Disagreement-Guided Agentic Data Curation for Multimodal Safety
Genglin Liu, Muye Zhang, Krishnamurthy Viswanathan, Nichole J. Hansen, Blaž Bratanič, Nathan L Clement, Shalini Ghosh, Ariel Fuxman
原文摘要
Written safety policies in multimodal settings define categories whose visual boundaries span many contexts. We study disagreement as a signal for identifying useful synthetic boundary cases. Our pipeline generates policy-conditioned images, selects cases with target--rater or inter-rater disagreement, conditionally escalates unresolved cases, and retrieves the resulting rater-labeled examples as in-context demonstrations. On a failure-enriched multimodal safety benchmark, policy grounding reduces false negative rate (FNR) from 0.875 to 0.413 and raises F1 from 0.220 to 0.731; retrieval from 50 disagreement-selected examples further reduces FNR to 0.245, compared with 0.297 from a 350-example Regular pool. The selected memory also improves a different target-model family, with a smaller gain on a proprietary benchmark with a refined policy. In a blinded human study, experts disagree on 24% of disagreement-guided images and 17% of baseline-curated images. These findings support traceable, disagreement-guided curation for multimodal safety evaluation and example selection.
cs.AI · cs.MA