PRISM: Pruning-Enhanced Intent Sampling for Recommendation Multi-List Generation
Abstract
Generative recommendation has emerged as a prominent paradigm, unifying retrieval and ranking through large-scale sequence transducers. However, existing generative recommenders are basically pointwise: they optimize for next-item prediction, leaving the construction of the final exposure list---a slate---to heuristic post-processing. This creates a structural gap between pointwise candidate generation and industrial listwise delivery, which demands holistic reasoning over contextual opportunity, position bias, and intent uncertainty. To bridge this gap, we introduce List Composer, a decoupled module that transforms pointwise candidate pools into deliverable exposure lists within generative recommender systems. We instantiate this concept through PRISM, an autoregressive listwise generation model featuring two core mechanisms. First, a CVAE-based IntentSampler module models user interests as a latent distribution, sampling distinct intent facets as initial states averts the Deterministic Collapse phenomenon, wherein standard autoregressive generators converge onto nearly identical sequences despite repeated decoding. Second, a Listwise Pruning strategy dynamically culls low-utility beam paths using position-discounted pointwise scores, concentrating computational resources on promising trajectories. Extensive offline experiments and real-world online A/B tests validate the effectiveness of PRISM, which achieves consistent improvements in recommendation quality and serving efficiency, including a 0.288% gain in video views on a major short-video platform.