Graph-of-Agents Framework Introduced for Multi-Agent LLM Collaboration
A novel framework known as Graph-of-Agents (GoA) has been introduced to overcome challenges in multi-agent LLM collaboration systems. This method initiates with node sampling, focusing on the most pertinent agents through model cards that outline each model's specialization and domain. Connections, or edges, are formed between these chosen agents by assessing their responses to establish a relevance ranking. Subsequently, directed message passing occurs, allowing highly relevant agents to communicate with those deemed less relevant, thereby improving responses. GoA seeks to advance coordination techniques like Mixture-of-Agents (MoA), which often face difficulties in agent selection, intra-agent communication, and response integration. The increasing variety of LLMs and benchmarks intensifies the need for effective model orchestration. This research was published on arXiv under identifier 2604.17148v1.
Key facts
- Graph-of-Agents (GoA) is a new graph-based framework for multi-agent LLM communication
- The framework begins with node sampling to select relevant agents using model cards
- Model cards summarize each model's domain, task specialization, and characteristics
- Edges are constructed between selected agents by evaluating responses against one another
- Directed message passing occurs from highly relevant agents to less relevant ones
- The approach addresses limitations in existing frameworks like Mixture-of-Agents (MoA)
- MoA frameworks often struggle with agent selection, intra-agent communication, and response integration
- The growing zoo of LLMs and benchmarks creates need for better multi-model orchestration
Entities
Institutions
- arXiv