BUZZY method reduces bias and speeds up multimodal QA

Researchers have developed BUZZY, a novel decoding method designed to mitigate text-induced bias in multimodal multiple-choice question answering (MCQA). This training-free approach corrects multimodal predictions by subtracting the text-only distribution, hypothesizing that models rely on visual evidence when their multimodal distribution significantly diverges from text-only predictions. Experiments on five benchmarks with five vision-language models demonstrated that BUZZY achieves superior accuracy while reducing inference latency by over 28% compared to existing contrastive decoding methods. AI

IMPACT This method could improve the accuracy and efficiency of vision-language models in question-answering tasks.

RANK_REASON The cluster describes a research paper detailing a new method for multimodal question answering. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• BUZZY
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• ScienceCast
• Taeyun Roh

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →