Work

Publications

1. arXiv 2026 Sep 2026

   AEGIS: Audio Endogenous Guarding via Internal Signals Against Large Audio-Language Model Jailbreaks

   Y. L. Liao , T. C. Chiu , Z. Y. Chen , C. L. Tsai , S. Y. Lo

   Under review at ICASSP 2027

   A defense for large audio-language models that detects diverse jailbreak attacks from the model's own internal representations, and analyzes how safety behavior is encoded inside the model.

   [arXiv]
2. LREC 2026 Mar 2026

   Do Multimodal LLMs Understand Order? Measuring the Fragility of Multimodal Reasoning under Input Order Perturbations

   S.L. Wei , Y. L. Liao , H. H. Huang , H. H. Chen

   Evaluates how large vision-language models respond to reordered multiple-choice options and modality inputs, revealing positional bias across MMMU, CVQA, and MMBench and testing strategies for more stable multimodal reasoning.

   [Paper]
3. Findings of EACL 2026 Feb 2026

   Bias in the Ear of the Listener: Assessing Sensitivity in Audio Language Models Across Linguistic, Demographic, and Positional Variations

   S.L. Wei* , Y. L. Liao* , Y. H. Chang , H. H. Huang , H. H. Chen

   *Equal contribution

   Investigates sensitivity biases in audio language models by evaluating how model responses shift across linguistic, demographic, and positional perturbations of the same audio content.

   [Paper] [arXiv]

Projects