Small Drafts, Big Verdict: Information-Intensive Visual Reasoning via Speculation

Researchers developed Speculative Verdict (SV), a training-free framework that improves large Vision-Language Models' ability to reason over information-dense images by combining multiple small draft models with a larger verdict model. The approach achieves better accuracy on visual question answering benchmarks while reducing computational costs compared to large proprietary models.

• →SV framework uses small VLMs as draft experts to generate diverse reasoning paths, then synthesizes them with a strong verdict model.
• →The approach addresses challenges in localizing critical information in dense visual layouts and multi-hop reasoning across dispersed evidence.
• →Consensus expert selection mechanism filters high-agreement reasoning paths to improve efficiency and accuracy.
• →Testing showed consistent improvements on challenging benchmarks including InfographicVQA, ChartMuseum, and HR-Bench 4K.
• →The training-free approach offers cost-efficiency advantages over large proprietary models while maintaining error correction capabilities.