APAO: Adaptive Prefix-Aware Optimization for Generative Recommendation

A research team led by Yuanqing Yu has introduced APAO (Adaptive Prefix-Aware Optimization), a novel framework designed to solve a critical problem in generative recommendation systems. These systems, which treat recommendation as an autoregressive generation task, suffer from a fundamental mismatch: they're trained with token-level objectives like cross-entropy loss but use multi-step beam search during inference. This creates a training-inference gap where correct items can be discarded early simply because their initial tokens (prefixes) receive low probability scores during beam search pruning. The standard training assumption that ground-truth history is always available doesn't hold in real inference scenarios.

APAO addresses this by introducing prefix-level optimization losses that better align training objectives with inference constraints. The framework's key innovation is an adaptive worst-prefix optimization strategy that dynamically identifies and focuses training on the most vulnerable prefixes—those most likely to be pruned incorrectly. This enhances the model's ability to retain correct candidates throughout the beam search process. Extensive experiments across multiple datasets demonstrate that APAO consistently improves performance when integrated with various generative recommendation backbones, while theoretical analyses confirm both effectiveness and computational efficiency. The publicly available implementation allows immediate application to existing recommendation systems.