RaG: Recommendation as Generation¶

Core D-SID interface only; no production video generator.

Paper information¶

Field	Details
Paper	arXiv 2606.25496
Institution	Kuaishou / Beihang University.
First public release	2026-06-24 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-09-02.
Local adapter or method key	`rag-generation`
Local reproduction code	`src/auto_research/reproductions/rag_generation/`

Paper method and evidence¶

Decoupled semantic IDs connect generative recommendation with interest-conditioned video generation; cross-domain rewards align candidate selection and content creation.

flowchart LR
  I[Item semantics] --> D[Decoupled semantic IDs]
  D --> R[Generative retrieval]
  D --> V[Interest-conditioned video generation]
  R --> A[Cross-domain reward]
  V --> A

Original paper figure¶

The original figure connects retrieval and creation. From the original paper; copyright remains with the authors.

Core equation¶

\[ z_i=(z_i^{content},z_i^{creative}),\qquad v\sim p_\theta(v|z_{interest}). \]

Reported results¶

Kuaishou ad A/B reports revenue +1.870% versus strong GRM and +5.462% versus DLRM. Paper evidence only.

Local reproduction¶

auto-research reproduce --paper rag-generation --dataset-dir data --seed 42

Same MovieLens-1M public-subset full-catalog protocol: NDCG@10 0.05401→0.04235 (-21.59%). Three-seed measurements, mean, SD and 95% CI. Proxy objectives verify only the interface, not private logs, traffic assignment, model scale, serving, or video synthesis.