Pengjie Zhou, Weiqin Yang, Haibo Xing, Hao Deng, Jinxin Hu
Long user behavioral sequences improve the performance of generative recommendation (GR), but computing full attention over long histories during autoregressive beam search decoding incurs substantial online inference latency. To accelerate long-sequence decoding, existing approaches typically rely on static history compression or dynamic top-k sparse attention; however, the former risks losing critical semantics, whereas the latter incurs significant overhead from repeated token selection and non-contiguous memory access. In this work, we thoroughly analyze attention patterns during GR decoding and reveal two key properties: (i) progressive attention focusing, where attention progressively focuses on a small subset of critical historical tokens; and (ii) cross-step ranking consistency, where their importance rankings remain relatively stable across subsequent decoding steps. Motivated by these observations, we propose Progressively Focused Attention (PFA), an efficient sparse decoding framework for long-sequence GR. PFA retains full attention at the initial decoding step to explore global historical preferences. It then establishes a shared importance ranking through cross-beam weighted aggregation and reorders the key-value cache of the selected tokens into a contiguous memory layout. Subsequent decoding steps reuse this ranking and geometrically reduce the token budget, preserving critical semantics while lowering computational and memory access overhead. Extensive experiments on two public benchmarks and a large-scale industrial dataset demonstrate that PFA outperforms state-of-the-art baselines. A 7-day online A/B test on a large-scale e-commerce advertising platform further confirms that PFA delivers substantial business gains while maintaining low inference latency. Our code is publicly available at https://anonymous.4open.science/r/PFA-review-088D.
Coordinating multiple recommendation objectives (e.g., click and conversion) is critical for real-world generative recommendation (GR). However, since a positive item in GR can be retrieved only when its tokens are successively selected at every decoding step, traditional item-wise methods lack fine-grained token-wise supervision to guide intermediate steps. Motivated by its success in foundation models, we consider introducing multi-objective on-policy distillation (MOPD) into GR to distill two single-objective teachers into a student at each token. Yet, we find that in vanilla MOPD, (i) teachers exhibit unequal distillation difficulty at each token (i.e., token-level imbalance), and (ii) the harder-to-distill teacher varies across tokens (i.e., cross-token inconsistency), leaving each objective poorly learned at certain steps and hindering positive item retrieval for both objectives, leading to suboptimal performance. To address this challenge, we propose KL-MOPD (KL-Weighted Multi-Objective On-Policy Distillation), a simple yet effective distillation method that allocates token-wise teacher weights proportional to the initial student-teacher token-wise KL divergences, thereby strengthening supervision for the harder-to-distill teacher at every decoding step. Despite its simplicity, we theoretically prove that KL-MOPD strictly outperforms vanilla MOPD in the additive joint recommendation objective under mild conditions. Moreover, KL-MOPD is plug-and-play and computationally efficient, requiring only a one-time pre-training weight calibration with negligible training overhead. Extensive experiments on three public benchmarks, a billion-scale industrial dataset, and online A/B tests on our e-commerce advertising platform demonstrate that KL-MOPD significantly improves both individual and joint recommendation performance. The anonymous code is provided for peer review.