Researchers introduce SJR architecture that improves multimodal moderation by 23.6% F1
A new arXiv paper presents Summarize‑Judge‑Refine (SJR), a two‑model system that separates multimodal content understanding from policy classification. The multimodal Content Model first creates structured text summaries of images, video or audio, which are then evaluated by a text‑only Policy Model against defined rules. An iterative co‑training loop refines the Content Model using GRPO, while…
Key points
- SJR decouples multimodal content analysis from policy classification via a content model and a text‑only policy model.
- On misleading ad detection, SJR improves non‑misleading F1 by 23.6% relative to a zero‑shot chain‑of‑thought baseline.
- A synthetic‑only training variant matches full‑data performance within 0.2% relative violating F1.
In experiments on misleading advertisement detection, SJR raises the non‑misleading F1 score by 23.6% relative to a zero‑shot chain‑of‑thought baseline and outperforms end‑to‑end fine‑tuning, STaR/RFT and RLFT approaches. Notably, a version trained without any real violating examples—using only synthetically generated positive data—matches the full‑data model within 0.2% relative on violating F1, suggesting new policies can be deployed without real violation data. The authors highlight that every decision is backed by a human‑readable summary, improving interpretability.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.
More in Research
All →- DeepInstructor agentic framework improves idea evaluation using experience graph from 58,607 peer reviews · 1 src
- Researchers identify context poisoning as extreme-value attention interference in long-context language models · 1 src
- Indy Autonomous Challenge features first road-course overtakes, Unimore wins · 1 src
- Researchers unveil TangleDiff to design entangled protein hydrogels · 1 src
- Anthropic reports Claude leads 26% of its AI research and development tasks · 3 src
Comments
via GitHub Discussions