DeAR: Decentralized Agentic Reasoning Boosts Multimodal AI Performance

Traditional agentic reasoning systems often rely on a centralized control mechanism, which can lead to bottlenecks and rigid role assignments, especially when dealing with complex, multimodal queries. This research introduces DeAR, a novel framework for Decentralized Agentic Reasoning, designed to overcome these limitations by fostering autonomous peer-to-peer collaboration among AI agents. DeAR operates on three core principles: decentralized capability grounding, which allows agents to specialize based on the specific query; thought map navigation, enabling targeted interactions between peers; and topology updates, which facilitate adaptive error correction. This decentralized approach allows agents to dynamically form collaborative networks tailored to the task at hand. Evaluations conducted across nine diverse multimodal reasoning and text-based QA benchmarks demonstrated that DeAR consistently surpassed recent baseline methods. The findings validate that a decentralized and adaptive collaborative architecture significantly enhances accuracy in knowledge-intensive reasoning tasks, offering a promising direction for more robust and scalable AI agent systems.