Abstract: Existing multimodal object re-identification (Re-ID) methods are largely limited to the shallow fusion of heterogeneous visual features, while largely overlooking the semantic logic embedded in textual descriptions and its potential to resolve ambiguities in visual representations. Although recent multimodal large language models have enabled automatic caption generation, the generated unstructured text still contains noise, making it difficult to provide reliable guidance for accurate cross-modal visual alignment. To address these issues, a multimodal object Re-ID method based on the synergistic deduction of semantic logic and topology is proposed. First, a standardized multimodal caption generation process is designed, in which DeepSeek-VL is combined with attribute constraints to extract semantically precise textual annotations from multimodal images. Second, a semantic logic deduction module is introduced to bridge the gap between textual semantics and visual representations. It transforms structured semantic attributes into visual guidance and exploits cross-modal responses to achieve semantic correction and adaptive alignment of identity-related regions in the feature space. Third, a semantic topology deduction module is developed to further refine the retrieval space by exploiting inter-sample relational topology. Extensive experiments on RGBNT201, MSVR310, and RGBNT100 demonstrate that the proposed method achieves competitive overall performance compared with representative state-of-the-art approaches.