GenRec: Page-Wise Generative Recommendation¶

Page-wise NTP, asymmetric token merging, GRPO-SR, and NLL run.

Paper information¶

Field	Value
Paper	arXiv 2604.14878
Affiliations	JD.com
First public release	2026-04-16 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`genrec`
Local code	`src/auto_research/reproductions/genrec/`

Original paper¶

Background and contribution¶

Item-at-a-time generation ignores complements and ordering on a page. Predict the next page, compress its inputs with an asymmetric merger, optimize group reward, and retain NLL to limit drift.

flowchart LR
  C[User context] --> M[Asymmetric merger] --> N[Page-wise prediction]
  N --> G[Group reward]
  N --> L[NLL constraint]
  G --> P[Page policy]
  L --> P

Key figure from the paper¶

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

Core mechanism¶

\[ \mathcal L=\mathcal L_{\rm GRPO-SR}+\lambda\mathcal L_{\rm NLL},\qquad A_i=R_i-\frac1G\sum_{j=1}^{G}R_j. \]

Paper results¶

JD.com's online results report +9.5% clicks and +8.7% transactions.

Local reproduction¶

Group three consecutive positive interactions into a page. Optimize position-aware advantage with NLL. Pointwise ranking NDCG@10 is 0.03540 vs. 0.07621 (+115.30%) locally; see metrics/movielens-100k-seed42.json.

auto-research reproduce --paper genrec --dataset-dir data --seed 42

Reproduction limits¶

Local page-proxy evidence is separate from production click and transaction measurements.