In this work, we analyze that both positive and negative prompts can amplify hallucinations and propose a training-free method called instruction contrastive decoding, which works on the fusion module.
Hi @Ethanscuter , your ICD paper is already in the paper list. Very interesting work!
Congrats for your paper acceptance! I will update the conference information.
Hi @JosephPai,
Thanks for your awesome repository, which serves as a great resource in this community.
Could you please consider adding our recent paper on ACL 2024 about multimodal hallucination?
ICD: Mitigating Hallucinations in Large Vision-Language Models with Instruction Contrastive Decoding ACL 2024 Mar. 23, 2024
In this work, we analyze that both positive and negative prompts can amplify hallucinations and propose a training-free method called instruction contrastive decoding, which works on the fusion module.
Thank you!