Paper deep dive
Cross-Country Code-Mixing for Generative Recommendation
Yuan Gao, Hao Deng, Haibo Xing, Yi Xu, Lingyu Mu, Jinxin Hu, Yu Zhang, Xiaoyi Zeng
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/25/2026, 4:32:31 AM
Summary
The paper introduces CMRec, a cross-country generative recommendation framework that addresses the lack of shared user/item IDs across different e-commerce markets. By creating a shared semantic codebook from multi-modal content and behavioral co-occurrence, CMRec synthesizes mixed-country user behavior sequences via token-level substitutions constrained by static content and dynamic market attributes. A context-aware loss function reweights these mixed samples to ensure plausibility. Experiments on industrial and Amazon M2 datasets demonstrate significant improvements in recommendation quality and business metrics like advertising revenue and orders.
Entities (10)
Relation Signals (8)
CMRec → developedby → Alibaba International Digital Commerce Group
confidence 95% · Affiliation: Alibaba International Digital Commerce Group... authors contributed equally
CMRec → employs → Code-Mixing
confidence 95% · we propose CMRec... that injects cross-country supervision at the data level via dual-constrained, context-aware code-mixing.
CMRec → evaluatedon → Amazon M2
confidence 95% · We evaluate CMRec on two multi-country datasets... The second is Amazon M2
CMRec → uses → Shared Semantic Codebook
confidence 95% · CMRec first learns a shared semantic codebook... It then uses this codebook to synthesize mixed-country sequences
Shared Semantic Codebook → derivedfrom → Multi-modal Content
confidence 90% · learns a shared semantic codebook from multi-modal content and behavioral co-occurrence
Shared Semantic Codebook → derivedfrom → Behavioral Co-occurrence
confidence 90% · learns a shared semantic codebook from multi-modal content and behavioral co-occurrence
Generative Recommendation → mitigates → Disjoint ID Spaces
confidence 90% · Generative recommendation (GR) mitigates this by mapping items into a shared token space
CMRec → outperforms →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Cross-country recommendation on modern e-commerce platforms is typically deployed with disjoint user and item ID spaces across markets, removing the shared anchors that conventional cross-domain methods rely on. Generative recommendation (GR) mitigates this by mapping items into a shared token space and training a unified model, but existing approaches keep behavior sequences strictly country-specific, so knowledge transfer occurs only at the parameter level and remains absent at the data level. Inspired by code-switching corpora in multilingual natural language processing, we propose CMRec, a cross-country GR framework that injects cross-country supervision at the data level via dual-constrained, context-aware code-mixing. CMRec first learns a shared semantic codebook from multi-modal content and behavioral co-occurrence across countries. It then uses this codebook to synthesize mixed-country sequences via token-level substitutions that satisfy both static (content) and dynamic (e.g., price, audience, popularity) constraints. Finally, it introduces a context-aware loss that reweights mixed samples according to their plausibility in the current sequence. Experiments on two real-world multi-country datasets and an online A/B test show that CMRec substantially improves recommendation quality in data-sparse countries while preserving performance in data-rich countries, achieving +1.77% advertising revenue and +2.64% orders on a large-scale e-commerce platform.
Tags
Links
- Source: https://arxiv.org/abs/2609.28972v1
- Canonical: https://arxiv.org/abs/2609.28972v1
Trouble viewing inline? Open PDF directly →
Full Text
37,875 characters extracted from source content.
Expand or collapse full text
Cross-Country Code-Mixing for Generative 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: 10.1145/3799682.3839862ISBN: 979-8-4007-2539-5/2026/11CCS: Information systems Recommender systems Yuan Gao Note: Both authors contributed equally to this research. email: jianbei.gy@alibaba-inc.com Affiliation: Alibaba International Digital Commerce Group, Beijing, China , Hao Deng email: denghao.deng@alibaba-inc.com Affiliation: Alibaba International Digital Commerce Group, Beijing, China , Haibo Xing email: xinghaibo.xhb@alibaba-inc.com Affiliation: Alibaba International Digital Commerce Group, Beijing, China , Yi Xu email: xy397404@alibaba-inc.com Affiliation: Alibaba International Digital Commerce Group, Beijing, China , Lingyu Mu email: moulingyu.mly@alibaba-inc.com Affiliation: Alibaba International Digital Commerce Group, Beijing, China , Jinxin Hu Note: Corresponding author. email: jinxin.hjx@alibaba-inc.com Affiliation: Alibaba International Digital Commerce Group, Beijing, China , Yu Zhang email: daoji@alibaba-inc.com Affiliation: Alibaba International Digital Commerce Group, Beijing, China and Xiaoyi Zeng email: yuanhan@taobao.com Affiliation: Alibaba International Digital Commerce Group, Hangzhou, China © c Abstract. Cross-country recommendation on modern e-commerce platforms is typically deployed with disjoint user and item ID spaces across markets, removing the shared anchors that conventional cross-domain methods rely on. Generative recommendation (GR) mitigates this by mapping items into a shared token space and training a unified model, but existing approaches keep behavior sequences strictly country-specific, so knowledge transfer occurs only at the parameter level and remains absent at the data level. Inspired by code-switching corpora in multilingual natural language processing, we propose CMRec, a cross-country GR framework that injects cross-country supervision at the data level via dual-constrained, context-aware code-mixing. CMRec first learns a shared semantic codebook from multi-modal content and behavioral co-occurrence across countries. It then uses this codebook to synthesize mixed-country sequences via token-level substitutions that satisfy both static (content) and dynamic (e.g., price, audience, popularity) constraints. Finally, it introduces a context-aware loss that reweights mixed samples according to their plausibility in the current sequence. Experiments on two real-world multi-country datasets and an online A/B test show that CMRec substantially improves recommendation quality in data-sparse countries while preserving performance in data-rich countries, achieving +1.77% advertising revenue and +2.64% orders on a large-scale e-commerce platform. Keywords: Generative Recommendation, Cross Country, Code Mixing †c-license: by 1. Introduction Two-panel schematic. Panel (a) shows the cross-country setting: several country-specific markets with disjoint user and item identifier spaces, whose items are quantized into one shared semantic token space. Panel (b) shows a user behaviour sequence in which one item is replaced by its cross-country counterpart through token-level substitution, together with the two failure cases this can cause: substitutions that violate static content attributes, and substitutions that violate time-varying attributes such as price, audience and popularity. Figure 1. (a) Cross-country setting. (b) Example and challenges of cross-country behavior-sequence code-mixing.Two-panel schematic. Panel (a) shows the cross-country setting: several country-specific markets with disjoint user and item identifier spaces, whose items are quantized into one shared semantic token space. Panel (b) shows a user behaviour sequence in which one item is replaced by its cross-country counterpart through token-level substitution, together with the two failure cases this can cause: substitutions that violate static content attributes, and substitutions that violate time-varying attributes such as price, audience and popularity. Modern e-commerce platforms increasingly operate across many countries (Xing et al., 2025; Yang et al., 2024; Li and Lyu, 2025). In practice, each country is deployed and logged as an independent system: user interactions are recorded only within a single country, and both user and item ID spaces are maintained separately (Starke et al., 2025). Consequently, both IDs and behavior sequences are strictly partitioned by country, with no natural overlap across countries, as illustrated in Figure 1 (a). This setup differs from the well-studied cross-domain recommendation (CDR) setting (Li et al., 2025a; Hu et al., 2026a; Zang et al., 2022; Zhang et al., 2023; Zhang et al., 2025), where knowledge transfer typically relies on some form of overlap across domains, such as shared users, shared items, or manually anchored entities that carry collaborative signals. Recent CDR methods based on generative recommendation (GR), such as GenCDR (Hu et al., 2026b) and GMC (Jin et al., 2025), relax this ID-level requirement by mapping items from different domains into a shared semantic token space and training a unified generative model. In this way, knowledge transfer occurs at the parameter level, mediated by the shared backbone and token space. However, GR learns primarily from item co-occurrence within training sequences (Hou et al., 2025; Li et al., 2025b; Huang et al., 2026). In the cross-country setting, these sequences remain strictly partitioned by country, so items from one country never appear in another country’s context, leaving cross-country reinforcement absent at the data level. Consequently, rich interaction signals in large markets cannot benefit small ones, and the potential for cross-country knowledge sharing remains largely untapped. A natural reference comes from multilingual natural language processing (NLP), where models learn cross-lingual alignment from code-switched corpora, in which tokens from different languages appear in the same training sequence (Blevins and Zettlemoyer, 2022; Briakou et al., 2023; Li et al., 2024). Motivated by this, we investigate whether an analogous construction is possible for cross-country GR, by building mixed-country sequences that interleave items from different markets, as illustrated in Figure 1 (b). However, no such corpus exists naturally; it must be synthesized by substitution. This raises a recommendation-specific question: what counts as a valid substitute? Existing substitution-based augmentation methods (Cui et al., 2025; Wang et al., 2023; Liu et al., 2025), mostly developed for CDR settings, match items by content (e.g., titles, descriptions, images) and replace them in user sequences. Such recipes can in principle be applied to cross-country GR, but they fall short along two complementary axes. (a) Substitution criteria are incomplete at the item level. Content similarity alone ignores two critical signals: behavioral (e.g., co-purchase and co-click patterns) and market (e.g., price, audience, popularity). Two items that look similar can play very different roles across markets (Starke et al., 2025), so a content-only criterion injects substantial noise. (b) Substitution granularity is too coarse at the context level. Even if a candidate is reasonable in isolation, it may still conflict with the user’s current intent. For example, in a sequence dominated by Apple devices, swapping an item with an Android pad that looks similar in content and basic attributes is syntactically plausible but contextually wrong, and this cannot be detected by any offline item-pair filter. To address these limitations, we propose CMRec, a cross-country GR framework that injects cross-country supervision at the data level via dual-constrained, context-aware code-mixing. CMRec first learns a shared semantic codebook from multi-modal content and behavioral co-occurrence across countries, which serves as both the GR tokenizer and a common semantic space for cross-country substitution. On top of this codebook, we perform token-level, dual-constrained code-mixing: substitutions are applied to semantic token sequences rather than country-specific item IDs, and candidates must satisfy both static (content) and dynamic (e.g., price, audience, popularity) constraints, enabling finer-grained, market-aware cross-country mixing. Finally, we introduce a context-aware loss that lets the GR model reweight mixed samples according to their plausibility in the current context, downweighting noisy substitutions while preserving useful cross-country signals. We conduct experiments on two real-world multi-country datasets and an online A/B test, achieving improvements of +1.77% advertising revenue and +2.64% orders, demonstrating the practical effectiveness of CMRec in production. 2. Problem Formulation We consider cross-country recommendation over a country set C, where each c∈c is an independent domain with disjoint user and item sets (c)U^(c) and ℐ(c)I^(c) (i.e., (c)∩(c′)=ℐ(c)∩ℐ(c′)=∅U^(c) ^(c )=I^(c) ^(c )= for c≠c′c≠ c ). For each user u∈(c)u ^(c), we denote the interaction sequence as xc=(i1,…,iT)x_c=(i_1,…,i_T) with it∈ℐ(c)i_t ^(c), and the target item as yc∈ℐ(c)y_c ^(c). Following the GR paradigm (Rajput et al., 2023; Xing et al., 2025), we map each item to a discrete token sequence via a codebook and formulate recommendation as conditional sequence generation in this unified token space. Let (c)=(xc,yc)D^(c)=\(x_c,y_c)\ collect the per-country training pairs from users u∈(c)u ^(c). For joint multi-country training, we pool samples from all countries into a unified set =⋃c∈(c)=(x,y)D= _c D^(c)=\(x,y)\, on which the GR model is trained to maximize (1) ℒGR(θ)=∑(x,y)∈logPθ(y∣x),L_GR(θ)= _(x,y) P_θ(y x), where θ are the model parameters. 3. Method Three-part pipeline diagram of the proposed framework. Part 1 trains a shared semantic codebook from multimodal item content and behavioural item-to-item co-occurrence, turning every item of every country into a token sequence that the generative recommendation model consumes. Part 2 uses that codebook together with side information to pick cross-country neighbour items and graft them into a user behaviour sequence, producing a code-mixed twin sequence and its rewritten target. Part 3 feeds both the original and the mixed sequence to the generative recommendation model itself, which scores their plausibility under the same user context and reweights each mixed sample so that noisy substitutions contribute less to the training loss. Figure 2. Overview of CMRec . Part 1 learns a shared semantic codebook from multimodal content and behavioral i2i co-occurrence, producing token sequences iq_i that are used by the GR model. Part 2 uses the codebook plus side-info to select cross-country neighbors and grafts them into the user sequence, yielding a code-mixed twin (′,y′)(x ,y ). Part 3 lets the GR model itself reweight each twin by its plausibility under the original context, downweighting noisy substitutions during training.Three-part pipeline diagram of the proposed framework. Part 1 trains a shared semantic codebook from multimodal item content and behavioural item-to-item co-occurrence, turning every item of every country into a token sequence that the generative recommendation model consumes. Part 2 uses that codebook together with side information to pick cross-country neighbour items and graft them into a user behaviour sequence, producing a code-mixed twin sequence and its rewritten target. Part 3 feeds both the original and the mixed sequence to the generative recommendation model itself, which scores their plausibility under the same user context and reweights each mixed sample so that noisy substitutions contribute less to the training loss. We train a unified cross-country GR model that shares information across countries while respecting country-specific dynamics. To this end, CMRec comprises three components: (i) a shared semantic codebook that tokenizes items from all countries using behavior and content signals; (i) a dual-constrained code-mixing mechanism that substitutes tokens across countries under static and dynamic constraints; and (i) a context-aware loss reweighting scheme that downweights noisy mixed samples during training. 3.1. Behavior- and Content-Grounded Shared Codebook To enable a unified cross-country GR backbone, we first construct a shared semantic codebook that is informed by both multi-modal content and behavioral signals. This codebook serves two roles at once: it tokenizes items for GR and provides a common semantic space for cross-country token substitution. We first encode each item’s multimodal side information (e.g., title, image) into a content embedding im_i using a pretrained model (Bai et al., 2025) shared across countries. This provides a unified semantic space where items from all countries are comparable at the content level. However, content alone does not capture how items actually behave in each country: items with similar visuals can have very different popularity or co-occurrence patterns, and these behavior signals are implicitly embedded in user interaction sequences. (McAuley et al., 2015; Yuan et al., 2023) To inject such behavioral information, we further train a single behavior-driven item-to-item (i2i) alignment model jointly across countries. On top of im_i, we adopt a two-tower architecture with sampled softmax (Deng et al., 2025; Xu et al., 2022): the user tower maps a behavior sequence to a user embedding iu_i, and the item tower maps im_i to an item embedding i=T(i)e_i=T(m_i). The joint i2i training objective is (2) ℒi2i=−∑ilogexp(i⊤i+)exp(i⊤i+)+∑j∈iexp(i⊤j),L_i2i=- _i (u_i e_i^+) (u_i e_i^+)+ _j _i (u_i e_j), where the negative pool iN_i is shared across countries (Huang et al., 2020), so ie_i aligns items that are both content-similar and behaviorally similar in user interaction patterns. We then apply a residual quantizer (e.g., RQ-VAE (Rajput et al., 2023)) to transform i\e_i\ into shared token sequences i=(qi(1),…,qi(M))q_i=(q_i^(1),…,q_i^(M)). These tokens form a unified codebook that we use both as the GR tokenizer and as the substrate for cross-country code-mixing. 3.2. Dual-Constrained Cross-Country Code-Mixing With the shared codebook in place, items from different countries can be matched in a common token space. We now introduce cross-country code-mixing at the token level to expose GR to mixed sequences across countries. A naive approach would substitute items based only on their codebook representations. However, the codebook is updated only infrequently and does not capture fast-changing attributes such as price discount, popularity, and audience tags, which are crucial for deciding whether two items are interchangeable in practice. To address this, we adopt a dual-constrained filter. Each item i is equipped with a side-information embedding is_i over these dynamic attributes, with each element discretized by per-country quantile bucketization so that values from different countries are placed on a comparable scale, and we define the cross-country candidate set as (3) ^τ,τs(i)=j∈⋃c′≠cℐ(c′)|d(i,j)≤τ∧sim(i,j)≥τs, N_τ, _s(i)= \j∈\!\! _c ≠ c\!\!I^(c )\, |\,d(q_i,q_j)≤τ\; \;sim(s_i,s_j)≥ _s \, where d is the Hamming distance (Chan et al., 2020) between token sequences and simsim is cosine similarity, enforcing static and dynamic consistency. Given a training pair (,y)(x,y), the operator Φτ,τs,k:(,y)↦(′,y′) _τ, _s,k:(x,y) (x ,y ) replaces the items at a rate-k fraction of slots in x and y with neighbors drawn from ^τ,τs(⋅) N_τ, _s(·), leaving the rest intact while injecting cross-country co-occurrence at the token level. We then apply Φ to a rate-p subsample of D to form the twin set ′D , with p controlling the cross-country mixing strength. 3.3. Context-Aware Adaptive Reweighting The filter ^τ,τs N_τ, _s ensures item-level plausibility, but it is blind to the user context and to how the model’s predictions evolve during training. Some synthetic substitutions can still be incompatible with the surrounding sequence, especially in long and noisy histories. We therefore let the GR model itself assess each substituted item during training. For a substituted example (′,y′)(x ,y ) derived from an original pair (,y)(x,y), we define the context-conditional substitutability as (4) w(y,y′∣)=sg[Pθ(y′∣)Pθ(y∣)+Pθ(y′∣)]∈[0,1],w(y,y )=sg\! [ P_θ(y )P_θ(y )+P_θ(y ) ]∈[0,1], where Pθ(⋅∣)P_θ(· ) is the autoregressive item likelihood under the original context and sg[⋅]sg[·] denotes stop-gradient, so w affects the loss but does not receive gradients. We deliberately evaluate y′y under the original context x rather than the mixed ′x , since PθP_θ is better calibrated on the observed x; this prevents w from being dominated by the model’s unstable response to out-of-distribution synthetic substitutions early in training. We then combine original and substituted supervision in the training objective: (5) ℒ(θ)=−∑(,y)∈logPθ(y∣)−λ∑(′,y′)∈′w(y,y′∣)logPθ(y′∣′),L(θ)=- _(x,y) P_θ(y )-λ _(x ,y ) w(y,y ) P_θ(y ), where λ controls the overall contribution of substituted examples. Because w is derived from the model’s own predictions, it acts as a context-aware weight: early in training most substitutions receive similar weights, while later the model downweights contextually invalid substitutes that the item-level filter cannot distinguish. 4. Experiments Table 1. Performance comparison on the industrial and Amazon M2 datasets. Each metric reports the 6-country/locale average (Avg6) plus three representative ones, with superscripts L/M/S marking large/medium/small volume (Amazon M2 has no M tier). Best results are in bold and second-best are underlined. “Improv.” shows the relative improvement (%) over the best baseline. Method Recall@10 Recall@100 NDCG@10 NDCG@100 (a) Industrial Dataset (6 countries) Avg6 AL BM CS Avg6 AL BM CS Avg6 AL BM CS Avg6 AL BM CS SASRec 0.1882 0.1993 0.1867 0.1705 0.2648 0.2807 0.2632 0.2404 0.0941 0.0995 0.0933 0.0853 0.1458 0.1546 0.1448 0.1322 HSTU 0.2117 0.2244 0.2098 0.1929 0.2983 0.3158 0.2957 0.2725 0.1058 0.1122 0.1049 0.0964 0.1642 0.1738 0.1629 0.1498 TIGER 0.2261 0.2396 0.2243 0.2062 0.3187 0.3377 0.3163 0.2909 0.1131 0.1198 0.1122 0.1031 0.1753 0.1858 0.1739 0.1600 REG4Rec 0.2372 0.2517 0.2354 0.2163 0.3342 0.3537 0.3318 0.3057 0.1187 0.1258 0.1177 0.1082 0.1838 0.1949 0.1824 0.1681 GenCDR 0.2423 0.2493 0.2407 0.2247 0.3413 0.3562 0.3393 0.3168 0.1212 0.1246 0.1204 0.1123 0.1877 0.1961 0.1866 0.1743 GenCDR+Cont 0.2404 0.2456 0.2371 0.2311 0.3432 0.3496 0.3359 0.3248 0.1202 0.1227 0.1186 0.1156 0.1891 0.1928 0.1847 0.1793 GenCDR+CF 0.2384 0.2528 0.2378 0.2256 0.3412 0.3575 0.3387 0.3179 0.1182 0.1264 0.1192 0.1109 0.1871 0.1969 0.1853 0.1768 CMRec 0.2496 0.2567 0.2494 0.2434 0.3613 0.3672 0.3567 0.3474 0.1236 0.1278 0.1239 0.1198 0.1972 0.2018 0.1957 0.1919 Improv. +3.01% +1.54% +3.61% +5.32% +5.27% +2.71% +5.13% +6.96% +1.98% +1.11% +2.91% +3.63% +4.28% +2.49% +4.88% +7.03% (b) Amazon M2 Dataset (6 locales) Avg6 UKL FRS ESS Avg6 UKL FRS ESS Avg6 UKL FRS ESS Avg6 UKL FRS ESS SASRec 0.2774 0.2293 0.3028 0.3095 0.4823 0.3986 0.5274 0.5389 0.1388 0.1147 0.1516 0.1548 0.2653 0.2192 0.2898 0.2963 HSTU 0.3212 0.2654 0.3506 0.3584 0.5584 0.4618 0.6099 0.6236 0.1606 0.1327 0.1753 0.1792 0.3070 0.2539 0.3354 0.3428 TIGER 0.3406 0.2816 0.3719 0.3803 0.5926 0.4898 0.6471 0.6616 0.1703 0.1408 0.1861 0.1902 0.3258 0.2694 0.3559 0.3638 REG4Rec 0.3523 0.2916 0.3846 0.3931 0.6124 0.5066 0.6688 0.6843 0.1762 0.1458 0.1921 0.1964 0.3369 0.2786 0.3679 0.3764 GenCDR 0.3603 0.2894 0.3928 0.4009 0.6266 0.5186 0.6839 0.6994 0.1798 0.1447 0.1963 0.2004 0.3446 0.2852 0.3761 0.3848 GenCDR+Cont 0.3579 0.2862 0.3907 0.3991 0.6291 0.5149 0.6893 0.7063 0.1784 0.1432 0.1951 0.1997 0.3463 0.2831 0.3793 0.3884 GenCDR+CF 0.3559 0.2922 0.3884 0.3961 0.6271 0.5198 0.6862 0.7035 0.1764 0.1465 0.1942 0.1985 0.3443 0.2860 0.3781 0.3865 CMRec 0.3748 0.2997 0.4137 0.4203 0.6671 0.5379 0.7432 0.7588 0.1847 0.1484 0.2033 0.2068 0.3637 0.2953 0.4058 0.4136 Improv. +4.02% +2.57% +5.32% +4.84% +6.04% +3.48% +7.82% +7.43% +2.73% +1.30% +3.57% +3.19% +5.03% +3.25% +6.99% +6.49% 4.1. Experimental Settings Datasets & Evaluation Metrics. We evaluate CMRec on two multi-country datasets with highly imbalanced data distributions. The first is an in-house industrial advertising dataset from an e-commerce platform. It contains over 1 billion interactions from 19 million users across 25 million advertisements across 6 countries (A–F). The largest country contributes about 10× more training interactions than the smallest, and user and item IDs are completely disjoint across countries. The second is Amazon M2 (Jin et al., 2023), the public KDD Cup 2023 multilingual session benchmark covering 6 locales (high-resource: UK/DE/JP; low-resource: IT/FR/ES). The high-resource locales account for roughly 91% of all training sessions, yielding a pronounced long-tail distribution. For offline evaluation, we report Recall@10/100 and NDCG@10/100 (Deng et al., 2025; Rajput et al., 2023) for each representative country/locale, and also report averages over all countries/locales. Baselines & implementation. We compare with five representative methods plus two augmented variants: SASRec (Kang and McAuley, 2018), GenCDR (Hu et al., 2026b), REG4Rec (Xing et al., 2025), HSTU (Zhai et al., 2024), and TIGER (Rajput et al., 2023). To study single-signal data augmentation, we follow SRA-CL (Cui et al., 2025) and build a content-augmented variant on top of GenCDR, denoted GenCDR+Cont. We also follow CoSeRec (Liu et al., 2021) and TiCoSeRec (Dang et al., 2023) to build a collaboration-augmented variant based on user-behavior signals, denoted GenCDR+CF. Both augmented variants use the same substitution ratio and sampling budget as CMRec for a controlled comparison. All baselines follow their original designs and are trained on the same dataset partition. For CMRec, we first encode each item’s content into a 64-dimensional vector using Qwen-VL-32B (Bai et al., 2025). We then tokenize this vector with an RQ-VAE (Rajput et al., 2023) into a token sequence of length L=4L=4. We adopt GenCDR as the backbone. We set p=0.1p=0.1 and k=0.1k=0.1 to balance performance and computational cost. We further set λ=0.5λ=0.5 and use relatively strict filtering thresholds, τ=1τ=1 and τs=0.8 _s=0.8. 4.2. Experimental Results Table 2. Ablation on the Industrial dataset (6-country avg.). Numbers in parentheses are relative gaps vs. Full. Variant Recall@100 NDCG@100 CMRec (Full) 0.3613 0.1972 w/o Token-level distance constraint 0.3447 (−-4.59%) 0.1894 (−-3.96%) w/o Side-info Similarity constraint 0.3573 (−-1.11%) 0.1947 (−-1.27%) w/o Context-aware Adaptive Reweighting 0.3531 (−-2.27%) 0.1928 (−-2.23%) Overall performance. Table 1 summarizes the main results, from which we draw three observations. (1) CMRec achieves the best Recall@10/100 and NDCG@10/100 on both datasets on all six countries, outperforming all baselines. (2) Token-based GR models (REG4Rec, GenCDR) outperform ID-based models (SASRec, HSTU), showing that tokenization and cross-country parameter sharing provide a stronger backbone, especially for small countries. (3) Single-signal augmentations based only on content or collaboration improve some countries but degrade others, while CMRec combines dual-constrained code-mixing with context-aware reweighting and delivers more consistent gains across all countries. Ablation study. We conduct an ablation study on three components of CMRec in Table 2: the token-level distance constraint, the side-information similarity constraint, and the context-aware reweighting. The token-level distance constraint contributes the largest improvement, indicating that codebook proximity is crucial for transferring cross-country behavioral structure. The side-information similarity constraint further refines substitutions by enforcing multi-modal coherence. Removing adaptive reweighting leads to clear degradation, showing that it effectively filters samples whose mixed tokens are incompatible with the original context. Hyperparameter sensitivity and robustness. To validate sensitivity and robustness of our method, we sweep (p,k)(p,k) for our dual-constrained code-mixing on two backbones (Figure 3). On both backbones, all (p,k)(p,k) combinations consistently outperform the un-augmented baseline, and the default (p=10%,k=10%)(p=10\%,\,k=10\%) achieves near-optimal Recall@100 in both cases, confirming that the augmentation is robust to both hyperparameter and backbone choice. Line charts of Recall at 100, averaged over the six countries of the industrial dataset, as the substitution ratio p and the neighbour pool size k are swept on two generative recommendation backbones. Every combination of p and k lies above the dashed horizontal line that marks the un-augmented baseline, and the default setting of p equal to ten percent and k equal to ten percent is close to the best value on both backbones. Figure 3. Sensitivity to p and k on Industrial (Recall@100, Avg6). Dashed line = baseline.Line charts of Recall at 100, averaged over the six countries of the industrial dataset, as the substitution ratio p and the neighbour pool size k are swept on two generative recommendation backbones. Every combination of p and k lies above the dashed horizontal line that marks the un-augmented baseline, and the default setting of p equal to ten percent and k equal to ten percent is close to the best value on both backbones. Online Experiments. We conducted an online A/B test on the industrial platform (10% live traffic) from May 10 to 20, 2026, covering all six countries. CMRec achieved a consistent increase of 1.77% in advertising revenue and 2.64% in orders on the six-country aggregate, demonstrating significant gains in real-world deployment. 5. Conclusion We presented CMRec, a cross-country generative recommendation (GR) framework that injects cross-country supervision at the data level through dual-constrained code-mixing. CMRec learns a shared semantic codebook from multi-modal content and behavioral co-occurrence across countries. It then performs token-level code-mixing under content-based and market-aware constraints to synthesize reliable mixed-country sequences without requiring item or user overlap. A context-aware loss reweights mixed samples by their contextual plausibility. This suppresses noisy substitutions while preserving useful cross-country signals. Extensive experiments on offline datasets and an online multi-country e-commerce platform show that CMRec achieves superior performance, effectively transfers knowledge from large to small countries, and offers practical insights for cross-country generative recommendation. 6. GenAI Usage Disclosure Generative AI was used in the preparation of this paper to assist with language polishing, grammar checking, and improving the clarity of the presentation. All technical ideas, including the design of the method, the implementation of experiments, and the interpretation of results, are solely contributed by the authors. References Bai et al. (2025) S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, J. Lin, et al. Qwen2.5-VL technical report. arXiv preprint arXiv:2502.13923. Cited by: §3.1, §4.1. Blevins and Zettlemoyer (2022) T. Blevins and L. Zettlemoyer Language contamination helps explains the cross-lingual capabilities of English pretrained models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), Abu Dhabi, United Arab Emirates, p. 3563–3574. External Links: Link, Document Cited by: §1. Briakou et al. (2023) E. Briakou, C. Cherry, and G. Foster Searching for needles in a haystack: on the role of incidental bilingualism in PaLM’s translation capability. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), Volume 1: Long Papers, Toronto, Canada, p. 9432–9452. External Links: Link, Document Cited by: §1. Chan et al. (2020) T. M. Chan, S. Golan, T. Kociumaka, T. Kopelowitz, and E. Porat Approximating text-to-pattern hamming distances. In Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing, p. 643–656. Cited by: §3.2. Cui et al. (2025) Z. Cui, Y. Wu, L. Bai, Y. Wang, X. Zhao, X. Lin, W. Wang, F. Feng, and T. Chua Semantic retrieval augmented contrastive learning for sequential recommendation. In Advances in Neural Information Processing Systems, Cited by: §1, §4.1. Dang et al. (2023) Y. Dang, E. Yang, G. Guo, L. Jiang, X. Wang, X. Xu, Q. Sun, and H. Liu Uniform sequence better: time interval aware data augmentation for sequential recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, Cited by: §4.1. Deng et al. (2025) H. Deng, H. Xing, K. Matsuyama, M. Zhang, J. Hu, H. Wen, Y. Zhang, X. Zeng, and J. Zhang CSMF: cascaded selective mask fine-tuning for multi-objective embedding-based retrieval. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 2122–2131. Cited by: §3.1, §4.1. Hou et al. (2025) Y. Hou, A. Zhang, L. Sheng, Z. Yang, X. Wang, T. Chua, and J. McAuley Generative recommendation models: progress and directions. In Companion Proceedings of the ACM on Web Conference 2025, p. 13–16. Cited by: §1. Hu et al. (2026a) P. Hu, W. Lu, and J. Wang From IDs to semantics: a generative framework for cross-domain recommendation with adaptive semantic tokenization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), Philadelphia, PA, USA. Cited by: §1. Hu et al. (2026b) P. Hu, W. Lu, and J. Wang From ids to semantics: a generative framework for cross-domain recommendation with adaptive semantic tokenization. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 14874–14882. Cited by: §1, §4.1. Huang et al. (2020) J. Huang, A. Sharma, S. Sun, L. Xia, D. Zhang, P. Pronin, J. Padmanabhan, G. Ottaviano, and L. Yang Embedding-based retrieval in facebook search. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, p. 2553–2561. Cited by: §3.1. Huang et al. (2026) Y. Huang, H. Deng, H. Xing, J. Hu, C. Xu, Z. Chen, Y. Zhang, and X. Zeng RCLRec: reverse curriculum learning for modeling sparse conversions in generative recommendation. arXiv preprint arXiv:2603.28124. Cited by: §1. Jin et al. (2025) J. Jin, Y. Zhang, F. Feng, and X. He Generative multi-target cross-domain recommendation. arXiv preprint arXiv:2507.12871. Cited by: §1. Jin et al. (2023) W. Jin, H. Mao, Z. Li, H. Jiang, C. Luo, H. Wen, H. Han, H. Lu, Z. Wang, R. Li, et al. Multilingual e-commerce session recommendation: the amazon m2 dataset and a baseline approach. In Proceedings of the ACM Web Conference 2023, Cited by: §4.1. 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: §4.1. Li et al. (2025a) H. Li, J. Li, W. Ma, P. Sun, H. Wu, J. Wang, Y. Yang, M. Zhang, and S. Ma CD-cdr: conditional diffusion-based item generation for cross-domain recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 1789–1798. Cited by: §1. Li et al. (2024) J. Li, S. Huang, A. Ching, X. Dai, and J. Chen PreAlign: boosting cross-lingual transfer by early establishment of multilingual alignment. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 10246–10257. Cited by: §1. Li and Lyu (2025) M. Li and W. Lyu Adaptive regularization for large-scale sparse feature embedding models. arXiv preprint arXiv:2511.06374. Cited by: §1. Li et al. (2025b) X. Li, B. Chen, J. She, S. Cao, Y. Wang, Q. Jia, H. He, Z. Zhou, Z. Liu, J. Liu, et al. A survey of generative recommendation from a tri-decoupled perspective: tokenization, architecture, and optimization. Cited by: §1. 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. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 1582–1592. Cited by: §1. Liu et al. (2021) Z. Liu, Y. Chen, J. Li, P. S. Yu, J. McAuley, and C. Xiong Contrastive self-supervised sequential recommendation with robust augmentation. arXiv preprint arXiv:2108.06479. Cited by: §4.1. 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 (SIGIR), p. 43–52. Cited by: §3.1. 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: §2, §3.1, §4.1, §4.1. Starke et al. (2025) C. Starke, L. Metikoš, N. Helberger, and C. de Vreese Contesting personalized recommender systems: a cross-country analysis of user preferences. Information, Communication & Society 28 (1), p. 41–60. Cited by: §1, §1. Wang et al. (2023) J. Wang, Z. Zeng, Y. Wang, Y. Wang, X. Lu, T. Li, J. Yuan, R. Zhang, H. Zheng, and S. Xia MISSRec: pre-training and transferring multi-modal interest-aware sequence representation for recommendation. In Proceedings of the 31st ACM International Conference on Multimedia, p. 6548–6557. Cited by: §1. Xing et al. (2025) H. Xing, H. Deng, Y. Mao, J. Hu, Y. Xu, H. Zhang, J. Wang, S. Wang, Y. Zhang, X. Zeng, et al. Reg4rec: reasoning-enhanced generative model for large-scale recommendation systems. arXiv preprint arXiv:2508.15308. Cited by: §1, §2, §4.1. Xu et al. (2022) Z. Xu, M. Zhao, L. Liu, L. Xiao, X. Zhang, and B. Zhang Mixture of virtual-kernel experts for multi-objective user profile modeling. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 4257–4267. Cited by: §3.1. Yang et al. (2024) W. Yang, Y. Jian, Y. Wang, S. Lu, L. Shen, B. Wang, H. Tang, and L. Zhang Not all embeddings are created equal: towards robust cross-domain recommendation via contrastive learning. In Proceedings of the ACM Web Conference 2024, p. 3195–3206. Cited by: §1. Yuan et al. (2023) Z. Yuan, F. Yuan, Y. Song, Y. Li, J. Fu, F. Yang, Y. Pan, and Y. Ni Where to go next for recommender systems? ID- vs. modality-based recommender models revisited. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), Cited by: §3.1. Zang et al. (2022) T. Zang, Y. Zhu, H. Liu, R. Zhang, and J. Yu A survey on cross-domain recommendation: taxonomies, methods, and future directions. ACM Transactions on Information Systems (TOIS) 41 (2), p. 1–39. 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: §4.1. Zhang et al. (2025) H. Zhang, M. Cheng, Q. Liu, J. Jiang, X. Wang, R. Zhang, C. Lei, and E. Chen A comprehensive survey on cross-domain recommendation: taxonomy, progress, and prospects. arXiv preprint arXiv:2503.14110. Cited by: §1. Zhang et al. (2023) W. Zhang, P. Zhang, B. Zhang, X. Wang, and D. Wang A collaborative transfer learning framework for cross-domain recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 5576–5585. Cited by: §1.