RecoReward: Train multimodal descriptions using recommender rewards¶

Core reward mechanism with public-data comparison, not the production reward pipeline.

Paper information¶

Field	Details
Paper	arXiv 2607.25901
Institution	Kuaishou / Nankai University / Chinese Academy of Sciences.
First public release	2026-07-28 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`reco-reward`
Local reproduction code	`src/auto_research/reproductions/reco_reward/`

Paper method and evidence¶

Freeze a two-tower recommender. Reward a description by how much more it appeals to target than non-target users. Online serving consumes descriptions without user profiles.

flowchart LR
  D[Candidate descriptions] --> S[Frozen two-tower affinity]
  U[Target and non-target users] --> S
  S --> R[Affinity difference reward]

Original paper figure¶

The original figure shows recommender-guided caption training. From the original paper; copyright remains with the authors.

Core equation¶

\[ R_{\mathrm{RAS}}(d,u)=s(f(d),g(u))-\lambda\,\mathbb E_{u^-}s(f(d),g(u^-)). \]

Reported results¶

The paper reports 31.7%–40.4% higher offline Recall than Qwen. One-week Kuaishou A/B: key-page effective user penetration +0.265%, outbound exposures +0.791%, outbound users +0.740%.

Local reproduction¶

Compared with content-only semantic retrieval, RAS selection yields Hit@10 -16.00%, NDCG@10 -9.68%, and head share -24.14%; this is a negative result. On MovieLens-100K, history-item centroids proxy the user tower. Run target/non-target subtraction, score 620 candidates, and serve content only.

auto-research reproduce --paper reco-reward --data-root data --seed 42

Stable result. No Qwen3.5-9B tuning, livestream video, or private Kuaishou behavior; verifies executable RAS, not ranking gains.