Reproduce “DARE to Mitigate Hallucination: Dual-path Auto-Regressive-aware Editing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jae-Ho Lee, Jeong-Eun Lee, Gyeong-Moon Park
DARE edits vision-language model weights with teacher-forced text contrasts, autoregressive hidden-state changes before hallucinated tokens, and paired-image visual differences. It applies text directions to feed-forward layers and visual directions to attention layers without decoding-time overhead. On LLaVA-1.5, it lowers caption-level CHAIR_S to 14.93 ± 1.68 and object-level CHAIR_I to 6.07 ± 0.82 versus Nullu, but does not eliminate hallucinations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.