RMIT-ADM+S at the MMU-RAG NeurIPS 2025 Competition
arXiv:2602.20735v1 Announce Type: cross Abstract: This paper presents the award-winning RMIT-ADM+S system for the Text-to-Text track of the NeurIPS~2025 MMU-RAG Competition. We introduce Routing-to-RAG (R2RAG), a research-focused retrieval-augmented generation (RAG) architecture composed of lightweight components that dynamically adapt the retrieval strategy based on inferred query complexity and evidence sufficiency. The system uses smaller LLMs, enabling operation on a single consumer-grade GPU while supporting complex research tasks. It builds on the G-RAG system, winner of the ACM~SIGIR~2025 LiveRAG Challenge, and extends it with modules informed by qualitative review of outputs. R2RAG won the Best Dynamic Evaluation award in the Open Source category, demonstrating high effectiveness with careful design and efficient use of resources.
arXiv:2602.20735v1 Announce Type: cross Abstract: This paper presents the award-winning RMIT-ADM+S system for the Text-to-Text track of the NeurIPS~2025 MMU-RAG Competition. We introduce Routing-to-RAG (R2RAG), a research-focused retrieval-augmented generation (RAG) architecture composed of lightweight components that dynamically adapt the retrieval strategy based on inferred query complexity and evidence sufficiency. The system uses smaller LLMs, enabling operation on a single consumer-grade GPU while supporting complex research tasks. It builds on the G-RAG system, winner of the ACM~SIGIR~2025 LiveRAG Challenge, and extends it with modules informed by qualitative review of outputs. R2RAG won the Best Dynamic Evaluation award in the Open Source category, demonstrating high effectiveness with careful design and efficient use of resources.
Executive Summary
This article presents the RMIT-ADM+S system, a research-focused retrieval-augmented generation (RAG) architecture, which won the Best Dynamic Evaluation award in the Open Source category of the NeurIPS 2025 MMU-RAG Competition. The system, called Routing-to-RAG (R2RAG), dynamically adapts the retrieval strategy based on inferred query complexity and evidence sufficiency, using smaller language models (LLMs) to operate on a single consumer-grade GPU. This approach enables complex research tasks while efficiently using resources. R2RAG builds on the G-RAG system, winner of the ACM SIGIR 2025 LiveRAG Challenge, and extends it with modules informed by qualitative review of outputs. The system's effectiveness demonstrates the value of careful design and resource management in RAG architectures.
Key Points
- ▸ R2RAG is a research-focused RAG architecture that dynamically adapts the retrieval strategy
- ▸ The system uses smaller LLMs to operate on a single consumer-grade GPU
- ▸ R2RAG won the Best Dynamic Evaluation award in the Open Source category of the NeurIPS 2025 MMU-RAG Competition
Merits
Strength in Resource Efficiency
R2RAG's use of smaller LLMs and efficient resource management enables complex research tasks while minimizing computational costs
Improved Adaptability
The system's ability to dynamically adapt the retrieval strategy based on inferred query complexity and evidence sufficiency enhances its effectiveness in various research tasks
Demerits
Potential Dependence on Specific Tasks
The system's performance may be task-dependent, and further research is needed to evaluate its effectiveness across various research domains
Limited Generalizability
The system's design and training data may not generalize well to other research tasks or domains, limiting its broader applicability
Expert Commentary
The article presents a significant contribution to the field of RAG architectures, highlighting the importance of careful design and resource management in achieving high effectiveness. The development of R2RAG demonstrates the potential of this approach in various research domains, and its focus on resource efficiency is particularly noteworthy. However, further research is needed to evaluate the system's performance across different tasks and domains, as well as its generalizability to broader applications. Overall, the article provides valuable insights into the design and development of efficient RAG architectures, with significant implications for researchers, developers, and policymakers.
Recommendations
- ✓ Future research should aim to evaluate R2RAG's performance across various research tasks and domains to assess its generalizability and effectiveness
- ✓ Developers and researchers should explore the application of R2RAG in other research domains to further demonstrate its potential and identify areas for improvement
