A Survey of Large Audio Language Models: Generalization, Trustworthiness and Outlook

A survey systematically evaluates the trustworthiness of large audio language models (LALMs), noting that their capabilities are growing markedly faster than assurance frameworks. The survey summarizes key vulnerabilities such as cross-modal jailbreaks, latent audio backdoors and biometric privacy leakage, and evaluates existing models along six dimensions: hallucination, robustness, safety, privacy, fairness and certification. The authors argue that unified end-to-end frameworks and continuous acoustic signals expand the attack surface, and propose a roadmap of defense in depth, causal auditory world modeling and intrinsic representation engineering.