Academic

Experience as a Compass: Multi-agent RAG with Evolving Orchestration and Agent Prompts

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. We identify two key limitations: the lack of continuously adaptive orchestration mechanisms and the absence of behavior-level learning for individual agents. To this end, we propose HERA, a hierarchical framework that jointly evolves multi-agent orchestration and role-specific agent prompts. At the global level, HERA optimizes query-specific agent topologies through reward-guided sampling and experience accumulation. At the local level, Role-Aware Prompt Evolution refines agent behaviors via credit assignment and dual-axes adaptation along operational and behavioral principles, enabling targeted, role-condi

S
Sha Li, Naren Ramakrishnan · · 1 min read · 3 views
#cs.AI

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. We identify two key limitations: the lack of continuously adaptive orchestration mechanisms and the absence of behavior-level learning for individual agents. To this end, we propose HERA, a hierarchical framework that jointly evolves multi-agent orchestration and role-specific agent prompts. At the global level, HERA optimizes query-specific agent topologies through reward-guided sampling and experience accumulation. At the local level, Role-Aware Prompt Evolution refines agent behaviors via credit assignment and dual-axes adaptation along operational and behavioral principles, enabling targeted, role-conditioned improvements. On six knowledge-intensive benchmarks, HERA achieves an average improvement of 38.69\% over recent baselines while maintaining robust generalization and token efficiency. Topological analyses reveal emergent self-organization, where sparse exploration yields compact, high-utility multi-agent networks, demonstrating both efficient coordination and robust reasoning.

Executive Summary

The article proposes a novel hierarchical framework, HERA, to address the limitations of Multi-agent Retrieval-Augmented Generation (RAG) in handling diverse and complex tasks. HERA evolves multi-agent orchestration and role-specific agent prompts through reward-guided sampling and experience accumulation, ensuring continuous adaptation and improvement. The framework achieves an average improvement of 38.69% over recent baselines on six knowledge-intensive benchmarks, demonstrating robust generalization, token efficiency, and emergent self-organization. The study highlights the importance of adaptive orchestration mechanisms and behavior-level learning for individual agents in multi-agent RAG. The proposed framework has significant implications for AI applications requiring complex reasoning and multi-hop tasks.

Key Points

  • ▸ HERA framework jointly evolves multi-agent orchestration and role-specific agent prompts
  • ▸ Reward-guided sampling and experience accumulation ensure continuous adaptation and improvement
  • ▸ Achieves an average improvement of 38.69% over recent baselines on six knowledge-intensive benchmarks

Merits

Strength in Adaptive Orchestration

HERA's ability to adaptively evolve multi-agent orchestration and role-specific agent prompts enables the framework to handle diverse and complex tasks effectively.

Improved Performance

The proposed framework achieves significant improvements over recent baselines on six knowledge-intensive benchmarks, demonstrating its potential in real-world applications.

Demerits

Complexity in Implementation

The HERA framework's hierarchical structure and adaptive mechanisms may introduce additional complexity in implementation, which could be a challenge for practitioners.

Limited Generalizability

The study's focus on knowledge-intensive benchmarks may limit the generalizability of the results to other domains or tasks.

Expert Commentary

The proposed HERA framework is a significant contribution to the field of multi-agent systems, addressing the limitations of existing approaches through adaptive orchestration and behavior-level learning. The study's results demonstrate the potential of HERA in handling diverse and complex tasks, with significant implications for AI applications. However, the framework's complexity in implementation and limited generalizability are notable limitations that require further attention. As the field continues to evolve, it is essential to explore the HERA framework's potential in various domains and tasks, as well as its explainability and transparency.

Recommendations

  • ✓ Future research should focus on developing more efficient and scalable implementation of the HERA framework.
  • ✓ The study's findings should be replicated and extended to other domains and tasks to ensure the framework's generalizability and robustness.

Sources

Original: arXiv - cs.AI

Related Articles