Experience as a Compass: Multi-agent RAG with Evolving Orchestration and Agent Prompts
Summary: arXiv:2604.00901v1 Announce Type: new
Abstract: Multi-agent Retrieval-Augmented Generation (RAG), wherein each agent takes on a specific role, supports hard queries that require multiple steps and sources, or complex reasoning. Existing approaches, however, rely on static agent behaviors and fixed orchestration strategies, leading to brittle performance on diverse, multi-hop tasks.
Introduction
Recent advancements in artificial intelligence have highlighted the utility of multi-agent systems for tackling complex queries. However, traditional methods often fall short due to their reliance on static behaviors and fixed strategies. This article introduces HERA, a novel hierarchical framework designed to enhance the orchestration of multi-agent systems and improve agent-specific prompts.
Challenges in Current Multi-agent RAG Systems
Existing multi-agent systems face several limitations that hinder their performance in dynamic environments:
- Static Agent Behaviors: Agents often operate under predefined roles without the ability to adapt to changing contexts.
- Fixed Orchestration Strategies: Current systems typically employ rigid coordination strategies that do not account for the complexities of multi-hop tasks.
The HERA Framework
HERA addresses these limitations by introducing two innovative mechanisms:
- Adaptive Orchestration Mechanism: At the global level, HERA optimizes agent topologies based on specific queries through reward-guided sampling and experience accumulation. This allows for a more flexible and responsive system capable of handling diverse tasks.
- Role-Aware Prompt Evolution: HERA refines agent behaviors by utilizing credit assignment and adapting along dual axes—operational and behavioral principles. This targeted approach promotes role-specific improvements, enhancing overall system performance.
Performance and Results
HERA has been evaluated on six knowledge-intensive benchmarks, demonstrating significant advancements over existing baselines:
- Average improvement of 38.69% in performance metrics.
- Maintains robust generalization across varied tasks.
- Demonstrates token efficiency, allowing for better resource utilization.
Emergent Self-Organization
One of the key findings from the HERA framework is the emergence of self-organizing structures within multi-agent networks. Through sparse exploration, the system has been able to develop compact, high-utility networks that facilitate:
- Efficient Coordination: Agents work more cohesively, leading to improved task execution.
- Robust Reasoning: Enhanced reasoning capabilities allow for better handling of complex queries.
Conclusion
The HERA framework represents a significant advancement in the field of multi-agent systems, addressing critical limitations of previous approaches. By fostering adaptive orchestration and role-specific behavior evolution, HERA not only enhances multi-agent collaboration but also sets the stage for future research in dynamic, knowledge-intensive environments.