This pilot study introduces a deductive debating approach titled Blueprint Debate on Graphs (BDoG) aimed at remedying issues in multimodal reasoning. Traditional inductive debates in AI often over-summarize information or get distracted by irrelevant details from images. BDoG confines debates to a blueprint graph, preventing opinion trivialization through word-level summarization and storing evidence in branches within the graph.

• Blueprint Debate on Graphs (BDoG) is introduced to multitask multimodal reasoning.
• BDoG prevents trivialization of opinions and distractions from irrelevant concepts.
• The blueprint graph confines debates and stores evidence effectively.
• Employs a top-down debating process as opposed to traditional bottom-up methods.
• Validated through extensive experiments, BDoG outperforms existing methods on benchmarks like Science QA and MMBench.

Understanding and reasoning about multimodal data is crucial in AI. The BDoG approach presented in this paper is important because it addresses fundamental challenges that AI faces when dealing with complex multimodal information. By structuring debates within a blueprint graph, BDoG could influence the development of AI systems that are better at focusing on relevant information and reasoning more effectively. This research has the potential to be a stepping stone towards AI that can understand and engage with the world in a way that’s more similar to human cognitive processing.