0182 0.0122 0.0288 0.0156 0.0267 0.0163 0.0407 0.0208 0.0228 0.0140 0.0381 0.0189 0.0226 0.0137 0.0378 0.0186 S3-Rec 0.0251 0.0161 0.0385 0.0204 0.0387 0.0244 0.0647 0.0327 0.0443 0.0294 0.0700 0.0376 0.0213 0.0130 0.0375 0.0182 Tokenization RecJPQ 0.0141 0.0076 0.0220 0.0102 0.0311 0.0167 0.0482 0.0222 0.0331 0.0182 0.0484 0.0231 0.0075 0.0046 0.0138 0.0066 VQ-Rec 0.0208 0.0144 0.0300 0.0173 0.0457 0.0317 0.0664 0.0383 0.0497 0.0346 0.0737 0.0423 0.0352 0.0238 0.0520 0.0292 TIGER 0.0264 0.0181 0.0400 0.0225 0.0454 0.0321 0.0648 0.0384 0.0521 0.0371 0.0712 0.0432 0.0492 0.0329 0.0748 0.0411 ETEGRec 0.0175 0.0114 0.0281 0.0149 0.0404 0.0277 0.0587 0.0337 0.0209 0.0136 0.0339 0.0178 0.0309 0.0204 0.0461 0.0253 HSTU 0.0258 0.0165 0.0414 0.0215 0.0469 0.0314 0.0704 0.0389 0.0433 0.0281 0.0669 0.0357 0.0417 0.0275 0.0638 0.0346 RPG 0.0296 0.0203 0.0428 0.0246 0.0533 0.0366 0.0753 0.0437 0.0509 0.0357 0.0765 0.0440 0.0486 0.0328 0.0693 0.0395 Tlow 0.0307 0.0207 0.0477 0.0261 0.0545 0.0377 0.0786 0.0454 0.0590 0.0395 0.0864 0.0482 0.0541 0.0362 0.0801 0.0446 Impr. (%) 3.72 1.97 11.45 6.10 2.25 3.01 4.38 3.89 11.52 9.70 9.80 11.36 9.96 10.03 7.09 8.52 3.3. Ablation Study We conduct further experiments to validate the effectiveness of two critical modules in Tlow: the transformed latent embeddings z and the codebook guidance. • Random z. Latent embeddings z are sampled directly from a standard normal distribution instead of being transformed from semantic embeddings by Tlow. • w/o ℒsimL_sim. The codebook guidance is removed by setting its corresponding degree λ=0λ=0. The results in Table 3 demonstrate the effectiveness of codebook guidance for enhanced recommendation through learning a semantically clearer latent space of token embeddings. Furthermore, the tokenization based on randomly sampled z leads to a significant performance drop. This confirms that Tlow retains rich semantic information necessary for behaviors modeling when transforming the embedding x. Table 3. Ablation study of Tlow. Model Sports Beauty Toys CDs R@10 N@10 R@10 N@10 R@10 N@10 R@10 N@10 Tlow 0.0477 0.0261 0.0786 0.0454 0.0864 0.0482 0.0801 0.0446 w/o ℒsimL_sim 0.0449 0.0248 0.0764 0.0436 0.0826 0.0471 0.0756 0.0425 Random z 0.0202 0.0103 0.0623 0.0363 0.0571 0.0341 0.0147 0.0074 3.4. Cross-domain and Multi-modal Recommendation In this section, we further validate the necessity and effectiveness of transforming semantic embeddings from different distributional spaces into a unified space for tokenization. 3.4.1. Cross-domain Recommendation We choose the current SOTA cross-domain model LLM4CDSR (Liu et al., 2025) as our baseline, which leverages LLMs to bridge the domain gap with item semantic embeddings and hierarchical user profiling. Besides, RPG is also included to demonstrate the effectiveness of Tlow when tokenizing items from different domains. For the implementation of RPG and Tlow, we directly use the mixed item sequences merged from two domains for model training. The results in Table 4 demonstrates the significant improvement of Tlow over both LLM4CDSR and RPG. It is effective for two reasons: first, tokenizing items to create a shared vocabulary of token IDs and training their embeddings naturally bridges the cross-domain gap. Second, Tlow’s ability to transform embedding distributions is highly effective for improving recommendations across different domains. Table 4. Performance comparison on cross-domain recommendation. Model Overall Cloth Sports R@10 N@10 R@10 N@10 R@10 N@10 LLM4CDSR 0.4620 0.2803 0.4220 0.2637 0.5507 0.3172 RPG 0.4766 0.3457 0.4487 0.3353 0.5385 0.3686 Tlow 0.5558 0.4395 0.5669 0.4568 0.5312 0.4014 3.4.2. Multi-modal Recommendation We chose the current SOTA multi-modal model HM4SR (Zhang et al., 2025) as our baseline, which leverages interactive and temporal mixture of experts to capture user dynamic interests. Since HM4SR utilizes a large amount of additional side-information (including interaction timestamp, item category, semantic embeddings as additional input), we replace the ID embeddings in HM4SR with the token embeddings e generated by Tlow and RPG to directly compare the effectiveness of tokenization. We conduct experiments on the “Sports” dataset and generate additional image embeddings using the advanced CLIP model OpenCLIP ViT-H/14 (Radford et al., 2021), where the image is downloaded from product URLs of each item. The tokenization on text and image embeddings are performed independently, and both token IDs are merged as final token IDs. The results in Table 5 show that RPG’s tokenization brings limited and inconsistent gains over HM4SR in this setting with combined text-image distributions (e.g., it ties HM4SR on R@5 and underperforms on N@5 and N@10). In contrast, Tlow consistently improves over both baselines on all four metrics after performing the distribution transformation, although the absolute margin over HM4SR (e.g., R@10 from 0.0469 to 0.0521) is moderate in this single-domain multi-modal setting; we expect larger gains with richer or more heterogeneous modalities. Table 5. Performance comparison on multi-modal recommendation. Model R@5 N@5 R@10 N@10 HM4SR 0.0326 0.0231 0.0469 0.0277 RPG 0.0326 0.0216 0.0501 0.0272 Tlow 0.0343 0.0233 0.0521 0.0290 3.5. Cold-Start Recommendation Focusing on item semantics, tokenization-based models inherently have an advantage in cold-start recommendation. Therefore, we further investigate whether Tlow’s tokenization can enhance recommendation performance in cold-start scenarios. We conduct our validation from both user and item perspectives. Specifically, users are grouped by their number of historical interactions (interaction depth), while items in the test set are grouped by their frequency in the training set (popularity). We then evaluate the recommendation performance for these different groups on the largest “CDs” dataset. Figure 3 shows the performance comparison between Tlow and RPG, where Tlow performs significantly better across all the groups. It indicates that due to Tlow’s ability to transform any embeddings (including long-tail ones) into a standard normal distribution, its tokenization on items across all popularity levels is more comprehensive. This further leads to more accurate behavioral modeling for users with varying interaction depths. Figure 3. Performance comparison under different user and item groups. 3.6. Online Experiments 3.6.1. Online Setup We verify Tlow’s online performance on China’s largest social media platform WeChat, in the picture recommendation scenario involving image-text multi-modal features, comparing sequential models that use Tlow’s tokenized embeddings e against those using randomly assigned item-ID embeddings. We construct users’ behavior sequences of length 500 with their recently clicked items, and feed the sequence of item embeddings into a 12-layer Transformer decoder to predict the user’s interaction tendency toward the next item. Both sequential models are integrated as retrieval pathways in addition to the primary DSSM-based real-time retrieval method, and the online serving scheme is shown in Figure 4. During the online serving phase, user embeddings and item embeddings are stored in dedicated embedding servers. Upon user access, the corresponding user embedding is retrieved via a real-time lookup operation, and candidate pictures with the highest similarity scores are then dynamically identified by the Similarity Server (Douze et al., 2024) to construct the retrieval results. We conduct both single-domain and cross-domain online experiments: • Single-domain: Only users’ clicked pictures are used to construct behavior sequences for model training. • Cross-domain: Both users’ clicked pictures and articles are mixed to construct behavior sequences for model training. This is to verify the effectiveness of Tlow’s tokenization in cross-domain scenarios. 3.6.2. Training and Inference In the stage of training Tlow, we find that the model converges after training on embeddings of approximately 1 million items (pictures and articles). Then Tlow can generate token IDs for other unseen items by performing semantic tokenization. As a result, the training time of Tlow is negligible, and the inference efficiency is such that tokenization for tens of millions of newly published items per day can be completed in just a few minutes on a single GPU. In the deployment, we set the number of codebooks C=16C=16 and codebook size S=256S=256. Therefore, the number of token embeddings that the Tlow-based sequential model needs to learn is only C×S=4096C× S=4096, far fewer than the tens of millions of item embeddings in the baseline model. The Tlow-based sequential model is initialized with all parameters of the baseline model except the item-ID embeddings. Before being deployed for online A/B test, both models are trained for two weeks on identical datasets with more than 70 and 200 million interaction records per day in single-domain and cross-domain scenarios, respectively. 3.6.3. Online Performance After deploying the aforementioned Tlow-based and ID-based models as supplemental retrieval pathways, we evaluate the conversion efficacy of both pathways across the entire picture corpus and specifically on cold-start pictures exposed on the same day as publication. The following metrics are reported: • CTR (Click-Through Rate) measures the click-through rate of items exposed via this specific retrieval pathway. • UCTR (User Click-Through Rate) averages all the users’ click-through rate of items exposed via this pathway. Results are shown in Table 6. Due to privacy considerations, we report only the relative difference for all metric comparisons between the models. As we can see, the sequential model employing Tlow’s tokenization demonstrates 4.79%4.79\% and 6.23%6.23\% higher CTR compared to the one using random IDs in single-domain and cross-domain scenarios, respectively. Furthermore, we observe a more substantial improvement of 10.32%10.32\% and 7.20%7.20\% in UCTR, indicating its effectiveness in retrieving potentially relevant content to a broader user base. These findings validate Tlow’s powerful capabilities on multi-modal and cross-domain embedding transformation in large-scale industrial recommender systems. To further validate Tlow’s effectiveness on cold-start recommendations, we perform evaluations exclusively on newly published pictures each day. The results reveal that the advantages of Tlow become markedly more pronounced when evaluated solely on these cold-start pictures, since random ID embeddings require extensive user interaction data for training, whereas Tlow utilizes inherent semantic features available immediately upon a picture’s publication. Figure 4. Online serving of Tlow’s tokenization. Table 6. Online metrics improvement (%) of pathway comparison in single-domain and cross-domain scenarios. Scenario Overall New Item CTR UCTR CTR UCTR Single-domain 4.79 10.32 8.46 11.64 Cross-domain 6.23 7.20 9.09 9.45 Beyond pathway comparisons, core online metrics from A/B testing also improve significantly: in the single-domain scenario, 24-hour overall and newly-published picture CTR increase by 0.78%0.78\% and 1.94%1.94\%, respectively; in the cross-domain scenario, per-capita picture CTR improves by 1.05%1.05\%, accompanied by a 1.15%1.15\% decrease in top-tier accounts’ exposure share, indicating that Tlow substantially benefits long-tail picture exposure. 4. Related Work 4.1. ID-based Recommendation ID-based recommendation originates from collaborative filtering, evolving from matrix factorization (Koren et al., 2009) to GNN-based models (Wang et al., 2019). For sequential recommendation, various architectures have been explored, including Markov Chains (Rendle et al., 2010), RNNs and CNNs (Tang and Wang, 2018; Hidasi et al., 2015; Xu et al., 2019; Yan et al., 2019), and self-attention mechanisms (Zhou et al., 2018; Kang and McAuley, 2018; Zhang et al., 2025). Multi-modal information such as text and images has also been incorporated to enrich item representations (Zhang et al., 2025; Tao et al., 2020). 4.2. Tokenization-based Recommendation TIGER (Rajput et al., 2023) pioneers the use of RQ-VAE to tokenize item semantic embeddings, followed by works that introduce alignment strategies to enhance tokenization quality (Wang et al., 2024a; Liu et al., 2024; Wang et al., 2024b). However, the hierarchical nature of RQ-VAE leads to low decoding efficiency due to codebook correlations. To enable parallel decoding, PQ-based independent tokenization methods have been proposed (Hou et al., 2023; Hou et al., 2025), while parameter-free tokenizers based on heuristic algorithms remain limited in performance (Hua et al., 2023; Petrov and Macdonald, 2023; Si et al., 2024). Despite this progress, existing independent tokenizers still face challenges from non-independent embedding dimensions and complex embedding distributions, which motivates our flow-based approach. 5. Conclusion We propose Tlow, a flow-based item tokenizer that transforms semantic embeddings into a standard normal distribution space, achieving dimensional independence and distributional simplicity for more accurate independent tokenization. Tlow significantly improves recommendation performance across general, cross-domain, multi-modal, and cold-start scenarios in both offline and online experiments. Acknowledgements. This work is supported by the National Natural Science Foundation of China under U24B20180. References Douze et al. (2024) M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P. Mazaré, M. Lomeli, L. Hosseini, and H. Jégou The faiss library. External Links: 2401.08281 Cited by: §3.6.1. Ge et al. (2013) T. Ge, K. He, Q. Ke, and J. Sun Optimized product quantization. IEEE transactions on pattern analysis and machine intelligence 36 (4), p. 744–755. Cited by: 1st item, §1. Hidasi et al. (2015) B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939. Cited by: §1, §3.1.2, §4.1. Hou et al. (2023) Y. Hou, Z. He, J. McAuley, and W. X. Zhao Learning vector-quantized item representation for transferable sequential recommenders. In Proceedings of the ACM Web Conference 2023, p. 1162–1171. Cited by: §1, §3.1.1, §3.1.2, §4.2. Hou et al. (2025) Y. Hou, J. Li, A. Shin, J. Jeon, A. Santhanam, W. Shao, K. Hassani, N. Yao, and J. McAuley Generating long semantic ids in parallel for recommendation. arXiv preprint arXiv:2506.05781. Cited by: 1st item, §1, §2.4, §3.1.1, §3.1.1, §3.1.2, §3.1.3, §3.1.4, §4.2. Hua et al. (2023) W. Hua, S. Xu, Y. Ge, and Y. Zhang How to index item ids for recommendation foundation models. In Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, p. 195–204. Cited by: §4.2. Jegou et al. (2010) H. Jegou, M. Douze, and C. Schmid Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence 33 (1), p. 117–128. Cited by: 2nd item, §1, §1. Kang and McAuley (2018) W. Kang and J. McAuley Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), p. 197–206. Cited by: §1, §3.1.1, §3.1.2, §4.1. Kingma and Dhariwal (2018) D. P. Kingma and P. Dhariwal Glow: generative flow with invertible 1x1 convolutions. Advances in neural information processing systems 31. Cited by: §1, §2.1. Koren et al. (2009) Y. Koren, R. Bell, and C. Volinsky Matrix factorization techniques for recommender systems. Computer 42 (8), p. 30–37. Cited by: §4.1. Krichene and Rendle (2020) W. Krichene and S. Rendle On sampled metrics for item recommendation. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, p. 1748–1757. Cited by: §3.1.3. Kuai et al. (2024) Z. Kuai, Z. Chen, H. Wang, M. Li, D. Miao, B. Wang, X. Chen, L. Kuang, Y. Han, J. Wang, et al. Breaking the hourglass phenomenon of residual quantization: enhancing the upper bound of generative retrieval. arXiv preprint arXiv:2407.21488. Cited by: §1. Li et al. (2020) B. Li, H. Zhou, J. He, M. Wang, Y. Yang, and L. Li On the sentence embeddings from pre-trained language models. arXiv preprint arXiv:2011.05864. Cited by: 2nd item. Liu et al. (2024) E. Liu, B. Zheng, C. Ling, L. Hu, H. Li, and W. X. Zhao Generative recommender with end-to-end learnable item tokenization. arXiv preprint arXiv:2409.05546. Cited by: §2.5, §3.1.2, §4.2. Liu et al. (2025) Q. Liu, X. Zhao, Y. Wang, Z. Zhang, H. Zhong, C. Chen, X. Li, W. Huang, and F. Tian Bridge the domains: large language models enhanced cross-domain sequential recommendation. arXiv preprint arXiv:2504.18383. Cited by: §3.1.1, §3.4.1. Ma et al. (2019) C. Ma, P. Kang, and X. Liu Hierarchical gating networks for sequential recommendation. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, p. 825–833. Cited by: §3.1.2. McAuley et al. (2015) J. McAuley, C. Targett, Q. Shi, and A. Van Den Hengel Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, p. 43–52. Cited by: §1, §3.1.1. Ni et al. (2021) J. Ni, G. H. Abrego, N. Constant, J. Ma, K. B. Hall, D. Cer, and Y. Yang Sentence-t5: scalable sentence encoders from pre-trained text-to-text models. arXiv preprint arXiv:2108.08877. Cited by: §3.1.4. Petrov and Macdonald (2023) A. V. Petrov and C. Macdonald Generative sequential recommendation with gptrec. arXiv preprint arXiv:2306.11114. Cited by: §4.2. Petrov and Macdonald (2024) A. V. Petrov and C. Macdonald RecJPQ: training large-catalogue sequential recommenders. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, p. 538–547. Cited by: §3.1.2. Radford et al. (2021) A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: §3.4.2. Rajput et al. (2023) S. Rajput, N. Mehta, A. Singh, R. Hulikal Keshavan, T. Vu, L. Heldt, L. Hong, Y. Tay, V. Tran, J. Samost, et al. Recommender systems with generative retrieval. Advances in Neural Information Processing Systems 36, p. 10299–10315. Cited by: §1, §3.1.1, §3.1.2, §3.1.3, §4.2. Rendle et al. (2010) S. Rendle, C. Freudenthaler, and L. Schmidt-Thieme Factorizing personalized markov chains for next-basket recommendation. In Proceedings of the 19th international conference on World wide web, p. 811–820. Cited by: §4.1. Si et al. (2024) Z. Si, Z. Sun, J. Chen, G. Chen, X. Zang, K. Zheng, Y. Song, X. Zhang, J. Xu, and K. Gai Generative retrieval with semantic tree-structured identifiers and contrastive learning. In Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, p. 154–163. Cited by: §4.2. Sun et al. (2019) F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang BERT4Rec: sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, p. 1441–1450. Cited by: §3.1.2. Tang and Wang (2018) J. Tang and K. Wang Personalized top-n sequential recommendation via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining, p. 565–573. Cited by: §3.1.2, §4.1. Tao et al. (2020) Z. Tao, Y. Wei, X. Wang, X. He, X. Huang, and T. Chua Mgat: multimodal graph attention network for recommendation. Information Processing & Management 57 (5), p. 102277. Cited by: §4.1. Wang et al. (2024a) W. Wang, H. Bao, X. Lin, J. Zhang, Y. Li, F. Feng, S. Ng, and T. Chua Learnable item tokenization for generative recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, p. 2400–2409. Cited by: §2.5, §4.2. Wang et al. (2019) X. Wang, X. He, M. Wang, F. Feng, and T. Chua Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval, p. 165–174. Cited by: §1, §4.1. Wang et al. (2024b) Y. Wang, J. Xun, M. Hong, J. Zhu, T. Jin, W. Lin, H. Li, L. Li, Y. Xia, Z. Zhao, et al. Eager: two-stream generative recommender with behavior-semantic collaboration. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 3245–3254. Cited by: §4.2. Xu et al. (2019) C. Xu, P. Zhao, Y. Liu, J. Xu, V. S. S. S. Sheng, Z. Cui, X. Zhou, and H. Xiong Recurrent convolutional neural network for sequential recommendation. In The world wide web conference, p. 3398–3404. Cited by: §4.1. Yan et al. (2019) A. Yan, S. Cheng, W. Kang, M. Wan, and J. McAuley CosRec: 2d convolutional neural networks for sequential recommendation. In Proceedings of the 28th ACM international conference on information and knowledge management, p. 2173–2176. Cited by: §4.1. Zeghidour et al. (2021) N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi Soundstream: an end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30, p. 495–507. Cited by: §1. Zhai et al. (2024) J. Zhai, L. Liao, X. Liu, Y. Wang, R. Li, X. Cao, L. Gao, Z. Gong, F. Gu, M. He, et al. Actions speak louder than words: trillion-parameter sequential transducers for generative recommendations. arXiv preprint arXiv:2402.17152. Cited by: §3.1.2. Zhang et al. (2022) J. Zhang, Q. Liu, D. Lian, Z. Liu, L. Wu, and E. Chen Anisotropic additive quantization for fast inner product search. In Proceedings of the AAAI conference on Artificial Intelligence, Vol. 36, p. 4354–4362. Cited by: 1st item. Zhang et al. (2025) S. Zhang, L. Chen, D. Shen, C. Wang, and H. Xiong Hierarchical time-aware mixture of experts for multi-modal sequential recommendation. In Proceedings of the ACM on Web Conference 2025, p. 3672–3682. Cited by: §3.4.2, §4.1. Zhang et al. (2019) T. Zhang, P. Zhao, Y. Liu, V. S. Sheng, J. Xu, D. Wang, G. Liu, X. Zhou, et al. Feature-level deeper self-attention network for sequential recommendation.. In IJCAI, p. 4320–4326. Cited by: §3.1.2. Zhao et al. (2022) W. X. Zhao, Z. Lin, Z. Feng, P. Wang, and J. Wen A revisiting study of appropriate offline evaluation for top-n recommendation algorithms. ACM Transactions on Information Systems 41 (2), p. 1–41. Cited by: §3.1.1. Zhou et al. (2025) G. Zhou, J. Deng, J. Zhang, K. Cai, L. Ren, Q. Luo, Q. Wang, Q. Hu, R. Huang, S. Wang, et al. OneRec technical report. arXiv preprint arXiv:2506.13695. Cited by: §1. Zhou et al. (2018) G. Zhou, X. Zhu, C. Song, Y. Fan, H. Zhu, X. Ma, Y. Yan, J. Jin, H. Li, and K. Gai Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, p. 1059–1068. Cited by: §1, §4.1. Zhou et al. (2020) K. Zhou, H. Wang, W. X. Zhao, Y. Zhu, S. Wang, F. Zhang, Z. Wang, and J. Wen S3-rec: self-supervised learning for sequential recommendation with mutual information maximization. In Proceedings of the 29th ACM international conference on information & knowledge management, p. 1893–1902. Cited by: §3.1.2.