BUZZY: Contrastive Scoring to Mitigate Text-Induced Bias in Multimodal Multiple-Choice QA

The research team proposed the BUZZY method, aimed at addressing the issue of text-induced bias in multimodal visual-language models in multiple-choice questions. This method relies on the assumption that candidate options serve as text priors and employs a training-free decoding strategy: by subtracting the distribution generated solely from text from the multimodal prediction results, it corrects the biases caused by insufficient visual signals. Experiments show that BUZZY achieved the most advanced level of average accuracy in five benchmark tests, while reducing推理 latency by more than 28%.