DeALOG: Decentralized Multi-Agents Log-Mediated Reasoning Framework

Researchers introduce DeALOG, a decentralized multi-agent framework that uses specialized AI agents coordinating through a shared natural-language log to answer complex questions spanning text, tables, and images. The system demonstrates competitive performance on multiple benchmarks while improving robustness through collaborative verification without central control.

DeALOG represents a meaningful shift in how AI systems approach multimodal reasoning tasks that demand integration across heterogeneous data sources. Rather than relying on monolithic models or centralized orchestration, the framework decomposes complex question-answering into specialized agents—each optimized for tables, context, visuals, summarization, or verification—that communicate asynchronously through persistent natural-language logs. This architectural choice mirrors principles from distributed systems and decentralized networks, trading some computational efficiency for improved interpretability and error resilience.

The significance lies in demonstrating that persistent, human-readable communication logs can function as effective coordination mechanisms in multi-agent AI systems. Traditional approaches often employ hidden state representations or centralized coordinators that obscure reasoning processes. By making agent interactions explicit through natural language, DeALOG enables collaborative error detection and verification without hierarchical control, addressing a critical pain point in complex AI systems where failures cascade unpredictably.

For the AI development community, this work validates modular, specialized agent architectures as competitive with larger monolithic models on challenging benchmarks including FinQA, TAT-QA, and MultiModalQA. This has practical implications for developers building production systems where interpretability and maintainability matter alongside accuracy. The log-based communication approach scales more naturally than centralized designs as agent count increases.

Looking forward, the framework's principles could extend beyond question-answering into other complex reasoning domains. Key questions include whether natural-language logs introduce latency bottlenecks at scale and how the approach performs as task complexity increases beyond current benchmarks. The research suggests that decentralized coordination patterns, borrowed from distributed systems, merit deeper exploration in AI architecture design.

• →DeALOG uses five specialized agents communicating through shared natural-language logs to handle multimodal question-answering without central control.
• →The log-based coordination mechanism enables collaborative error detection and improves system robustness compared to centralized alternatives.
• →Competitive performance across six major benchmarks validates the effectiveness of modular, specialized agent architectures for complex reasoning.
• →Natural-language communication between agents provides explicit, interpretable reasoning traces—a critical advantage for production deployments.
• →The decentralized framework demonstrates scalability through modular components, suggesting applicability beyond question-answering tasks.