Title: RedDebate: Enhancing Safety via Multi-Agent Red Teaming Debates

Abstract:

This paper presents RedDebate, an innovative multi-agent debate framework designed to help Large Language Models (LLMs) detect and address unsafe behaviors. Current AI safety strategies frequently depend on expensive human assessments or isolated single-model evaluations, methods that suffer from limited scalability and are vulnerable to oversight errors. To overcome these challenges, RedDebate utilizes collaborative argumentation among several LLMs in various debate contexts. This approach allows models to critically scrutinize each other’s reasoning and systematically expose unsafe failure modes through a completely automated red-teaming process.

To facilitate this, we introduce distinct long-term memory modules that retain safety-critical insights gained from debate interactions. These insights are then utilized during future inference stages, enabling the continuous refinement of model behavior. Our empirical tests on safety benchmarks involving a wide range of models show that RedDebate significantly decreases the production of unsafe outputs. Although debate mechanisms alone help LLMs improve their conduct, incorporating memory leads to additional reductions in errors. As far as we are aware, RedDebate represents the first fully automated framework to integrate multi-agent debate with red-teaming, progressively boosting LLM safety without requiring human involvement.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC