MBGR: Multi-Business Generative Recommendation¶

Core business-aware SID, shared experts, per-business prediction, and label dynamic routing.

Paper information¶

Field	Value
Paper	arXiv 2604.02684
Affiliations	Meituan
First public release	2026-04-03 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`mbgr`
Local code	`src/auto_research/reproductions/mbgr/`

Original paper¶

Background and contribution¶

Ordinary next-item labels cover only the next logged business. A business-conditioned autoencoder retains SID semantics, shared MoE provides business representations, and LDR finds each business's nearest future interaction, masking missing labels.

flowchart LR
  I[Cross-business history] --> B[BID encoder and reconstruction] --> M[Shared MoE] --> D[Per-business SID decoding]
  I --> L[LDR future labels] --> D

Key figure from the paper¶

Original Figure 1: MBGR. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ e^b=\sum_{k=1}^K g_k^b\operatorname{FFN}^{exp}_k([e,b]),\qquad i_{u,t+1}^{(k)}=i_{u,t'},\ t'=\min\{t''>t:b_{u,t''}=b_k\}. \]

Paper results¶

Meituan live A/B reports CTCVR +3.98%.

Local reproduction¶

Versus shared next-item baseline, use 19 genre proxies and BID/MBP/LDR. Hit@10 −4.17%, NDCG@10 −5.92%. Validation chose blend 0.1. LDR masked 140,235 missing business targets. See metrics.

auto-research reproduce --paper mbgr --seed 42

Reproduction limits¶

Genres do not replace private multibusiness logs or production generator scale.