DeAR: Decentralized Agentic Reasoning via Capability Grounding and Collaborative Thought Navigation

Details

Current multi-agent reasoning systems depend on a central coordinator to route queries and assign roles, creating bottlenecks and rigid structure that degrades on complex multimodal tasks. DeAR removes the central controller entirely, letting agents specialize dynamically based on query requirements and communicate directly with relevant peers. The framework rests on three mechanisms: capability grounding assigns agents roles based on the incoming query rather than static assignment; thought map navigation determines which agents interact; and topology update allows the network to correct errors adaptively without central intervention. Evaluated across 9 benchmarks spanning multimodal reasoning and text-based QA, DeAR consistently outperforms recent baselines. Source code will be released upon paper acceptance — no public implementation is available yet. The paper is a preprint submitted August 18, 2026. Relevant to practitioners building production multi-agent pipelines who face scaling issues with centralized orchestration. The peer-to-peer approach reduces single-point-of-failure risks in complex agentic systems.