DEGR: Diversity constraints and adaptive reward for generative reranking¶

Next-item CE, cohort diversity, reward-weighted ORPO and greedy reranking.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	JD.com.
First public release	2026-08-05 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`degr`
Local reproduction code	`src/auto_research/reproductions/degr/`

Paper method and evidence¶

Add squared cosine penalties within a cohort to prevent repeated similar exposures. Exploration feedback adjusts ORPO reward weights, and inference selects greedily with diversity constraints.

flowchart LR
  A[Generate slate] --> B[Penalize cohort similarity]
  B[Penalize cohort similarity] --> C[Reward-adaptive ORPO]
  C[Reward-adaptive ORPO] --> D[Diversity-aware inference]

Original paper figure¶

Figure 3 shows the reranking architecture. From the original paper; copyright remains with the authors.

Core equation¶

\[ \mathcal L=\mathcal L_{CE}+\lambda_d\sum_{i<j}\cos^2(e_i,e_j)+\lambda_p\,w(r)\log\!\left(1+e^{-(s^+-s^-)}\right). \]

Reported results¶

JD online A/B: UCTR +1.22%, PV +0.20%; qualifies under the industrial evidence criterion.

Local reproduction¶

Matched CE generator vs DEGR on MovieLens-1M, 260 users/420 items, 50 steps, seed 42. Both NDCG@10 0.01546 (+0.00%); head-share@10 0.06346→0.06269 (-1.21%). 2,400 reward-ORPO and 67,200 diversity-pair updates execute. Metrics. No cross-request exposure/reward; inverse popularity proxies exploration. No JD billion-request reward model or online gain.

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