Stochastic primal–dual decoding for multi-objective generative recommendation¶
Core public-data decoding mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Spotify. |
| First public release | 2026-05-26 (arXiv v1) |
| Author code | No public author code found as of 2026-09-02. |
| Local adapter or method key | primal-dual-decoding |
| Local reproduction code | src/auto_research/reproductions/primal_dual_decoding/ |
Paper method and evidence¶
Without retraining, add auxiliary objective contributions to relevance logits at each decoding position. Mirror descent updates Lagrange multipliers from remaining constraint slack.
flowchart LR
A[Relevance and auxiliary logits] --> B[Constrained token selection]
B[Constrained token selection] --> C[Mirror-descent multiplier updates]Original paper figure¶
Figure shows multi-objective decoding. From the original paper; copyright remains with the authors.
Core equation¶
\[
j^*=\arg\max_j\\{p_\theta(t_j\mid x,t_{<k})+\lambda_k b(t_j)\\},\qquad \lambda_{k+1}=\lambda_k e^{-\eta(b(t_{j^*})-R/K)}.
\]
Reported results¶
Spotify large-scale A/B reports +1.8% on the auxiliary objective with no user satisfaction loss.
Local reproduction¶
MovieLens-1M full catalog: NDCG@10 0.05401→0.02109 (-60.95%). Three-seed metrics, mean, SD, and 95% CI. No private logs, production models, bucketing or serving. Public proxies cannot establish online lift.