AI Summary: MiRA introduces a 'divide and conquer' strategy for multimodal tasks by using three distinct agents for vision, text, and final judgment. By forcing the visual and textual agents to reason independently before a 'judge' combines their findings, the framework prevents the 'hallucination leakage' often seen in unified multimodal models.
We propose Mixture-of-Reasoning Agents (MiRA), a zero-shot multimodal framework that decomposes reasoning across three specialized agents: Visual Analyzing, Text Comprehending, and Judge. By consolidating multimodal evidence through independent reasoning paths, MiRA achieves 96.0% accuracy on ScienceQA, surpassing all few-shot and zero-shot baselines. This modular approach provides highly accurate, interpretable results without the need for task-specific adaptation, establishing a new paradigm for generalizable AI in multimodal answering.
Share your opinion to help other learners triage faster.
Write a reviewInvite someone by email to share an invited review for MiRA: A Zero-Shot Mixture-of-Reasoning Agents Framework.