時間：115年9月17日(星期四) 

地點：

主持人：蔡銘峰老師

演講者:  台大資工博士候選人 魏聖倫先生

演講題目: Input Bias in Large Language Models: Investigating Robustness Across Text, Vision, and Speech

演講摘要: Large language models (LLMs) achieve remarkable performance across diverse tasks, yet their predictions remain surprisingly fragile to superficial perturbations of the input. This talk presents a unified line of research on input bias in LLMs across three modalities. It first revisits selection bias in text-based LLMs (ACL 2024), introducing the Fluctuation Rate metric and mitigation strategies for option-order and token sensitivity. The discussion then extends to large vision–language models (LREC 2026), revealing how choice order, modality order, and visual complexity jointly modulate robustness. Finally, the talk turns to speech with BiasInEar (EACL 2026), the first systematic study of bias in multilingual speech-integrated LLMs across language, accent, gender, and option order. Together, these studies argue that input bias is not a modality-specific artifact but a fundamental challenge for building robust and trustworthy multimodal language models.

3. 個人簡介
Sheng-Lun (Kevin) Wei is a CS PhD candidate at National Taiwan University, working on LLM biases and fairness, multimodal AI, and LLM evaluation. With 6+ years of industry experience at ShopBack, Junyi Academy, and KKStream, he bridges the gap between research and real-world applications. He also serves as an adjunct instructor at NTU's Center of General Education and Department of Economics, and is the founder of ccClub (社團法人攜曦程式推廣學會), a non-profit organization dedicated to programming education that has served over 4,000 learners since 2016.