Reproduce “Self-Evolving Memory for Generative Recommendation” on OpenResearch

About this paper

Xinyu Lin, Zhuosong Jiang, Zixiao Suo, Siqin Wang, Hanqing Zeng, Hanchao Yu, Yinglong Xia, Jiang Zhang and 7 more

LION introduces a sparse trainable memory for continual generative recommendation, routing users to a small subset of shared key-value slots while applying an auxiliary consolidation objective. Across three Amazon Reviews datasets, it improves TIGER’s Recall@10 by 15.6–35.5% and produces larger gains for inactive users, suggesting that separating behavioral update pathways can reduce interference during streaming adaptation.

Key insights

• The paper reframes continual adaptation as an evolution-conflict problem: common and uncommon behaviors can generate opposing gradients, with frequent patterns dominating updates because they occur more often.
• Sparse routing limits cross-pattern interference to the memory slots shared by their activation sets, while fixed memory size avoids parameters growing in proportion to the user population.
• The auxiliary loss is effective only when paired with isolated memory pathways; applying similar supervision to a fully shared backbone can worsen gradient conflict.

Summary by alphaXiv.