Figureure 2 · : The training pipeline, where (a) presents the training data construcFigure 2: The training pipeline, where (a) presents the training data construction strategy designed for item-to-item retrieval, including retrieval sample pairs and hard negative mining strategy; (b) represents the training stage 1: contrastive supervised fine-tuning (SFT); and (c) represents the training stage 2: relative rerank via reinforcement learning (RL).这张图概括 UniNote 的整体方法流程。阅读时先看模块之间传递的训练信号,再看作者如何把目标拆成可优化的子问题。Figureure 1 · : We propose UniNote, a unified retrieval-ranking framework specificalFigure 1: We propose UniNote, a unified retrieval-ranking framework specifically designed for I2I retrieval, which is capable of encoding various modality inputs into a unified space with cross-modal alignment, localized retrieval, and relevance-aware ranking capabilities.这张图概括 UniNote 的整体方法流程。阅读时先看模块之间传递的训练信号,再看作者如何把目标拆成可优化的子问题。