This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory
Created by
Haebom
Author
Jun Liu, Zhenglun Kong, Changdi Yang, Fan Yang, Tianqi Li, Peiyan Dong, Joannah Nanjekye, Hao Tang, Geng Yuan, Wei Niu, Wenbin Zhang, Pu Zhao, Xue Lin, Dong Huang, Yanzhi Wang
Outline
This paper proposes RCR-Router, a novel routing framework for efficient collaboration in multi-agent large-scale language model (LLM) systems. To overcome the limitations of existing static or full-context routing strategies, RCR-Router adopts a modular, role-aware approach that dynamically selects semantically relevant memory subsets based on each agent's role and task stage. This is achieved while adhering to a strict token budget. A lightweight scoring policy guides memory selection, and agent outputs are iteratively merged into a shared memory store, enabling incremental context refinement. Furthermore, we present the Answer Quality Score metric, which captures the explanations generated by the LLM, to better evaluate model performance. Experimental results on three multi-hop QA benchmarks—HotPotQA, MuSiQue, and 2WikiMultihop—show that RCR-Router maintains or improves answer quality while reducing token usage by up to 30%.
Takeaways, Limitations
•
Takeaways:
◦
We present RCR-Router, a novel routing framework for efficient collaboration in multi-agent LLM systems.
◦
Role-aware and dynamic memory routing reduces token usage and maintains or improves response quality.
◦
Improving the evaluation method by proposing an Answer Quality Score indicator that takes into account the LLM creation description.
◦
Emphasize the importance of structured memory routing and output-aware evaluation.
•
Limitations:
◦
The performance of the proposed RCR-Router is based on experimental results limited to a specific benchmark. Additional experiments on a variety of tasks and datasets are required.
◦
Further research is needed to determine the generalizability and objectivity of the Answer Quality Score indicator.
◦
Lack of detailed explanation of the design and optimization of the lightweight scoring policy.
