MM-VeriRec：面向可验证智能体多模态推荐的失败引导融合
原标题：MM-VeriRec: Failure-Guided Fusion for Verifiable Agentic Multimodal Recommendation
AI 摘要
研究者提出 MM-VeriRec，一个面向可验证智能体多模态推荐的基准协议与失败引导融合方法，针对隐藏视觉约束、图文不匹配和不可能任务弃权三类失败模式。该方法基于真实电影海报与商品图像数据集构建任务，用确定性视觉属性验证推荐结果，并将失败转化为可操作的标签（如文本陷阱跟随、视觉忽视、错误接受）。实验表明，更强的文本与视觉嵌入能提升检索但无法消除这些失败模式，而失败引导融合可以；在非对齐 CLIP 检测器下跨域迁移仍优于 VBPR 基线和普通融合。
正文节选
MM-VeriRec: Failure-Guided Fusion for Verifiable Agentic Multimodal Recommendation Abstract. Images often carry the recommendation constraints that text metadata only hints at: a movie may need to look dark, a product may need a minimal style, and a visually impossible request should be rejected rather than politely hallucinated. Agentic multimodal recommenders must therefore do more than retrieve plausible items; they must reason over text-image evidence, decide when visual evidence is decisive