Automatic Hard Example Synthesis with Multi-Level Agentic Data Curation

2026-07-15

AI Summary

This work addresses the vulnerability of multimodal large language models (MLLMs) in content safety moderation to adversarial attacks and out-of-distribution edge cases, which are challenging for conventional human annotation and active learning approaches to handle efficiently. To overcome this limitation, the authors propose the first fully automated multi-agent red-teaming framework that operates without human intervention. The framework integrates an Architect reasoning agent, an advanced image generator, and an LLM-based scoring committee to iteratively synthesize high-difficulty adversarial examples. Coupled with a test-time retrieval-augmented mechanism, it enables automated curation of high-quality safety-critical data. Evaluated on public image safety benchmarks, the method significantly reduces the target model’s false negative rate from 41.2% to 24.5%, substantially enhancing moderation robustness.

Abstract

Multimodal Large Language Models (MLLMs) are increasingly deployed for nuanced content safety and moderation tasks, yet they remain vulnerable to adversarial attacks and out-of-distribution edge cases. Traditional active learning and manual annotation fail to scale against the complexity and volume of novel multimodal threats. In this paper, we propose an automated, agentic red-teaming framework that systematically synthesizes difficult examples using an iterative strategy that proposes novel hypotheses as well as mutating on past attempts. Leveraging a multi-agent architecture that consists of a high-reasoning Architect agent, an advanced image generator, and a multi-level verification committee of LLM raters, our system autonomously uncovers boundary-pushing violations and ambiguous policy edge cases without any human intervention. By employing these carefully synthesized adversarial examples as in-context demonstrations via test-time Retrieval, we substantially improve the target model's robustness, reducing the False Negative Rate (FNR) from 41.2% to 24.5% in a public image safety benchmark without relying on any human labeling.

Problem

• Multimodal Large Language Models
• adversarial attacks
• out-of-distribution edge cases
• content safety
• hard example synthesis

Innovation

• Agentic Red-Teaming
• Hard Example Synthesis
• Multi-Agent Architecture
• Multimodal Adversarial Robustness
• Automated Data Curation