GRC: generation, reflection, and correction for generative recommendation¶
Fidelity: core mechanism. Runs structured SFT, full-trajectory GRPO, and entropy-guided reflection scheduling.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2602.23639 |
| Organization | Alibaba International / Wuhan University |
| First public release | 2026-02-27 (arXiv v1) |
| Authors' code | The paper did not provide official or author code when checked on 2026-08-09. |
| Adapter / method | grc |
| Local reproduction code | src/auto_research/reproductions/grc/ |
Original paper¶
Motivation and method¶
Early errors in autoregressive semantic-ID generation can propagate. GRC structures a draft SID, reflection on the first error and attribute consistency, and a corrected SID. It applies GRPO to whole trajectories and allocates a fixed correction budget to the most uncertain beams using reflection-token entropy.
flowchart LR
A[User history] --> B[Generate draft SID]
B --> C[Reflect on first error and semantics]
C --> D[Corrected SID]
D --> E[Trajectory GRPO]
C --> F[Fixed-budget EGRS scheduling]
Key figure from the paper¶
Figure 2: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.
Core equations¶
Paper results¶
From 2026-01-02 through January 12, with 15% traffic per arm, revenue rose +1.79%, CTR +2.11%, and GMV +2.04%, all significant.
Local reproduction¶
Local comparison: The baseline is the GR backbone with the same SID. GRC adds 1,320 structured trajectories, 10 GRPO rounds, and EGRS. Relative Hit@10 is -12.50% and NDCG@10 -11.12%.
Mean GRPO reward rises 8.3989→8.4273 while full-catalog test ranking degrades. See the stable metrics.
Reproduction boundaries¶
MovieLens genre stands in for category, seller, and brand. Industrial LLM parameter scale is not reproduced.
