Meta AI 提出 SJR 双模型架构解耦多模态内容审核
原标题：Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation
AI 摘要
Meta AI 提出 Summarize-Judge-Refine（SJR）双模型架构，将多模态内容理解与政策分类解耦：多模态 Content Model 生成结构化文本摘要，纯文本 Policy Model 依据政策定义对摘要分类。通过 GRPO 迭代协同训练和文本空间增强，在误导性广告检测任务上，SJR 相比零样本思维链基线非误导类 F1 相对提升 23.6%，并超过端到端 SFT、STaR/RFT 和 RLFT。值得注意的是，仅用合成正类数据、零真实违规样本训练的变体，在违规类 F1 上与全量数据模型相对差距仅 0.2%，表明新政策可在无真实违规数据下启动。
正文节选
Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation Abstract Content moderation systems traditionally entangle multimodal understanding with policy-specific classification, requiring full pipeline retraining for every policy change and suffering from label scarcity since multimedia cannot be meaningfully augmented. We propose Summarize-Judge-Refine (SJR), a two-model architecture that decouples these concerns via a natural language inte