CoGen: A Two-Stage Unified Multi-modal Framework for Bundle Construction via Discrete Semantics Transfer

Bundle construction aims to assemble sets of mutually compatible items that collectively satisfy user preferences. This problem naturally encompasses two interconnected tasks: bundle completion, which captures intra-bundle compatibility, and bundle generation, which constructs personalized bundles. However, existing approaches typically treat these tasks in isolation, neglecting their inherent synergy: completion models learn compatibility in entangled continuous spaces that are difficult to transfer, while generation models often suffer from sparse supervision and lack explicit compatibility constraints, leading to semantically incoherent bundles. To bridge this gap, we propose CoGen, a two-stage unified multi-modal framework for bundle construction. The core idea of CoGen is to transform compatibility-aware continuous representations learned from completion into discrete semantics to guide bundle generation. Specifically, CoGen first learns disentangled multi-modal item representations and performs bundle completion to capture intra-bundle compatibility. The learned representations are then discretized via residual quantization (RQ) to obtain explicit discrete semantics that make compatibility patterns reusable for generation. To further improve bundle generation quality, CoGen adopts a semantic-anchored non-autoregressive generator to generate bundles based on the learned discrete semantics. Extensive experiments on four public multi-modal datasets verify the effectiveness of CoGen compared to several state-of-the-art approaches. Our code is available at https://github.com/y2q109/CoGen.