Shenghao Yang, Zhiqiang Guo, Weizhi Ma, Min Zhang, Lei Wang · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840848
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 has attracted widespread attention as a new paradigm in recommender systems. Its core idea is to discretize item semantic representations into semantic IDs and leverage the generative model's sequence modeling capability to predict the next item the user may prefer. Early approaches mainly focused on textual semantics, and recent multimodal generative recommendation methods incorporate additional modalities to provide richer semantics. However, existing methods typically handle modalities independently during both tokenization and generation, leading to insufficient modality fusion in semantic ID construction and inadequate cross-modal integration during recommendation. Moreover, the modality tokenizers are optimized with representation-level reconstruction or alignment objectives and lack recommendation-aware supervision, resulting in sub-optimal performance. To address these issues, we propose Accumulative Modality Tokenization for Multimodal Generative Recommendation, namely AMTGR. Specifically, we harness the Multimodal Large Language Model as the unified tokenizer to progressively accumulate and tokenize multimodal information, and design a collaborative-aware reconstruction task as the tokenizer's optimization objective to achieve recommendation-oriented multimodal semantic IDs, which will be jointly fed into the generative model to achieve effective cross-modal integration during recommendation. Experiments on public datasets demonstrate that AMTGR significantly outperforms baselines, and further analyses verify the effectiveness of high-quality multimodal semantic IDs and the unified modality integration.
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.