SMCS: Scalable Multi-LLM Collaboration System Outperforms GPT-4.1
A new system called SMCS (Scalable Multi-LLM Collaboration System) has been proposed to address scalability challenges in integrating new LLMs and tasks. It consists of a Retrieval-based Prior Selection (RPS) module for dynamic LLM selection and an Exploration-Exploitation-Driven Posterior Enhancement (EPE) module for response diversity and quality scoring. On eight benchmarks, SMCS integrating fifteen open-source LLMs outperformed GPT-4.1 by 5.36% and GPT-o3-mini by 5.28%, even exceeding the average of best results on different datasets.
Key facts
- SMCS is a Scalable Multi-LLM Collaboration System
- It has two core components: RPS and EPE
- RPS dynamically selects suitable LLMs per input
- EPE fosters response diversity and selects high-quality outputs
- Experiments on eight mainstream benchmarks
- Integrated fifteen open-source LLMs
- Outperformed GPT-4.1 by 5.36%
- Outperformed GPT-o3-mini by 5.28%
Entities
—