Merge and Branch (M&B) is a tree‑structured supervised fine‑tuning method for large language models (LLMs) in recommendation systems. It compresses each historical item’s semantic identifier (SID) and text into a few learned tokens (Merge) and then uses the compressed history to predict multiple target SIDs (Branch), organizing each example as a shallow, wide tree. The approach includes a tree FlashAttention mechanism for efficient training and a tree‑structured key/value cache for constrained decoding, achieving better recommendation performance and improved training and decoding efficiency compared to strong baselines.
arXiv:2608. 07816v1 Announce Type: cross Abstract: Recent advances in generative recommendation (GR) leverage large language models (LLMs) as recommender backbones, enabling LLMs to directly generate recommendations conditioned on item-interaction histories.
arXiv:2608. 09605v1 Announce Type: cross Abstract: Large Language Models (LLMs) have emerged as powerful tools for improving recommendation systems.
arXiv:2607. 24845v1 Announce Type: cross Abstract: Large language models (LLMs) have been applied to sequential recommendation by formulating it as a natural language task.
arXiv:2608. 11980v2 Announce Type: replace-cross Abstract: Semantic-ID generative recommenders represent each item as a short sequence of discrete semantic tokens and predict the next item by autoregressively generating this token sequence.
arXiv:2608. 11980v1 Announce Type: cross Abstract: Semantic-ID generative recommenders represent each item as a short sequence of discrete semantic tokens and predict the next item by autoregressively generating this token sequence.
arXiv:2606. 25147v1 Announce Type: cross Abstract: User modeling in industrial recommender systems typically produces dense embeddings, which suffer from representational constraints inherent to fixed-dimensional vectors.
arXiv:2607. 25209v1 Announce Type: cross Abstract: Generative recommendation commonly represents items using fixed-length semantic identifiers (SIDs) constructed through clustering and quantization.
arXiv:2512. 10388v3 Announce Type: replace-cross Abstract: Conventional Sequential Recommender Systems (SRS) typically assign unique hash IDs (HID) to construct item embeddings, which mainly capture collaborative signals from historical user-item interactions.
arXiv:2511.22707v2 Announce Type: replace-cross Abstract: In web environments, user preferences are often refined progressively as users move from browsing broad categories to exploring specific item...
arXiv:2607. 24865v1 Announce Type: cross Abstract: Large-scale recommendation systems face "Memory Wall" bottlenecks due to massive, dense embedding tables.
JevRec introduces a variable‑length linguistic encoding for items in generative recommendation, assigning shorter semantic identifiers (SIDs) to popular items and richer SIDs to long‑tail items. It separates length prediction and content generation into distinct heads while sharing the interaction‑history representation, mitigating gradient conflicts. Experiments on three public datasets demonstrate that JevRec surpasses baseline models in recommendation metrics and reduces training and inference time compared to TIGER.
REGER is a unified framework that enhances generative recommendation by explicitly modeling inter-item relations through semantic pretraining, history encoding, and SID generation. It pretrains the model to generate semantic neighbors’ IDs, uses combined global and local attention to capture historical context, and introduces a hierarchical preference decomposition over the SID prefix tree. The approach also distills a discriminative teacher’s full conditional distributions to improve next-token predictions, yielding consistent performance gains on real-world datasets.