Jinwei Luo, Dugang Liu, Xing Jia Tang, Xiuqiang He, Renjie Wei, Zhe Wu, Xin Huang, Siliang Huang, Hongcheng Fu · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840168
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).
Autoregressive retrieval (AR) has emerged as a leading paradigm for the first stage of industrial recommender systems. However, existing AR methods (i) overlook long-term action perception, training their generation only against the immediate next interaction, and (ii) lack the controllability needed to align candidate generation with diverse business goals without retraining. We propose Controllable Autoregressive Retrieval with long-term action perception (CAR), a simple yet deployable framework that addresses both issues. A long-term supervision module attaches multiple action-sliding prediction heads at different future horizons and transfers their knowledge to the main model via parameter sharing (P-CAR) or cascading (C-CAR), injecting a long-horizon perspective. An action-induced input module reorders the input as an action--item sequence and exposes a target action token at inference time, enabling goal-aware control over the recall set. Experiments on three public benchmarks and two industrial datasets, along with online A/B tests on QQ Music and WeSing, which now serve hundreds of millions of daily users in production, show consistent, statistically significant gains over strong AR baselines.
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.