HiGR: Hierarchical Generative Slate Recommendation¶

Learns two residual SID levels and runs coarse-to-fine decoding with ORPO-style reranking.

Paper information¶

Field	Details
Paper	arXiv 2512.24787
Institution	Tencent.
First public release	2025-12-31 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`higr`
Local reproduction code	`src/auto_research/reproductions/higr/`

Paper method and evidence¶

Itemwise generation overlooks slate interactions. PCRQ-VAE produces hierarchical semantic IDs. The hierarchical decoder picks clusters before items, and ORPO aligns whole-page preferences.

flowchart LR
  I[Items] --> Q[Hierarchical semantic IDs]
  Q --> D[Decode cluster then item]
  D --> R[Whole-slate ORPO]

Original paper figure¶

The original figure shows hierarchical slate generation. From the original paper; copyright remains with the authors.

Core equation¶

\[ i\mapsto(c_i^{(1)},c_i^{(2)}),\qquad \mathcal L_{\mathrm{ORPO}}=-\log p(y^+|x) -\lambda\log\sigma(\logit p(y^+|x)-\logit p(y^-|x)). \]

Reported results¶

On public KuaiRec and Tencent data, hierarchy and ORPO outperform flat generation. At 5% online traffic, stay time +1.03%, watch time +1.22%, video views +1.73%, request count +1.57%.

Local reproduction¶

Versus a flat item slate on MovieLens-1M, NDCG@10 0.07514→0.06588 (-12.32%), head share 0.21762→0.15714. Diversity improves while accuracy falls; online gains were not reproduced. Metrics.

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

Single next-item feedback is not video-slate feedback; k-means residual IDs are not production PCRQ-VAE.