Skip to content
InfoResearchPreprintLLM-specific

Attacker Diversity Reduction in Adaptive LLM Red Teaming

Published
Record updated
View JSON

Summary

Co-training an attacker and defender with GRPO reduced attacker diversity in automated LLM red teaming, reproduced here with a Qwen2.5-14B attacker and a Phi-4-mini defender trained by DPO. Across three runs, Attack Diversity Retention fell to 0.63, 0.61 and 0.63, below the 0.75 target, at round 12, while attack success kept improving. The 26 AdvBench behaviors and the nominal intervals, given training ended after 12 rounds on development checkpoints, bound the result.