GLIDE: Generative Retrieval with Short- and Long-Term Interests¶

Autoregressive residual semantic IDs and two prompt timescales run.

Paper information¶

Field	Value
Paper	arXiv 2603.17540
Affiliations	Spotify
First public release	2026-03-18 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`glide`
Local code	`src/auto_research/reproductions/glide/`

Original paper¶

Background and contribution¶

Recent-only histories retrieve habitual content. Generate item IDs conditioned on recent behavior and a long-term user embedding, balancing immediate intent and discovery.

flowchart LR
  H[Recent history] --> D[Prompted decoder]
  U[Long-term embedding] --> D
  C[Residual semantic IDs] --> D --> R[Constrained retrieval]

Key figure from the paper¶

Original Figure 1: GLIDE. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ p(c_{1:L}\mid h_{\rm recent},u_{\rm long}) =\prod_{\ell=1}^L p(c_\ell\mid c_{<\ell},h_{\rm recent},u_{\rm long}). \]

Paper results¶

Spotify's live experiment reports +5.4% non-habitual streaming and +14.3% new-show discovery.

Local reproduction¶

Three-level residual quantization uses public item features, with code transitions and short/long profiles. Traditional transition/content retrieval gives NDCG@10 0.03540 versus 0.03437 (−2.89%) for dual-prompt retrieval; see metrics/movielens-100k-seed42.json.

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

Reproduction limits¶

The public-data regression is not the Spotify production result.