BUZZY: New Method Reduces Text-Induced Bias in Multimodal MCQA

A recent study presents BUZZY, a method for decoding that does not require training and aims to reduce bias stemming from text in multimodal multiple-choice question answering (MCQA). The research, titled "BUZZY: Contrastive Scoring to Mitigate Text-Induced Bias in Multimodal Multiple-Choice QA," is accessible on arXiv under the identifier 2603.28026, replacing a prior version. The authors contend that in MCQA, candidate answers serve as textual priors, leading vision-language models (VLMs) to prefer linguistically plausible choices over visual content. They propose that a model utilizes visual evidence only when its multimodal distribution significantly differs from the text-only distribution. BUZZY adjusts multimodal predictions by subtracting the text-only distribution. Tests involving five VLMs across five multimodal MCQA benchmarks validate the method's effectiveness. The paper can be found at https://arxiv.org/abs/2603.28026.