一作ACL 2026Poster· Natural Language Processing
Knowledge Injection Exists in M o E ? Exploring Expert-Aware Contrast Decoding in M o E for Mitigating LLM s’ Hallucinations
Xinyue Fang · Zhiliang Tian · Zhen Huang · Ziyi Pan · Zhihua Wen · Xi Wang · Quntian Fang · Dongsheng Li
摘要
Existing LLM hallucination mitigation methods, including prompt engineering and model optimization, either hardly alter models’ internal knowledge or have poor cross-domain generalization. Contrastive decoding mitigates hallucinations by using layer-wise differences in LLMs. However, prior studie...