arxiv: 2604.02684 · v1 · submitted 2026-04-03 · 💻 cs.IR

Recognition: 1 theorem link

· Lean Theorem

MBGR: Multi-Business Prediction for Generative Recommendation at Meituan

Changhao Li, Haitao Wang, Junwei Yin, Senjie Kou, Shuli Wang, Wenshuai Chen, Xingxing Wang, Yinhua Zhu, Zhilin Zeng

Authors on Pith no claims yet

Pith reviewed 2026-05-13 19:12 UTC · model grok-4.3

classification 💻 cs.IR

keywords generative recommendationmulti-business recommendationsemantic IDnext token predictionindustrial recommendation systemfood delivery platform

0 comments

The pith

MBGR resolves the seesaw phenomenon and representation confusion in multi-business generative recommendation through business-aware semantic IDs, specific prediction structures, and dynamic label routing.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that standard generative recommendation methods, which rely on next-token prediction and a single semantic ID space, fail in settings with multiple distinct businesses because they cannot capture cross-business behavior patterns and mix up semantic signals. The authors introduce MBGR as a tailored framework with three components: a business-aware semantic ID module that tokenizes separately per domain, a multi-business prediction module that supports business-specific outputs, and a label dynamic routing module that converts sparse labels into denser ones. Experiments on Meituan's food delivery platform show gains in both offline metrics and online A/B tests, with the system moved into production. A reader would care because most large-scale recommendation platforms run several businesses at once and need methods that scale without forcing trade-offs between them.

Core claim

MBGR is the first generative recommendation framework built for multi-business scenarios. It uses a Business-aware semantic ID (BID) module for domain-aware tokenization to preserve semantic integrity, a Multi-Business Prediction (MBP) structure to supply business-specific prediction heads, and a Label Dynamic Routing (LDR) module that turns sparse multi-business labels into dense ones. Together these address the seesaw effect, where next-token prediction cannot jointly optimize across businesses, and representation confusion, where a unified semantic ID space blurs distinct business semantics.

What carries the argument

The three-component architecture of BID for domain-aware tokenization, MBP for business-specific prediction heads, and LDR for densifying sparse labels inside a next-token-prediction generative model.

If this is right

Simultaneous optimization across businesses becomes possible without one service degrading another.
Semantic integrity is maintained per business through separate tokenization rather than a single shared space.
Sparse labels from multiple businesses can be made dense enough for effective next-token generation.
The framework supports production deployment on large industrial platforms with measurable gains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same separation of tokenization and prediction heads could extend to other generative tasks that mix domains, such as multi-category search or advertising.
Explicit business routing may prove necessary for any next-token-prediction model once the number of distinct services grows beyond a small threshold.
Automating the choice of routing weights in LDR could allow the model to adapt when new businesses are added without manual redesign.

Load-bearing premise

That the seesaw phenomenon and representation confusion are the dominant obstacles in multi-business generative recommendation and that the BID, MBP, and LDR modules resolve them without creating new problems or needing heavy per-platform retuning.

What would settle it

Running controlled A/B tests on a second multi-business recommendation platform and finding no reduction in the seesaw effect or no improvement in cross-business metrics when MBGR replaces a standard generative baseline.

Figures

Figures reproduced from arXiv: 2604.02684 by Changhao Li, Haitao Wang, Junwei Yin, Senjie Kou, Shuli Wang, Wenshuai Chen, Xingxing Wang, Yinhua Zhu, Zhilin Zeng.

**Figure 1.** Figure 1: The overall architecture of MBGR. 1. Contextual Fusion: Concatenate the item embedding with business context: z 𝑏 = [e, b] ∈ R 𝑑𝑒+𝑑𝑏 (9) where b ∈ R 𝑑𝑏 is the learnable business type embedding that encodes domain-specific characteristics. 2. Adaptive Gating: Compute business-specific attention weights: g 𝑏 = SiLu(FFN𝑔𝑎𝑡𝑒 (z 𝑏 )) ∈ R 𝐾 (10) where FFN𝑔𝑎𝑡𝑒 is a shared gating network that learns to allocate e… view at source ↗

**Figure 2.** Figure 2: Comparison of embedding distributions between [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Architecture of the online deployment with MBGR. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

read the original abstract

Generative recommendation (GR) has recently emerged as a promising paradigm for industrial recommendations. GR leverages Semantic IDs (SIDs) to reduce the encoding-decoding space and employs the Next Token Prediction (NTP) framework to explore scaling laws. However, existing GR methods suffer from two critical issues: (1) a \textbf{seesaw phenomenon} in multi-business scenarios arises due to NTP's inability to capture complex cross-business behavioral patterns; and (2) a unified SID space causes \textbf{representation confusion} by failing to distinguish distinct semantic information across businesses. To address these issues, we propose Multi-Business Generative Recommendation (MBGR), the first GR framework tailored for multi-business scenarios. Our framework comprises three key components. First, we design a Business-aware semantic ID (BID) module that preserves semantic integrity via domain-aware tokenization. Then, we introduce a Multi-Business Prediction (MBP) structure to provide business-specific prediction capabilities. Furthermore, we develop a Label Dynamic Routing (LDR) module that transforms sparse multi-business labels into dense labels to further enhance the multi-business generation capability. Extensive offline and online experiments on Meituan's food delivery platform validate MBGR's effectiveness, and we have successfully deployed it in production.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MBGR adds three targeted modules to generative recommendation for multi-business settings and claims production deployment at Meituan, but the abstract supplies no metrics or baselines to judge whether the fixes deliver real gains.

read the letter

The main thing here is a practical extension of generative recommendation to platforms that run several distinct services at once. The authors identify the seesaw effect, where training on one business degrades another, and representation confusion from a shared semantic ID space. They respond with business-aware semantic IDs that tokenize per domain, a multi-business prediction structure that keeps separate paths, and a label dynamic routing module that densifies sparse cross-business signals. These pieces are presented as new relative to prior GR work, and the paper notes successful offline tests, online A/B results, and live deployment on Meituan's food delivery system. That deployment is the strongest concrete signal the work offers. The architecture description is clear enough that a reader could sketch an implementation from it. The focus on real multi-business data is also useful, since most GR papers stay within single-domain settings. The soft spots sit in the evidence. No performance deltas, no baseline comparisons, and no module ablations appear in the abstract, so it is impossible to tell how large the gains are or whether the added components create new latency or tuning costs. The central assumption—that these three changes cleanly resolve the named problems without side effects—rests on Meituan-specific data and would need verification on other distributions. The citation list follows the usual GR references but could usefully expand to more recent multi-task or domain-adaptation work. This paper is aimed at industrial recsys teams that already run generative models and now face heterogeneous services. A practitioner who wants concrete module ideas for cross-business scaling would get value from the design. It deserves peer review because the production claim is worth checking in detail, even if the experimental section will probably need expansion and clearer numbers before acceptance.

Referee Report

2 major / 2 minor

Summary. The paper proposes MBGR, the first generative recommendation (GR) framework tailored for multi-business scenarios. It diagnoses two issues in existing GR methods—seesaw phenomenon arising from Next Token Prediction's inability to capture complex cross-business behavioral patterns, and representation confusion caused by a unified Semantic ID (SID) space—and introduces three components to address them: Business-aware semantic ID (BID) for domain-aware tokenization, Multi-Business Prediction (MBP) structure for business-specific predictions, and Label Dynamic Routing (LDR) module to convert sparse multi-business labels into dense labels. The framework is evaluated via offline and online experiments on Meituan's food delivery platform and has been deployed in production.

Significance. If the experimental claims hold, the work would be significant for industrial-scale recommendation systems, as it extends the GR paradigm (with its Semantic IDs and NTP scaling) to multi-business settings where cross-domain patterns and semantic distinctions matter. The reported production deployment provides evidence of practical impact beyond academic benchmarks.

major comments (2)

Abstract: the central claim that 'extensive offline and online experiments validate MBGR's effectiveness' is load-bearing, yet the abstract (and by extension the high-level description) supplies no metrics, baselines, ablation results, or error analysis, making it impossible to assess whether the data actually supports superiority over prior GR methods.
The weakest assumption—that the seesaw phenomenon and representation confusion are the primary problems and that BID/MBP/LDR fully resolve them without new issues or Meituan-specific tuning—is not accompanied by concrete evidence (e.g., no ablation isolating each module's contribution to cross-business pattern capture or semantic separation).

minor comments (2)

The paper would benefit from explicit definitions or citations for the 'seesaw phenomenon' and 'representation confusion' in the introduction or related-work section to ground the motivation.
Implementation details for BID tokenization, MBP architecture, and LDR routing (e.g., exact tokenization rules, prediction heads, or routing equations) should be expanded for reproducibility, ideally with pseudocode or diagrams.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive feedback on our manuscript. We address each major comment below and outline the revisions we will make to strengthen the presentation of our results.

read point-by-point responses

Referee: Abstract: the central claim that 'extensive offline and online experiments validate MBGR's effectiveness' is load-bearing, yet the abstract (and by extension the high-level description) supplies no metrics, baselines, ablation results, or error analysis, making it impossible to assess whether the data actually supports superiority over prior GR methods.

Authors: We agree that the abstract would benefit from concrete quantitative support. In the revised version we will add specific metrics (e.g., relative CTR and GMV lifts over the strongest generative and multi-task baselines) together with a brief mention of the ablation results. The full experimental details, baselines, and error analysis remain in Sections 4 and 5. revision: yes
Referee: The weakest assumption—that the seesaw phenomenon and representation confusion are the primary problems and that BID/MBP/LDR fully resolve them without new issues or Meituan-specific tuning—is not accompanied by concrete evidence (e.g., no ablation isolating each module's contribution to cross-business pattern capture or semantic separation).

Authors: Section 3 explains the motivation for each component and how BID, MBP, and LDR target the two diagnosed issues. Section 4.3 already reports module-level ablations that quantify gains on cross-business metrics. To make the evidence more explicit we will add, in the revision, quantitative measures of semantic separation (e.g., inter-business embedding cosine distances) and confirm that the improvements hold across multiple business subsets rather than being Meituan-specific. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper proposes MBGR as a novel framework with three explicitly introduced components (BID for domain-aware tokenization, MBP for business-specific prediction, and LDR for label routing) to mitigate the stated seesaw and representation confusion problems. These modules are defined independently of the target outcomes and are not shown to reduce to self-definitions, fitted parameters renamed as predictions, or load-bearing self-citations. The central claims rest on offline/online experiments and production deployment on Meituan data, which constitute external validation rather than internal equivalence. No uniqueness theorems, ansatzes smuggled via citation, or renamings of known results appear in the provided text. The derivation chain is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 3 invented entities

Only the abstract is available, so no concrete free parameters, axioms, or invented entities can be extracted in detail. The framework introduces three new modules treated as invented entities without independent evidence beyond the claimed experiments.

invented entities (3)

Business-aware semantic ID (BID) no independent evidence
purpose: Preserve semantic integrity via domain-aware tokenization across businesses
New module introduced to solve representation confusion
Multi-Business Prediction (MBP) structure no independent evidence
purpose: Provide business-specific prediction capabilities to avoid seesaw effects
New structure introduced to handle cross-business patterns
Label Dynamic Routing (LDR) module no independent evidence
purpose: Transform sparse multi-business labels into dense labels
New module introduced to enhance multi-business generation

pith-pipeline@v0.9.0 · 5549 in / 1406 out tokens · 59508 ms · 2026-05-13T19:12:23.877256+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
We propose Multi-Business Generative Recommendation (MBGR)... BID module... MBP structure... LDR module... InfoNCE loss... reconstruction loss

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · 2 internal anchors

[1]

Anirudhan Badrinath, Prabhat Agarwal, Laksh Bhasin, Jaewon Yang, Jiajing Xu, and Charles Rosenberg. 2025. PinRec: Outcome-Conditioned, Multi-Token Generative Retrieval for Industry-Scale Recommendation Systems.arXiv preprint arXiv:2504.10507(2025)

work page arXiv 2025
[2]

Jianxin Chang, Chenbin Zhang, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, and Kun Gai. 2023. Pepnet: Parameter and embedding personalized network for infusing with personalized prior information. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3795–3804

work page 2023
[3]

Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al

work page
[4]

InProceedings of the 1st workshop on deep learning for recommender systems

Wide & deep learning for recommender systems. InProceedings of the 1st workshop on deep learning for recommender systems. 7–10

work page
[5]

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)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[6]

Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction.arXiv preprint arXiv:1703.04247(2017)

work page arXiv 2017
[7]

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.arXiv preprint arXiv:2505.18654(2025)

work page arXiv 2025
[8]

Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating long semantic IDs in parallel for recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 956–966

work page 2025
[9]

Yanhua Huang, Yuqi Chen, Xiong Cao, Rui Yang, Mingliang Qi, Yinghao Zhu, Qingchang Han, Yaowei Liu, Zhaoyu Liu, Xuefeng Yao, et al . 2025. Towards Large-scale Generative Ranking.arXiv preprint arXiv:2505.04180(2025)

work page arXiv 2025
[10]

Pengcheng Li, Runze Li, Qing Da, An-Xiang Zeng, and Lijun Zhang. 2020. Improv- ing multi-scenario learning to rank in e-commerce by exploiting task relation- ships in the label space. InProceedings of the 29th ACM International Conference on Information & Knowledge Management. 2605–2612

work page 2020
[11]

Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xdeepfm: Combining explicit and implicit feature in- teractions for recommender systems. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1754–1763

work page 2018
[12]

Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. InProceedings of the 29th ACM International Conference on Information & Knowledge Management. 2685–2692

work page 2020
[13]

Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al

work page
[14]

Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

work page 2023
[15]

Xiang-Rong Sheng, Liqin Zhao, Guorui Zhou, Xinyao Ding, Binding Dai, Qiang Luo, Siran Yang, Jingshan Lv, Chi Zhang, Hongbo Deng, et al. 2021. One model to serve all: Star topology adaptive recommender for multi-domain ctr prediction. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 4104–4113

work page 2021
[16]

Zihua Si, Zhongxiang Sun, Xiao Zhang, Jun Xu, Xiaoxue Zang, Yang Song, Kun Gai, and Ji-Rong Wen. 2023. When search meets recommendation: Learning disentangled search representation for recommendation. InProceedings of the 46th international ACM SIGIR conference on research and development in information retrieval. 1313–1323

work page 2023
[17]

Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. InProceedings of the 14th ACM Conference on Recommender Systems. 269–278

work page 2020
[18]

Yichao Wang, Huifeng Guo, Bo Chen, Weiwen Liu, Zhirong Liu, Qi Zhang, Zhicheng He, Hongkun Zheng, Weiwei Yao, Muyu Zhang, et al. 2022. Causalint: Causal inspired intervention for multi-scenario recommendation. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4090–4099

work page 2022
[19]

Yuhao Wang, Yichao Wang, Zichuan Fu, Xiangyang Li, Wanyu Wang, Yuyang Ye, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2024. Llm4msr: An llm- enhanced paradigm for multi-scenario recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2472–2481

work page 2024
[20]

Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, Shuanglong Li, et al. 2025. Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations.arXiv preprint arXiv:2503.02453(2025)

work page arXiv 2025
[21]

Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. 2021. Soundstream: An end-to-end neural audio codec.IEEE/ACM Transactions on Audio, Speech, and Language Processing30 (2021), 495–507

work page 2021
[22]

Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152(2024)

work page internal anchor Pith review arXiv 2024
[23]

Kexin Zhang, Yichao Wang, Xiu Li, Ruiming Tang, and Rui Zhang. 2024. Incmsr: An incremental learning approach for multi-scenario recommendation. InPro- ceedings of the 17th ACM International Conference on Web Search and Data Mining. 939–948

work page 2024
[24]

Zijian Zhang, Shuchang Liu, Jiaao Yu, Qingpeng Cai, Xiangyu Zhao, Chunxu Zhang, Ziru Liu, Qidong Liu, Hongwei Zhao, Lantao Hu, et al . 2024. M3oe: Multi-domain multi-task mixture-of experts recommendation framework. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 893–902

work page 2024
[25]

Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448

work page 2024
[26]

Zuowu Zheng, Ze Wang, Fan Yang, Jiangke Fan, Teng Zhang, Yongkang Wang, and Xingxing Wang. 2025. EGA-V2: An End-to-end Generative Framework for Industrial Advertising.arXiv preprint arXiv:2505.17549(2025)

work page arXiv 2025
[27]

Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 5941–5948

work page 2019
[28]

Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1059–1068

work page 2018
[29]

Jie Zhou, Xianshuai Cao, Wenhao Li, Lin Bo, Kun Zhang, Chuan Luo, and Qian Yu. 2023. HiNet: Novel Multi-Scenario & Multi-Task Learning with Hierarchi- cal Information Extraction. In2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2969–2975

work page 2023
[30]

Jiachen Zhu, Yichao Wang, Jianghao Lin, Jiarui Qin, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. M-scan: A Multi-Scenario Causal-driven Adaptive Network for Recommendation. InProceedings of the ACM Web Conference 2024. 3844–3853

work page 2024