RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates

We introduce RedDebate, a novel multi-agent debate framework that provides the foundation for Large Language Models (LLMs) to identify and mitigate their own unsafe behaviors. Existing AI safety approaches often rely on costly human evaluation or isolated single-model assessment, both constrained by scalability and prone to oversight failures. RedDebate employs collaborative argumentation among multiple LLMs across diverse debate scenarios, enabling them to critically evaluate one another’s reasoning and systematically uncover unsafe failure modes through fully automated red-teaming. We further integrate distinct long-term memory modules that preserve safety-relevant insights from debate interactions and leverage them during subsequent inference, facilitating continuous refinement of model behavior. Empirical evaluation on safety benchmarks across a diverse set of models demonstrates that RedDebate substantially reduces unsafe outputs. While debate alone allows LLMs to refine their behavior, the addition of memory modules yields further significant reductions. To the best of our knowledge, RedDebate is the first fully automated framework to unify multi-agent debate and red-teaming to progressively enhance LLM safety without human intervention.

Large language models (LLMs) are becoming increasingly capable, but they can still produce harmful, biased, or unsafe responses. Improving their safety often depends on extensive human oversight, which can be expensive, slow, and difficult to scale. In this work, we introduce RedDebate, a fully automated system that helps AI models identify and reduce their own unsafe behavior by debating with one another. In RedDebate, multiple AI agents engage in structured discussions where they challenge, critique, and evaluate each other’s responses across a wide range of harmful or risky scenarios. Through these debates, the models are able to uncover weaknesses and unsafe reasoning that a single model might miss on its own. The system also includes long-term memory components that store important safety lessons learned during previous debates and reuse them in future interactions, allowing the models to continuously improve over time. We evaluate RedDebate on several AI safety benchmarks using different language models and show that it significantly reduces unsafe outputs. Our results demonstrate that debate alone can improve model safety, while adding memory leads to even greater improvements. To the best of our knowledge, this is the first fully automated framework that combines multi-agent debate and AI red-teaming to progressively improve language model safety without requiring human intervention.