SJR
An architectural paradigm that decouples content understanding from policy classification in multimodal content moderation. A content model generates textual summaries, while a policy model performs classification based on those summaries. The two components are jointly trained using GRPO, supporting few-shot cold-start scenarios.
Timeline (last 90 days)
The SJR (Summarize-Judge-Refine) architecture decouples content understanding and policy judgment into two independent models, using natural language as the interface between them.
The SJR architecture consists of two components: a content model (a multimodal model that converts raw content into structured text summaries) and a policy model (a text-only model that performs classification judgments against policy definitions).
SJR uses GRPO (Group Relative Policy Optimization) to refine the content model so that it generates policy-relevant summaries.
SJR performs data augmentation by generating adversarial summary variants in the text space, enabling few-shot policy cold-start.
On a misleading advertisement detection task, SJR achieved a 23.6% relative improvement in F1 score on the 'non-misleading' category compared to the zero-shot chain-of-thought baseline.
SJR comprehensively outperformed multiple training methods including end-to-end SFT, STaR/RFT, and RLFT in experiments.
One SJR training variant uses no real violating samples at all (all positive-class data is synthetically generated), and its violation-class F1 score differs from the model trained on full real data by less than 0.2% in relative terms.
The SJR architecture enables new policies to be launched without any real violating data.
In the SJR architecture, every moderation decision is grounded in a human-readable text summary, making interpretability a natural product of the architectural design.
When policies change, the SJR architecture requires only adjusting or replacing the text-only policy model, without touching the multimodal content model.
All Facts (10)
The SJR (Summarize-Judge-Refine) architecture decouples content understanding and policy judgment into two independent models, using natural language as the interface between them.
50%UnverifiedThe SJR architecture consists of two components: a content model (a multimodal model that converts raw content into structured text summaries) and a policy model (a text-only model that performs classification judgments against policy definitions).
50%UnverifiedSJR uses GRPO (Group Relative Policy Optimization) to refine the content model so that it generates policy-relevant summaries.
50%UnverifiedSJR performs data augmentation by generating adversarial summary variants in the text space, enabling few-shot policy cold-start.
50%UnverifiedOn a misleading advertisement detection task, SJR achieved a 23.6% relative improvement in F1 score on the 'non-misleading' category compared to the zero-shot chain-of-thought baseline.
50%UnverifiedSJR comprehensively outperformed multiple training methods including end-to-end SFT, STaR/RFT, and RLFT in experiments.
50%UnverifiedOne SJR training variant uses no real violating samples at all (all positive-class data is synthetically generated), and its violation-class F1 score differs from the model trained on full real data by less than 0.2% in relative terms.
50%UnverifiedThe SJR architecture enables new policies to be launched without any real violating data.
50%UnverifiedIn the SJR architecture, every moderation decision is grounded in a human-readable text summary, making interpretability a natural product of the architectural design.
50%UnverifiedWhen policies change, the SJR architecture requires only adjusting or replacing the text-only policy model, without touching the multimodal content model.
50%