Abstract
Conversion-oriented generative recommendation (GR) is challenged by sparse conversion supervision, where the decoder must learn conversion targets from long interaction histories with limited positive signals. Existing multi-behavior GR methods enhance behavior-aware representation learning, but still optimize the original full-history prediction task directly, leaving the difficulty of sparse conversion modeling insufficiently addressed. Recent reverse-curriculum-inspired methods further show that conversion-related intermediate behaviors are useful, yet they mainly instantiate the curriculum as one-shot decoder-prefix conditioning, which does not provide a staged reverse curriculum process and will increase online serving latency. To address these limitations, we propose SRCRec, a staged reverse curriculum learning framework for conversion-oriented GR. SRCRec constructs difficulty-aware curriculum contexts over the historical sequence, trains the model from target-near easy contexts to harder historical contexts, and finally recovers the original no-prefix conversion objective. This staged optimization enables the model to progressively learn from curricula with increasing difficulty while keeping inference identical to standard GR. Experiments on public and industrial datasets show that SRCRec achieves superior offline performance, and online A/B tests further improve orders by 3.52% and Gross Merchandise Volume by 2.98%.