Shunhang Zheng, Pengfei Jiao, Pinlong Zhao, Haocheng Yang, Dalin Zhang, Yuling Wang · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840997
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Generative recommendation leverages Large Language Models (LLMs) to model item semantics and user preference dynamics through sequence-to-token generation, but deep Transformer layers and wide Feed Forward Network (FFN) dimensions introduce substantial memory and latency overhead in online recommendation scenarios. Existing LLM pruning methods are mainly designed to preserve general language modeling ability, often overlooking the recommendation-specific states shaped by item semantics, collaborative signals, and temporal preference dynamics. To address this limitation, we propose Co-JDPRec, a collaborative joint dynamic pruning framework that shifts the pruning objective from parameter removal to recommendation-state preservation. Specifically, Co-JDPRec introduces a task-aware saliency criterion that combines collaborative alignment and sequence-order sensitivity, enabling the model to identify neurons critical for user-item relations and evolving preference dynamics. Guided by this saliency, Co-JDPRec jointly searches FFN width and layer depth to derive a compact heterogeneous architecture, followed by distillation recovery to restore recommendation performance. Theoretically, Co-JDPRec characterizes structural importance via first-order Taylor sensitivity under recommendation-aware objectives. Empirically, it achieves superior accuracy–efficiency trade-offs on both title-based and semantic-ID-based generative recommendation backbones, validating the effectiveness of recommendation-state-aware pruning.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.