S2GR: Stepwise Semantic Guidance in Latent Space¶

Core mechanism evidence. Paper and local evaluations are separate.

Paper information¶

Field	Value
Paper	arXiv 2601.18664
Affiliations	Kuaishou Technology
First public release	2026-01-26 (arXiv v1)
Authors' code	No public author implementation found as of 2026-09-05.
Adapter / method	`s2gr`
Local code	`src/auto_research/reproductions/s2gr/`

Original paper¶

Background and contribution¶

Insert interpretable thinking tokens before each SID level and align with codebook distributions contrastively. Co-occurrence and load balancing improve the coarse-to-fine hierarchy.

flowchart LR
  H[User history] --> T[Thinking token] --> S[Next semantic ID level] --> R[Repeat]
  C[Codebook and co-occurrences] --> T

Key figure from the paper¶

Original Figure 3: S2GR. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ p(z_l\mid z_{<l},t_{\le l},u),\qquad \mathcal L=\mathcal L_{gen}+\lambda\mathcal L_{think}+\gamma\mathcal L_{balance}. \]

Paper results¶

Paper core online recommendation metric +5.25%, not local gain.

Local reproduction¶

Matched full-catalog MovieLens NDCG@10 0.05401→0.05463 (+1.15%). Three-seed metrics, SD, and 95% CI.

auto-research reproduce --paper s2gr --dataset-dir data --seeds 42,43,44

Reproduction limits¶

Fixed public MovieLens 100K subset and auditable proxy only. No private logs, production base models, bucketing, or serving.