Learning to Attack and Defend: Adaptive Red Teaming of Language Models via GRPO

Name: Learning to Attack and Defend: Adaptive Red Teaming of Language Models via GRPO
Uploaded: 2026-10-11T07:57:02.989207+00:00
Duration: 333 s
Description: AdvGRPO introduces a co-training method for joint attacker-defender optimization using Group Relative Policy Optimization (GRPO). It uses dense, multi-channel rewards and decoupled advantage normalization to make GRPO work effectively. This results in highly transferable attacks and defenders that significantly outperform baselines on safety benchmarks.

AdvGRPO introduces a co-training method for joint attacker-defender optimization using Group Relative Policy Optimization (GRPO). It uses dense, multi-channel rewards and decoupled advantage normalization to make GRPO work effectively. This results in highly transferable attacks and defenders that significantly outperform baselines on safety benchmarks.

Open the full episode