Generative retrieval (GR) is an emerging paradigm for industrial search that frames candidate generation as an autoregressive task. A new arXiv paper, "WeGR: Towards Unified Generative Retrieval via Behavior-aware Tokenization and Preference Alignment," addresses a key limitation: how to effectively leverage user behaviors in this framework.
The authors propose WeGR, which introduces behavior-aware tokenization to incorporate user actions into the retrieval process. They also employ preference alignment to better align the model's output with user intent. The paper is listed as a replacement (v2) on arXiv, indicating revised content, though the abstract does not provide further experimental details.