SRPFN: One model pretrained on synthetic data works across recommendation domains

Existing sequential recommendation models are trained per dataset, making generalization to new domains expensive — requiring full retraining from scratch. A new paper from KAIST researchers introduces SRPFN (Prior-data Fitted Network), which sidesteps this limitation entirely. SRPFN is pretrained offline on 25.6M synthetic sequences that simulate diverse item-to-item transition patterns. Instead of fitting to a specific item catalog, it learns to produce posterior predictive distributions for next-item recommendations. At inference, the model adapts to a new domain using only a support set of item-item transition examples from the target dataset — no gradient updates, no fine-tuning.

Compared against 10 baseline models across 5 real-world benchmarks, SRPFN consistently delivers the best or second-best performance on nearly all metrics. Crucially, this performance comes with substantially lower computational cost, as the model operates in a single forward pass. The work, accepted at KDD 2026, demonstrates that a single pretrained model can generalize across diverse domains without any update step. This opens the door to zero-shot recommendation systems that can be deployed instantly on new user bases or product catalogs, dramatically reducing the engineering overhead of personalized AI.