Multi-Business Generative Recommendation

• MBGR is a framework where one generative system is conditioned on explicit business objectives to optimize recommendations across various platforms.
• It reformulates retrieval by embedding business signals as control tokens or budgets, dynamically balancing personalization with commercial goals.
• Industrial implementations (e.g., Pinterest, Meituan) demonstrate significant gains in recall, engagement, and revenue while addressing engineering challenges.

Multi-Business Generative Recommendation (MBGR) denotes the use of one generative retrieval or recommendation system to flexibly serve multiple, possibly competing business objectives across heterogeneous surfaces, markets, or scenarios. In this framing, business goals are not merely post hoc weights on a fixed candidate set; they become conditioning signals, control tokens, business labels, budgets, or value functions inside the generative process itself. Recent instantiations include outcome-conditioned continuous-vector generation at Pinterest, behavior-first autoregressive generation in multi-behavior sequential recommendation, bid-aware control-token decoding for joint organic and sponsored delivery, business-aware semantic IDs and multi-business prediction at Meituan, and one-model generative advertising and e-commerce systems such as GPR and OneMall (Badrinath et al., 9 Apr 2025, Liu et al., 2024, Jiang et al., 23 Mar 2026, Zhang et al., 13 Nov 2025, Li et al., 3 Apr 2026, Zhang et al., 29 Jan 2026).

1. Definition, scope, and terminology

In the industrial literature, MBGR is a response to the fact that modern recommendation systems rarely optimize a single scalar objective. Pinterest describes MBGR as “the idea of using one generative retrieval system to flexibly serve multiple, possibly competing business objectives (e.g., clicks, saves/repins, outbound product clicks, diversity/exploration) across heterogeneous surfaces and markets,” and operationalizes it through outcome conditioning and multi-token generation (Badrinath et al., 9 Apr 2025). Meituan uses the same term for “the first GR framework tailored for multi-business scenarios,” with explicit attention to cross-business interference and representation separation (Li et al., 3 Apr 2026). Kuaishou’s OneMall formulates an analogous problem as one model serving product-card, short-video, and live-streaming scenarios, while GPR frames a one-model advertising system that can be extended to a multi-business generative recommendation paradigm (Zhang et al., 29 Jan 2026, Zhang et al., 13 Nov 2025).

A recurring ambiguity in the literature is that the acronym MBGR is also used for Multi-Behavior Generative Recommendation in MBGen, where the task is multi-behavior sequential recommendation rather than multi-business platform optimization (Liu et al., 2024). The two usages are closely related but not identical. Multi-behavior models focus on behavior types such as click, add-to-cart, add-to-favorite, and purchase; multi-business models focus on heterogeneous surfaces, markets, monetization regimes, and business policies. This suggests that MBGR is best understood as a family of conditional generative formulations rather than a single canonical architecture.

2. Conditional-generation formulations

A central shift in MBGR is the reformulation of retrieval from static similarity scoring to conditional generation. In PinRec, if $x$ denotes the user/context sequence and $c$ denotes a conditioning signal that captures the business goal, the generator learns

$$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$

At serving time, the business policy is expressed through a per-action budget distribution $B: A \to [0,1]$ with $\sum_{a \in A} B(a)=1$, and the total number of recommended items $N$ is split per action as $B(a)\cdot N$ (Badrinath et al., 9 Apr 2025). The important consequence is that personalization and business policy are decoupled: the same user history can be steered toward repins, grid clicks, outbound clicks, or exploration by changing $c$ and $B$.

MBGen formulates multi-behavior sequential recommendation as a two-step generative process:

$$p(b_{t+1}, i_{t+1} \mid S_t) = p(b_{t+1} \mid S_t)\cdot p(i_{t+1} \mid S_t, b_{t+1}),$$

with a unified token-level autoregressive objective

$c$0

Because behavior tokens are part of the interleaved sequence, the model jointly optimizes next-behavior and next-item generation without separate losses or weights (Liu et al., 2024). Although this formulation is not “multi-business” in the Meituan or Pinterest sense, it establishes a general principle: latent intent or policy variables can be generated first and then used to condition item generation.

Meituan generalizes Next Token Prediction to multi-business conditional generation. Given mixed-business history $c$1 with $c$2, the model generates the SID for the next possible item for each business $c$3 in parallel:

$c$4

Its Label Dynamic Routing (LDR) further densifies supervision by routing each business head to the nearest future interaction of the same business,

$c$5

and masking the loss if no such $c$6 exists (Li et al., 3 Apr 2026). Here, the conditioning variable is the business label itself, not only a behavior or objective token.

GEM-Rec extends the same logic to organic and sponsored delivery. Given history $c$7, it factorizes the next segment as

$c$8

where $c$9 is a control token that decides whether the next slot is organic or sponsored (Jiang et al., 23 Mar 2026). The common pattern across these formulations is explicit factorization of business state, behavior type, or monetization mode from the content-generation process.

3. Representations, tokenization, and network architectures

MBGR systems differ sharply in how they represent items. PinRec does not emit semantic ID tokens; it generates real-valued item representations suitable for nearest-neighbor retrieval. Each item is embedded via an item-type-specific MLP, the transformer is a GPT-2 style decoder over heterogeneous user activity sequences, and an output head maps the hidden state plus auxiliary conditioning to an item representation in the same space as item embeddings (Badrinath et al., 9 Apr 2025). This continuous-output design removes a fixed item-token vocabulary from the generation loop, but it makes embedding quality and retrieval infrastructure central.

By contrast, MBGen, GEM-Rec, GPR, OneMall, and Meituan all rely on discrete tokenization, usually through semantic IDs. MBGen interleaves behavior and item-code tokens and introduces a position-routed sparse architecture with deterministic routing by token position/type and a behavior-injected feed-forward mechanism (Liu et al., 2024). GEM-Rec uses hierarchical semantic IDs learned with a Residual Quantized VAE and prepends control tokens $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$0 to each item-code sequence (Jiang et al., 23 Mar 2026). Meituan retains a shared SID vocabulary but learns business-aware encodings and decodings through a Business-aware semantic ID (BID) module, then applies Multi-Business Prediction (MBP) and Label Dynamic Routing (LDR) on top of a Transformer-based autoregressive backbone (Li et al., 3 Apr 2026). GPR serializes U-Token, O-Token, E-Token, and I-Token in a unified schema and maps both ads and organic content into a shared multi-level semantic ID space through RQ-KMeans+, then uses the Heterogeneous Sequence-wise Decoder, Progressive Token-wise Decoder, and Hierarchical Token-wise Evaluator (Zhang et al., 13 Nov 2025). OneMall uses a multimodal semantic tokenizer based on residual K-means plus FSQ, a decoder-only Transformer with Query-Former sequence compression, Cross-Attention for multi-behavior fusion, and Sparse MoE for scalable autoregressive generation (Zhang et al., 29 Jan 2026).

System	Core conditioning	Output form
PinRec	action embeddings and budget $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$1	continuous item representations
MBGen	behavior token or target behavior	interleaved behavior and item tokens
GEM-Rec	$$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$2 plus live bids	semantic ID codes
Meituan MBGR	business embedding, MBP, LDR	business-aware SID tokens
GPR	U/O/E/I schema, value guidance, HEPO	multi-level semantic IDs
OneMall	scenario-aware semantic tokenizer and RL	three-level SIDs

These architectural choices are linked to different operating constraints. Continuous-vector generation emphasizes nearest-neighbor scalability, KV-cached decoding, and retrieval-system compatibility, whereas token-based systems emphasize legality constraints, prefix-aware decoding, trie pruning, and vocabulary balance. This suggests that “generative recommendation” in MBGR is not architecturally monolithic; it spans both continuous-vector autoregression and discrete semantic-ID generation.

4. Control mechanisms, optimization, and inference-time steering

The distinctive feature of MBGR is not only generative modeling, but controllability at training time or inference time. PinRec exposes two primary knobs. The first is the outcome budget $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$3: increasing $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$4 raises repin-oriented recall and online repin rates while reducing click-oriented recall and rates, and vice versa. The second is the multi-token parameter $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$5 per step: increasing $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$6 improves performance, diversity, and reduces latency because multiple embeddings are generated while reusing the KV cache. PinRec also compresses near-duplicate continuous outputs by cosine threshold before Faiss IVF-HNSW retrieval, with compression ratios of approximately $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$7 for unconditioned PinRec and approximately $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$8 for outcome-conditioned multi-token variants (Badrinath et al., 9 Apr 2025).

MBGen’s control mechanism is behavior-first generation. At inference, it can first predict the next behavior token, then generate item tokens conditioned on that behavior, or it can accept a provided target behavior and start decoding from a prompt such as $$p(y_{1:T} \mid x, c) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x, c).$$9. It further introduces behavior-aware sampling, which reweights beams per behavior proportion and improves behavior-item retrieval without increasing beam size (Liu et al., 2024). In this setting, control is achieved through prompts and beam allocation rather than explicit business budgets.

GEM-Rec makes inference-time steering explicit through Bid-Aware Decoding. It modulates the logit of $B: A \to [0,1]$0 by

$B: A \to [0,1]$1

and modulates item-code logits inside the sponsored branch by

$B: A \to [0,1]$2

The paper proves allocation monotonicity—exposure is weakly increasing in an ad’s bid under fixed assumptions—and states Organic Integrity: item-level modulation applies only when $B: A \to [0,1]$3, so the ranking among organic items is invariant to $B: A \to [0,1]$4 (Jiang et al., 23 Mar 2026). Here, MBGR becomes a generative policy with an explicit economic control parameter.

GPR and OneMall push control further into training. GPR uses Multi-Token Prediction pretraining, Value-Aware Fine-Tuning, and Hierarchy Enhanced Policy Optimization, with a final value

$B: A \to [0,1]$5

and value-guided trie-based beam search at inference (Zhang et al., 13 Nov 2025). OneMall connects retrieval to ranking through RL, using reward fusion such as $B: A \to [0,1]$6, with $B: A \to [0,1]$7, $B: A \to [0,1]$8, and $B: A \to [0,1]$9 in experiments, and combines language-model loss, contrastive loss, and RL loss in a joint objective (Zhang et al., 29 Jan 2026). Meituan uses business weights $\sum_{a \in A} B(a)=1$0, time decay $\sum_{a \in A} B(a)=1$1, and dense routed labels to strengthen smaller or fast-changing businesses (Li et al., 3 Apr 2026). Across these systems, business steering may occur through prompt tokens, explicit budgets, live bids, value-aware loss reweighting, or RL-based policy alignment.

5. Empirical evidence and industrial deployment

The empirical record is strongly industrial. PinRec is built for Pinterest’s scale of 550M+ MAU and 1.5B weekly saves, with an NVIDIA Triton ensemble that includes a Signal Service in a Lambda architecture, a TorchScript Learned ID Embeddings service, GPT-2 style autoregressive inference with KV Cache, CUDA Graphs and torch.compile, cosine-threshold compression, and Faiss IVF-HNSW retrieval (Badrinath et al., 9 Apr 2025). Offline, unordered recall@10 on Homefeed rises from 0.461 for PinnerFormer to 0.608 for PinRec-UC, 0.625 for PinRec-OC, and 0.676 for PinRec-{MT,OC}; Related Pins rises from 0.412 to 0.631; Search rises from 0.257 to 0.450. Outcome conditioning increases unordered recall for matched actions, including +1.9% for repins and +6.2% for outbound clicks, and at $\sum_{a \in A} B(a)=1$2 multi-token generation improves unordered recall@10 by approximately +16.0%, unique proportion by approximately +21.3%, with approximately 10× lower generation latency compared to next-token generation with the same total number of outputs. Online A/B on Homefeed reports, for the best group per variant, fulfilled sessions +0.02% for PinRec-UC, +0.21% for PinRec-OC, and +0.28% for PinRec-{MT,OC}; time spent −0.02%, +0.16%, and +0.55% respectively; and Homefeed grid clicks +1.87%, +4.01%, and +3.33% respectively. Search with PinRec-UC reports Search fulfillment +0.48%, Search repins +0.84%, and Search grid clicks +1.46% (Badrinath et al., 9 Apr 2025).

MBGen evaluates on two public datasets with full-catalog ranking. On Retail, MBGen (SID) reaches NDCG@10 = 0.3954 for target behavior item prediction, versus 0.2449 for the best baseline SASRec_M, with improvements of +61.45% in NDCG@10, +70.08% in NDCG@5, and +44.89% in HR@10. On IJCAI, MBGen (SID) reaches NDCG@10 = 0.3949 versus 0.2788 for the best baseline TIGER, with improvements of +41.64% in NDCG@10, +34.45% in NDCG@5, and +45.82% in HR@10. It also reports strong gains on behavior-specific item prediction and behavior-item joint prediction, and states that all gains are statistically significant with $\sum_{a \in A} B(a)=1$3 by paired $\sum_{a \in A} B(a)=1$4-test (Liu et al., 2024).

GEM-Rec evaluates on Steam and Amazon Beauty, Sports, and Toys. As $\sum_{a \in A} B(a)=1$5 increases, Ad Rate and Revenue increase while Strict Policy Fit declines proportionally, and Conditional Organic NDCG and Recall stay essentially flat. On Steam, GEM-Rec moves from Ad Rate 2.5%; Revenue 535; NDCG 0.1411 at $\sum_{a \in A} B(a)=1$6 to Ad Rate 4.7%; Revenue 1,173; NDCG 0.1381 at $\sum_{a \in A} B(a)=1$7; the corresponding TIGER baseline has Ad Rate 0.0%; NDCG 0.1442. The paper also reports 100% of generated ad IDs are valid across all settings and large revenue response under simulated bid shocks (Jiang et al., 23 Mar 2026).

The large-scale one-model systems likewise report production deployment. GPR is fully deployed in Tencent Weixin Channels, serving hundreds of millions of users and tens of millions of ads. Offline, the full GPR reaches HitRate@100 = 27.32%, compared with 18.98% for HSTU and 19.85% for OneRec. In online A/B tests, sequential launches report GMV +2.11% for v0.1, +0.70% for v0.2, +0.63% for v0.3, +0.71% for v0.4, and +0.58% for v0.5; v0.1 also reports CTR +1.69%, CVR +1.15%, CTCVR +3.16%, and GMV +2.11% (Zhang et al., 13 Nov 2025). Meituan trains on 38,258,649 users and 54,875,570 merchants across four businesses and reports offline Hit@10 gains from 0.0214 for HSTU to 0.0410 for HSTU-MBGR overall, with especially large gains on smaller businesses such as business B from 0.0127 to 0.0554. Downstream CTCVR GAUC rises from 0.7748 to 0.8040 overall, and online A/B over one week on 30% traffic reports CTCVR +3.98% overall (p < 0.01), with business-wise gains of A +3.0%, B +7.5%, C +4.5%, and D +5.2% (Li et al., 3 Apr 2026). OneMall has been deployed to 400M+ DAU at Kuaishou and reports headline improvements of +13.01% GMV in Product-card, +15.32% Orders in Short-Video, and +2.78% Orders in Live-Streaming; a detailed snapshot reports Product-card +14.71% GMV, Short-video +10.33% GMV, and Live-streaming +4.90% GMV (Zhang et al., 29 Jan 2026).

6. Limitations, controversies, and open directions

The main controversy in MBGR is whether multi-business capability follows automatically from scaling a generic generative recommender. Meituan argues that it does not: a unified SID space and standard NTP over mixed-business sequences produce a seesaw phenomenon and representation confusion, and therefore require explicit business-aware tokenization, prediction, and label routing (Li et al., 3 Apr 2026). This directly challenges the assumption that a single generative backbone is sufficient without business-specific structure.

A second limitation concerns reliance on observed logs. PinRec notes that conditioning relies on observed outcomes for training and that counterfactual alignment to future goals is limited; exploration and policy budgets help, but learning to optimize unseen objectives may require additional alignment methods (Badrinath et al., 9 Apr 2025). GEM-Rec similarly learns a feasibility policy from historical “survivor” logs and states that if historical ad placement is suboptimal or biased, the model will inherit those patterns; achieving DSIC payments over practical beam search is left for future work (Jiang et al., 23 Mar 2026). GPR addresses some of this with off-policy correction, simulation, and Anticipatory Request Rehearsal, but still treats broader multi-objective tuning and constraint handling as an extensible direction rather than a solved problem (Zhang et al., 13 Nov 2025).

Fairness, long-tail coverage, cold-start handling, and latency remain open engineering and research issues. PinRec explicitly recommends guardrails such as category caps, long-tail quotas, deduping, blocklists or safelists, and exploration budgets via $\sum_{a \in A} B(a)=1$8; it also notes that pure nearest-neighbor retrieval may still favor head items without diversity constraints (Badrinath et al., 9 Apr 2025). OneMall highlights token collisions, invalid SIDs, and RL instability under excessive RL weight, while GEM-Rec notes that high $\sum_{a \in A} B(a)=1$9 regimes reduce ad relevance and strict policy fit (Zhang et al., 29 Jan 2026, Jiang et al., 23 Mar 2026). Across the surveyed systems, future work includes generative ranking, sequence “prompt” optimization, RLHF-like alignment techniques, multiple control tokens such as $N$0, $N$1, and $N$2, principled multi-objective optimization, and tighter unification of retrieval and ranking (Badrinath et al., 9 Apr 2025, Jiang et al., 23 Mar 2026, Zhang et al., 13 Nov 2025).

Taken together, the literature positions MBGR as a transition from single-objective retrieval toward controllable generative policies. The shared result is not one fixed recipe, but a recurrent design pattern: encode heterogeneous user context, make business intent explicit inside the generation process, and expose serving-time or training-time control variables that can trade off relevance, monetization, diversity, and operational constraints within one model family (Badrinath et al., 9 Apr 2025, Li et al., 3 Apr 2026).