aiqbench analysis
This research introduces Summarize-Judge-Refine (SJR), a novel architecture that decouples multimodal understanding from policy classification to solve label scarcity in content moderation. By generating human-readable summaries and leveraging text-space augmentation, the system enables few-shot learning for new policies without requiring real violation data. The approach significantly improves detection performance while providing inherent interpretability through structured textual grounding.
Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation
arXiv:2609.22094v1 Announce Type: new 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 interface: a multimodal Content Model produces structured text summaries, and a text-only Policy Model classifies them against policy definitions. An iterative co-training loop refines the Content Model via GRPO to produce policy-relevant summaries, while text-space augmentation generates adversarial summary variants---an augmentation pathway impossible on raw multimedia---enabling few-shot policy bootstrap. Every decision is grounded in a human-readable summary, providing interpretability as a structural byproduct. On misleading advertisement detection, SJR achieves +23.6\% relative non-misleading F1 over a zero-shot chain-of-thought baseline, outperforming end-to-end SFT, STaR/RFT, and RLFT. Notably, a variant trained on zero real violating examples---with all positive-class data synthetically generated---matches the full-data model within 0.2\% relative on violating F1, demonstrating that new policies can launch without any real violation data.
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 interface: a multimodal Content Model produces structured text summaries, and a text-only Policy Model classifies them against policy definitions. An iterative co-training loop refines the Content Model via GRPO to produce policy-relevant summaries, while text-space augmentation generates adversarial summary variants---an augmentation pathway impossible on raw multimedia---enabling few-shot policy bootstrap. Every decision is grounded in a human-readable summary, providing interpretability as a structural byproduct. On misleading advertisement detection, SJR achieves +23.6\% relative non-misleading F1 over a zero-shot chain-of-thought baseline, outperforming end-to-end SFT, STaR/RFT, and RLFT. Notably, a variant trained on zero real violating examples---with all positive-class data synthetically generated---matches the full-data model within 0.2\% relative on violating F1, demonstrating that new policies can launch without any real violation data.
Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2609.22094 [cs.CL]
(or arXiv:2609.22094v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.22094 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 258 words
