Orchestra-o1 Enables Omnimodal Agent Orchestration for Complex Multi-Agent Systems.

The field of AI agents is moving towards multi-agent systems, where effective orchestration is key for task decomposition and collaboration. Current orchestration frameworks often struggle with tasks involving multiple modalities, such as text, images, audio, and video, limiting their application in complex, real-world scenarios. Orchestra-o1 addresses this by providing an omnimodal agent orchestration framework designed for efficient collaboration across diverse input types. It incorporates a unified mechanism for modality-aware task decomposition, allowing sub-agents to specialize dynamically and execute sub-tasks in parallel. This scalable design enables agent systems to process heterogeneous information sources effectively, achieving a 10.3% accuracy improvement over the next best approach on the OmniGAIA benchmark. The research also introduces DA-GRPO, an agentic reinforcement learning method used to train Orchestra-o1-8B, which sets new state-of-the-art performance for open-source omnimodal agents.