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

This survey reviews large audio language models, with a focus on generalization, trustworthiness risks, and defenses for safer audio intelligence.

It examines LALM architectures, alignment methods, and key failure modes such as hallucination, robustness gaps, and safety issues.

The paper introduces a risk taxonomy covering cross-modal jailbreaking, latent acoustic backdoors, biometric privacy leakage, and other threats.

It argues that audio AI has advanced faster than its security and reliability safeguards, making trust, privacy, and safety central to deployment.