Paper deep dive
CRAB: Codebook Rebalancing for Bias Mitigation in Generative Recommendation
Zezhong Fan, Ziheng Chen, Luyi Ma, Jin Huang, Lalitesh Morishetti, Kaushiki Nag, Sushant Kumar, Kannan Achan
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/10/2026, 3:01:43 AM
Summary
CRAB is a post-hoc debiasing strategy for Generative Recommendation (GeneRec) that addresses popularity bias by rebalancing the semantic codebook. It identifies over-popular tokens and splits them using a regularized K-means approach while preserving hierarchical semantic structure, complemented by a tree-structured regularizer to enhance representations for unpopular tokens.
Entities (5)
Relation Signals (3)
CRAB → mitigates → popularity bias
confidence 100% · CRAB, a post-hoc debiasing strategy for GeneRec that alleviates popularity bias
CRAB → improves → GeneRec
confidence 95% · Experiments on real-world datasets demonstrate that CRAB significantly improves recommendation performance
GeneRec → suffersfrom → popularity bias
confidence 95% · existing GeneRec approaches still suffer from severe popularity bias
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Generative recommendation (GeneRec) has introduced a new paradigm that represents items as discrete semantic tokens and predicts items in a generative manner. Despite its strong performance across multiple recommendation tasks, existing GeneRec approaches still suffer from severe popularity bias and may even exacerbate it. In this work, we conduct a comprehensive empirical analysis to uncover the root causes of this phenomenon, yielding two core insights: 1) imbalanced tokenization inherits and can further amplify popularity bias from historical item interactions; 2) current training procedures disproportionately favor popular tokens while neglecting semantic relationships among tokens, thereby intensifying popularity bias. Building on these insights, we propose CRAB, a post-hoc debiasing strategy for GeneRec that alleviates popularity bias by mitigating frequency imbalance among semantic tokens. Specifically, given a well-trained model, we first rebalance the codebook by splitting over-popular tokens while preserving their hierarchical semantic structure. Based on the adjusted codebook, we further introduce a tree-structured regularizer to enhance semantic consistency, encouraging more informative representations for unpopular tokens during training. Experiments on real-world datasets demonstrate that CRAB significantly improves recommendation performance by effectively alleviating popularity bias.
Tags
Links
- Source: https://arxiv.org/abs/2604.05113v1
- Canonical: https://arxiv.org/abs/2604.05113v1
Trouble viewing inline? Open PDF directly →
Full Text
27,798 characters extracted from source content.
Expand or collapse full text
CRAB: Codebook Rebalancing for Bias Mitigation in Generative Recommendation Ziheng Chen albertchen1993pokemon@gmail.com Walmart Global TechSunnyvaleUSA , Zezhong Fan zezhong.fan@walmart.com Walmart Global TechSunnyvaleUSA , Luyi Ma luyi.ma@walmart.com Walmart Global TechSunnyvaleUSA , Jin Huang jin.huang@stonybrook.edu Stony Brook University, NY, USAStony BrookUSA , Lalitesh Morishetti lalitesh.morishetti@walmart.com Walmart Global TechSunnyvaleUSA , Kaushiki Nag kaushiki.nag@walmart.com Walmart Global TechSunnyvaleUSA , Sushant Kumar sushant.kumar@walmart.com Walmart Global TechSunnyvaleUSA and Kannan Achan kannan.achan@walmart.com Walmart Global TechSunnyvaleUSA Abstract. Generative recommendation (GeneRec) has introduced a new paradigm that represents items as discrete semantic tokens and predicts items in a generative manner. Despite its strong performance across multiple recommendation tasks, existing GeneRec approaches still suffer from severe popularity bias and may even exacerbate it. In this work, we conduct a comprehensive empirical analysis to uncover the root causes of this phenomenon, yielding two core insights: 1) imbalanced tokenization inherits and can further amplify popularity bias from historical item interactions; 2) current training procedures disproportionately favor popular tokens while neglecting semantic relationships among tokens, thereby intensifying popularity bias. Building on these insights, we propose CRAB, a post-hoc debiasing strategy for GeneRec that reduces popularity bias by rebalancing semantic token frequencies. Specifically, given a well-trained model, we first identify and split over-popular tokens while preserving the overall hierarchical structure of the codebook. Based on the adjusted codebook, we further introduce a hierarchical regularizer to enhance semantic consistency, encouraging more informative representations for unpopular tokens during training. Experiments on real-world datasets demonstrate that CRAB effectively alleviates popularity bias while maintaining competitive performance. †copyright: none 1. Introduction Generative recommendation (GeneRec) (Wang et al., 2024; Han et al., 2025; Liu et al., 2025) has emerged as a promising paradigm for sequential recommendation, demonstrating strong empirical performance across diverse domains (Deng et al., 2025; Kong et al., 2025; Rajput et al., 2023). It formulates recommendation as a sequence-to-sequence generation task: a tokenizer maps each item into a sequence of discrete codebook tokens (Yu et al., 2021), and the model autoregressively predicts the next item based on the concatenated tokenized interaction history (Lv et al., 2024). This design enables modeling over a compact token vocabulary, facilitating efficiency and adaptation to new items. Moreover, powerful sequence models such as LLMs can be naturally incorporated to enhance long-sequence modeling performance. Despite their strong performance, GeneRec often exhibit popularity bias. To quantify this effect, we follow (Lu et al., 2025) and define the top 20%20\% of items by historical interaction frequency as popular items, which collectively account for a disproportionately large share of interactions, and the remaining items as unpopular ones. As shown in the left part of Figure 1, generative recommenders tend to recommend popular items more frequently than traditional methods. Specifically, compared with SASRec, the recommendation frequency of popular items increases by 6.7%6.7\% and 7.2%7.2\% for Mini-OneRec (MOR) (Kong et al., 2025) and TIGER (Rajput et al., 2023), respectively, while the frequency for unpopular items decreases by 3.8%3.8\% and 4.1%4.1\%. However, existing debiasing methods for LLM-based recommendation (Lu et al., 2025; Jiang et al., 2024) do not generalize well to GeneRec. The key reason is that these approaches focus solely on mitigating bias at the model level, while overlooking the bias inherited and amplified by over-popular tokens in the codebook (see Section 3). Although some methods attempt to construct balanced codebooks (Kuai et al., 2024; Deng et al., 2025; Hui et al., 2025), they typically constrain the number of items assigned to each token to prevent highly concentrated mappings, rather than addressing token popularity driven by historical interactions. In this paper, we propose CRAB, a post-hoc approach for mitigating popularity bias in GeneRec. CRAB decouples the debiasing process into two stages: Codebook Rebalancing and Hierarchical Semantic Alignment. In the first stage, we identify over-popular tokens in the existing codebook and split them while preserving the hierarchical semantic structure of the remaining tokens; this process is formulated as a regularized K-means problem. In the second stage, we introduce a hierarchy-aware regularizer to enhance the representation of unpopular tokens by leveraging richer supervision signals from the semantics induced by the codebook. Overall, our key contributions are summarized as follows: • Problem: We present the first investigation demonstrating that imbalanced codebooks give rise to over-popular tokens, which in turn bias the model toward popular items containing these tokens, thereby exacerbating popularity bias. • Method: We propose CRAB, an effective and efficient framework for mitigating popularity bias in GeneRec. It identifies and splits over-popular tokens in the codebook, and further enhances the representations of unpopular tokens through a hierarchical semantic regularizer. • Performance: Extensive experiments on real-world datasets demonstrate that CRAB reduces popularity bias by 16.5% while improving recommendation performance. 2. Background and Preliminaries We consider the sequential recommendation task. Let U denote the set of users and ℐI the set of items. Given a user u∈u with a chronologically ordered interaction history ℋu=i1,…,iTH_u=\i_1,…,i_T\, the goal is to predict the next item iT+1i_T+1. Following recent GeneRec frameworks, each item is first converted into a sequence of discrete semantic tokens from its textual description via residual quantization methods (RQ-KMeans or RQ-VAE). An LLM is then trained to generate the token sequence of the next item. 2.1. Residual Quantization For each item i∈ℐi , its textual description is first passed through a frozen encoder to obtain a continuous embedding i z_i. We aim to quantize it using L hierarchical tokens with L codebooks in a coarse-to-fine generation manner. At the l-th level, the codebook l=kl|k=1,⋯KC^l= \ c_k^l|k=1,·s K \ consists of K codeword embeddings kl c_k^l. When l=1l=1, the initial residual is defined as i1=i r_i^1= z_i. Then at l-th level, the item token sils_i^l can be defined as the index of the closest codeword embedding from the codebook lC^l, and the residual at l+1l+1-th level il+1 r_i^l+1 is updated as follows: (1) sil=argmink‖il−ckl‖2,il+1=il−sill s_i^l= _k|| r_i^l-c^l_k||^2, r_i^l+1= r_i^l- c^l_s_i^l The above process is repeated recursively L times to get a tuple of L codewords that represent the Semantic ID(SID) for item i. In RQ-KMeans, the codeword embedding kl∈l c_k^\,l ^l is defined as the centroid of the k-th cluster obtained by applying K-means to the residual set ℛl=il∣i∈ℐR^l=\ r_i^\,l i \ at level l (Wei et al., 2025). 2.2. Autoregressive Generation By applying the residual quantization to each item in ℋuH_u, the input is transformered into a flattened token sequence X. Accordingly, the target next item iT+1i_T+1 is also encoded as Y: =[s11,s12,…,s1L⏟i1,…,sT1,sT2,…,sTL⏟iT],=[sT+11,sT+12,⋯sT+1L⏟iT+1] X= [ s_1^1,s_1^2,…,s_1^L_i_1,…, s_T^1,s_T^2,…,s_T^L_i_T ], Y=[ s_T+1^1,s^2_T+1,·s s^L_T+1_i_T+1] where sils^l_i denotes the l-th token of item i. The LLM is then trained to predict the SID of Y by minimizing ℒRecL_Rec (2) ℒRec=−∑l=1LlogF(l∣,<l).L_Rec=- _l=1^L F ( Y_l X, Y_<l ). Here, l Y_l denotes the l-th token of Y, and F(⋅)F(·) represents the conditional probability modeled by the LLM. 3. Motivation Figure 1. Left: Popularity bias of GeneRec on the industrial dataset, with the x-axis representing item groups by popularity. Right: The GU of item groups grouped by token popularity. In this section, we empirically analyze how an imbalanced codebook amplifies item popularity bias in historical interactions. With a slight abuse of notation, let cklc_k^l denote the k-th token in the l-th level codebook lC^l. We define token popularity as the total frequency of its associated items in the training data. Specifically, given item frequency fif_i and the item set ℐcklI_c_k^l whose l-th semantic token is cklc_k^l, the popularity score P(ckl)P(c_k^l) is defined as (Zhu et al., 2021; Kokkodis and Lappas, 2020) (3) P(ckl)=∑i∈ℐcklfi,ℐckl=i∈ℐ∣sil=cklP(c_k^l)= _i _c_k^lf_i, _c_k^l=\i s_i^l=c_k^l\ At each level, we rank tokens by popularity and categorize them into the top 5%5\% over-popular tokens TpopT_pop, 5%−95%5\%-95\% neural tokens TneuT_neu and the remaining 5%5\% unpopular tokens TunpT_unp. Items associated with TpopT_pop and TunpT_unp form two groups, denoted as GpopG_pop and GunpG_unp, respectively. We measure Group Unfairness (GU) (Jiang et al., 2024), defined as the discrepancy between recommendation exposure and historical interaction frequency, to quantify bias amplification. As shown in Figure 1, tokens in TpopT_pop exhibit significantly stronger bias amplification. On MOR, the GU gap between GpopG_pop and GunpG_unp reaches 0.420.42, indicating that items associated with popular tokens receive disproportionately higher exposure—1.8×1.8× that of SASRec. Mechanistically, this issue stems from the codebook construction process. Semantically similar items are mapped to the same token; when such items are popular, their interactions accumulate on that token, further increasing its frequency. After training on these token sequences, the recommender becomes biased toward generating items associated with this dominant token, thereby amplifying popularity bias. 4. Method In this section, we introduce the two stages of CRAB respectively. Figure 2. Illustration of CRAB with a three-level codebook in MOR. Over-popular tokens are split by redistributing their child tokens via regularized Kmeans. For clarity, we denote ci1c_i^1, cj2c_j^2, and ck3c_k^3 as AiA_i, BjB_j, and CkC_k, respectively. 4.1. Rebalancing the Codebook The hierarchical token assignment in Equation 1 induces a parent–child relation between tokens at consecutive levels, denoted by ckl→cjl+1\c_k^l→ c_j^l+1\ if there exists at least one item whose l-th and (l+1)(l+1)-th tokens are cklc_k^l and cjl+1c_j^l+1, respectively. Accordingly, for any token ckl∈lc_k^l ^l, its children set is defined as (4) Ch(ckl)=cjl+1∈l+1|ckl→cjl+1.Ch(c_k^l)= \c_j^\,l+1 ^l+1\; |\;c_k^l→ c_j^\,l+1 \. Following Equation 3, we identify over-popular tokens cklc_k^l and split each of them into M new tokens by clustering the residual representations of its associated items. Since codebook tokens encode hierarchical semantic information from textual descriptions, the splitting procedure should ensure semantic coherence within each newly created token, while preserving the semantic integrity of the (l+1)(l\!+\!1)-th level tokens. To this end, we further impose a hard constraint that items sharing the same (l+1)(l\!+\!1)-th level semantic token are assigned to the same new token in the l-th level. Under the induced tree structure shown in Figure 2, this constraint implies that parent tokens are split by redistributing their child tokens, while each child token cjl+1∈Ch(ckl)c_j^l+1 (c_k^l), together with its associated items, remains intact. Moreover, to mitigate over-popular tokens, we aim to approximately balance the popularity scores of the newly generated tokens. Based on the definition of popularity score(Equation 3), we have P(ckl)=∑cjl+1∈Ch(ckl)P(cjl+1)P(c_k^l)= _c^l+1_j (c_k^l)P(c^l+1_j), indicating that the popularity score is preserved between a parent token and its children. Formally, for each cjl+1∈Ch(ckl)c^l+1_j (c_k^l), we introduce a one-hot vector j∈ℝM z_j ^M to indicate the assignment of it to one of the M new tokens. Hence the popularity score for each new token ck(m)c_k(m) can be calculated as follows, and we thereby introducing a balanced loss ℒbalL_bal (5) P(ck(m)l)=∑j=1|Ch(ckl)|j[m]P(cjl+1),ℒbal=∑m=1M(P(ck(m)l)−P¯)2P(c^l_k(m))=Σ^|Ch(c_k^l)| _j=1 z_j[m]P(c^l+1_j),\>L_bal= _m=1^M (P(c^l_k(m))- P )^2 Here j[m]∈0,1 z_j[m]∈0,1 denotes the m-th element of j z_j, and P¯ P represents the average score. Accordingly, splitting an over-popular token cklc_k^l can be formulated as a regularized K-means problem that redistributes its child tokens cjl+1∈Ch(ckl)c_j^\,l+1 (c_k^l) into M new parent tokens. (6) min _ z ∑m=1M∑j=1|Ch(ckl)|j[m]nj‖¯jl−μ¯m‖2+λℒbal,nj=|ℐcjl+1| _m=1^MΣ^|Ch(c_k^l)|_j=1 z_j[m]n_j \| r^l_j- μ_m \|^2+ _bal, n_j=|I_c_j^l+1| s.t. j[m]∈0,1,∑m=1Mj[m]=1,∀cjl+1∈Ch(ckl), z_j[m]∈\0,1\, _m=1^M z_j[m]=1, ∀\,c_j^\,l+1 (c_k^l), Where ¯jl r^l_j denotes the mean residual of items i∈ℐcjl+1i _c_j^l+1 at the l-th level, and μ¯m μ_m is the centroid of the newly formed cluster corresponding to ck(m)c_k(m). Eq. 6 is derived from the variance decomposition in K-means (Blömer et al., 2016): (7) ∑i=1nj‖il−μ¯m‖2=∑i=1nj‖il−¯jl‖2+nj‖¯jl−μ¯m‖2,¯jl=∑i=1njilnj _i=1^n_j\|r^l_i- μ_m\|^2= _i=1^n_j\|r^l_i- r^l_j\|^2+n_j\| r^l_j\ - μ_m\|^2,\> r^l_j= _i=1^n_jr^l_in_j Since all i∈ℐcjl+1i _c_j^l+1 are assigned to same cluster, the first term in Eq. 7 is constant. Hence, the optimization reduces to the second term. The regularized K-means objective can be efficiently optimized using the framework of (Raymaekers and Zamar, 2022). Note that Eq. 5 holds when the tree structure is strictly maintained in the codebook(e.g., RQ-Kmeans) (Deng et al., 2025). To accommodate settings where a child token may have multiple parents (e.g., RQ-VAE) (Kong et al., 2025), we revise Eq. 5 by aggregating frequencies only over items associated with both tokens cjl+1c^l+1_j and cklc_k^l: (8) P(ck(m)l)=∑j=1|Ch(ckl)|j[m]P(cjl+1|ckl),P(cjl+1|ckl)=∑i∈ℐcjl+1∩ℐcklfiP(c^l_k(m))=Σ^|Ch(c_k^l)| _j=1 z_j[m]P(c^l+1_j|c^l_k),\>P(c^l+1_j|c^l_k)= _i _c^l+1_j _c_k^lf_i 4.2. Hierarchical Semantic Alignment Splitting the token cklc_k^l introduces M new tokens. To adapt the LLM to the rebalanced codebook and mitigate bias, we introduce a tree-structure-aware regularizer ℒTL_T that promotes representation consistency among tokens sharing the same parent. (9) ℒT=∑l=1L−1∑k=1|l|1|Ch(ckl)|∑c∈Ch(ckl)‖e(c)−e¯kl‖22 _T=Σ^L-1_l=1Σ^|C^l|_k=1 1|Ch(c^l_k)| _c (c^l_k) \|e(c)- e^l_k \|_2^2 where e(c)e(c) denotes the LLM embedding of token c, and e¯kl e_k^l is the mean embedding of the child tokens of cklc_k^l. The regularizer applies to both new and existing tokens, serving two purposes: (1) enhancing under-represented tokens via supervision from semantically related siblings, and (2) enabling efficient knowledge transfer to newly introduced tokens after rebalancing. Table 1. Performance and Efficiency Comparison on Industrial and Office Datasets. Industrial Office Metric MOR Tiger RW R D2LR CRAB MOR Tiger RW R D2LR CRAB HR@10 ↑ 0.152 0.132 0.126 0.141 0.147 0.152 0.161 0.137 0.131 0.146 0.153 0.160 NDCG@10 ↑ 0.116 0.090 0.070 0.107 0.113 0.117 0.122 0.100 0.090 0.119 0.119 0.122 DGU@10 ↓ 0.418 0.423 0.367 0.406 0.410 0.356 0.423 0.427 0.386 0.410 0.414 0.368 MGU@10 ↓ 0.109 0.112 0.105 0.109 0.106 0.091 0.111 0.113 0.108 0.110 0.110 0.093 Time (h) ↓ - - 3.11 0.21 2.75 0.28 - - 3.25 0.28 3.16 0.38 4.3. Model Optimization To mitigate popularity bias in GeneRec, we jointly optimize the recommendation loss ℒRecL_Rec and the hierarchical regularizer ℒTL_T: (10) ℒ=ℒRec+γℒT =L_Rec+ _T where γ is a hyperparameter controlling the strength of regularization. During optimization, we update the embedding layers for both existing and newly introduced tokens. To improve efficiency, LoRA adapters are applied only to the attention layers of the LLM. 5. Experiments In this section, we evaluate the performance of CRAB. 5.1. Experimental Settings Datasets Experiments are carried out on two real-world datasets Office and Industrial. Following (Kong et al., 2025), we first filter out users and items with fewer than five interactions. For each dataset, interactions are split chronologically into training, validation, and test sets with an 8:1:1 ratio. Evaluation Metrics Following (Kong et al., 2025), we evaluate recommendation performance using NDCG@K and HR@K, and measure bias amplification with MGU@K and DGU@K (Jiang et al., 2024). We also report time cost to assess efficiency. Baselines To evaluate CRAB, we compare it against two backbone generative models: Tiger (Rajput et al., 2023) and MiniOneRec (MOR) (Kong et al., 2025). We also implement three SOTA popularity debiasing methods atop MOR for a fair comparison: (1) Reweighting (Jiang et al., 2024) balances loss contribution via item popularity; (2) Reranking (Jiang et al., 2024) penalizes popular items during post-processing; and (3) D2LR (Lu et al., 2025) employs propensity score weighting for LLMs. Implementation We implement CRAB based on the MOR framework using Qwen2-0.5B as the backbone. The model is trained on 4× NVIDIA A100 GPUs for 10 epochs. We employ the AdamW optimizer with a global batch size of 128. The learning rate is set to 1×10−41× 10^-4 with a weight decay of 0.01. To efficiently train the model, we utilize LoRA (Hu et al., 2022) with the rank r=8r=8, α=16α=16. For our codebook rebalancing strategy, we perform a hierarchical split at each level of the codebook with a 10% splitting ratio. The number of new tokens M is determined by the ratio between the frequency of the target token and the average frequency at the same layer, with an upper bound of M≤3M≤ 3. 5.2. Experimental Results Overall Performance We compare CRAB with the baselines in Section 5.1, as shown in Table 1. Overall, CRAB achieves performance comparable to MOR while significantly mitigating popularity bias. Specifically, on the Industrial dataset, CRAB improves HR@10 by 15.2%, 7.8%, and 3.4% over Tiger, RW, and R, respectively, demonstrating its ability to alleviate bias without sacrificing recommendation quality. In terms of debiasing effectiveness, CRAB achieves the best results, reducing DGU@10 by 14.8% compared to MOR and 13.2% compared to D2LR, while lowering MGU@10 by 16.5% and 14.2%, respectively. Although RW also mitigates bias, it causes severe performance degradation. In contrast, CRAB achieves a better trade-off between accuracy and fairness. Moreover, CRAB is highly efficient, requiring only about 1/11 and 1/10 of the training time of RW and D2LR. Although R is more efficient due to re-ranking, CRAB achieves superior performance and fairness. Figure 3. Left: MOR performance under different splitting ratios (5%5\%–20%20\%) at each level on the Industrial dataset. Right: MOR performance when splitting the top 5%5\% most popular tokens at levels A, B, and C separately. Figure 4. Left: Effect of γ Right: Effect of LoRA 5.3. In-depth Analysis Impact of Splitting Ratio We analyze how the proportion of tokens to split at each level affects CRAB. As shown in the left part of Figure 3, as the proportion increases, the performance first improves and then declines, showing a similar trend to bias amplification. This suggests that splitting a small fraction(less than 10%10\%) of popular tokens can smooth token popularity and enhance the representation of unpopular tokens, thereby improving long-tail items exposure. However, over-splitting tokens may disrupt the semantic integrity of the codebook and lead to performance degradation. Impact of Splitting Position We analyze splitting at different levels by splitting only the top-5%5\% popular tokens from level A, B, and C in MOR separately. From the right part of Figure 3, splitting over-popular tokens at level B improves representation while significantly mitigating bias, suggesting that intermediate-level tokens concentrate excessive semantic information, consistent with the “Hourglass” phenomenon (Kuai et al., 2024). In contrast, splitting only the last level may hurt performance, as its semantics are already fine-grained. Ablation Study We first examine the effect of the hyperparameter γ, which balances the recommendation objective and representation consistency. As shown in Figure 4(left), as γ increases, CRAB places more emphasis on ℒTL_T. When γ≤0.2γ≤ 0.2, the recommendation performance remains stable. However, for γ>0.2γ>0.2, NDCG drops sharply. To achieve a proper trade-off, we set γ=0.2γ=0.2. We then evaluate the necessity of LoRA and observe that removing LoRA leads to noticeable drops in both NDCG and HR from the Figure. 6. Conclusion In this paper, we present the first systematic investigation of popularity bias in GeneRec and propose CRAB, a novel method that mitigates bias by rebalancing the codebook. References (1) Blömer et al. (2016) Johannes Blömer, Christiane Lammersen, Melanie Schmidt, and Christian Sohler. 2016. Theoretical analysis of the k-means algorithm–a survey. In Algorithm Engineering: Selected Results and Surveys. Springer, 81–116. Deng et al. (2025) Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian 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). Han et al. (2025) Ruidong Han, Bin Yin, Shangyu Chen, He Jiang, Fei Jiang, Xiang Li, Chi Ma, Mincong Huang, Xiaoguang Li, Chunzhen Jing, et al. 2025. Mtgr: Industrial-scale generative recommendation framework in meituan. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management. 5731–5738. Hu et al. (2022) Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR 1, 2 (2022), 3. Hui et al. (2025) Zheng Hui, Xiaokai Wei, Reza Shirkavand, Chen Wang, Weizhi Zhang, Alejandro Peláez, and Michelle Gong. 2025. Semantics Meet Signals: Dual Codebook Representationl Learning for Generative Recommendation. arXiv preprint arXiv:2511.20673 (2025). Jiang et al. (2024) Meng Jiang, Keqin Bao, Jizhi Zhang, Wenjie Wang, Zhengyi Yang, Fuli Feng, and Xiangnan He. 2024. Item-side fairness of large language model-based recommendation system. In Proceedings of the ACM Web Conference 2024. 4717–4726. Kokkodis and Lappas (2020) Marios Kokkodis and Theodoros Lappas. 2020. Your hometown matters: Popularity-difference bias in online reputation platforms. Information Systems Research 31, 2 (2020), 412–430. Kong et al. (2025) Xiaoyu Kong, Leheng Sheng, Junfei Tan, Yuxin Chen, Jiancan Wu, An Zhang, Xiang Wang, and Xiangnan He. 2025. Minionerec: An open-source framework for scaling generative recommendation. arXiv preprint arXiv:2510.24431 (2025). Kuai et al. (2024) Zhirui Kuai, Zuxu Chen, Huimu Wang, Mingming Li, Dadong Miao, Wang Binbin, Xusong Chen, Li Kuang, Yuxing Han, Jiaxing Wang, et al. 2024. Breaking the Hourglass Phenomenon of Residual Quantization: Enhancing the Upper Bound of Generative Retrieval. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 677–685. Liu et al. (2025) Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al. 2025. Onerec-think: In-text reasoning for generative recommendation. arXiv preprint arXiv:2510.11639 (2025). Lu et al. (2025) Sijin Lu, Zhibo Man, Fangyuan Luo, and Jun Wu. 2025. Dual Debiasing in LLM-based Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2685–2689. Lv et al. (2024) Zheqi Lv, Shaoxuan He, Tianyu Zhan, Shengyu Zhang, Wenqiao Zhang, Jingyuan Chen, Zhou Zhao, and Fei Wu. 2024. Semantic codebook learning for dynamic recommendation models. In Proceedings of the 32nd ACM International Conference on Multimedia. 9611–9620. Rajput et al. (2023) 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. Raymaekers and Zamar (2022) Jakob Raymaekers and Ruben H Zamar. 2022. Regularized k-means through hard-thresholding. Journal of Machine Learning Research 23, 93 (2022), 1–48. Wang et al. (2024) 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. Wei et al. (2025) Tianxin Wei, Xuying Ning, Xuxing Chen, Ruizhong Qiu, Yupeng Hou, Yan Xie, Shuang Yang, Zhigang Hua, and Jingrui He. 2025. CoFiRec: Coarse-to-Fine Tokenization for Generative Recommendation. arXiv preprint arXiv:2511.22707 (2025). Yu et al. (2021) Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. 2021. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627 (2021). Zhu et al. (2021) Ziwei Zhu, Yun He, Xing Zhao, Yin Zhang, Jianling Wang, and James Caverlee. 2021. Popularity-opportunity bias in collaborative filtering. In Proceedings of the 14th ACM international conference on web search and data mining. 85–93.