Rebellion: Noise-Robust Reasoning Training for Audio Reasoning Models

The researchers propose Rebellion, a noise-robust reasoning training method for audio reasoning models (ARMs) to defend against jailbreak attacks on audio language models. The authors state that no prior work has studied the safety of ARMs against jailbreak attacks; the paper shows that standard reasoning training with safety reasoning data can defend against ordinary audio jailbreaks but cannot defend against the simple yet effective advanced jailbreak proposed by the authors, because a significant representation drift between the two types of jailbreaks forces the target model to output harmful responses. Based on this, Rebellion trains models to remain robust to worst-case representation drift. All results are based on Qwen2-Audio: Rebellion defends against advanced audio jailbreaks without harming performance on benign tasks, and significantly improves the accuracy-safety trade-off relative to standard reasoning training.