GrowthGR: Generative retrieval for long-term value of new products¶

ItemLTV, three-level SIDs, next-token prediction, clipped MoPO/GRPO, and valid-SID constraints execute locally.

Paper information¶

Field	Details
Paper	arXiv 2605.17994
Institution	Alibaba Group / Taobao & Tmall.
First public release	2026-05-18 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`growthgr`
Local reproduction code	`src/auto_research/reproductions/growthgr/`

Paper method and evidence¶

Optimizing immediate sales can starve new items of exposure. Counterfactual ItemLTV estimates incremental long-term orders from clicks. An RQ-VAE SID and decoder-only generator retrieve items. MoPO combines purchase, click, exposure, ranking, and uplift labels with clipped inverse propensity for rare valuable items.

flowchart LR
  C[Counterfactual ItemLTV] --> R[Multiobjective reward]
  I[Three-level semantic IDs] --> G[Decoder-only retrieval]
  G --> R
  R --> U[Clipped MoPO update]

Original paper figure¶

The original figure shows long-term value and generative retrieval. From the original paper; copyright remains with the authors.

Core equation¶

\[ \hat y=G_{base}(X)+W\,G_{uplift}(X),\qquad r_i=\operatorname{Clip}(-\log\pi_{old}(o_i|x),1,M)\sum_kw_ks_k. \]

\[ \mathcal J^{CLIP}=\min(\rho A,\operatorname{clip}(\rho,1-\epsilon,1+\epsilon)A). \]

Reported results¶

Taobao/Tmall reports new-product GMV +5.3% and sitewide search GMV +0.3% online.

Local reproduction¶

auto-research reproduce --paper growthgr --seed 42

On the same transition policy, add ItemLTV and eight MoPO rounds to NTP retrieval. Local Hit@10 +0.00%, NDCG@10 +2.05%, head share down 3.06%; validation selects blend 0.1. Metrics. Future positive-frequency proxies seven-day orders; no actual search funnel or multimodal product embeddings.