GoAgent: Group-of-Agents Communication Topology Generation for LLM-based Multi-Agent Systems
New method treats agent groups as atomic units, achieving state-of-the-art results while reducing communication overhead.
A research team led by Hongjiang Chen has introduced GoAgent, a novel framework that fundamentally rethinks how communication topologies are generated for LLM-based multi-agent systems. Unlike existing node-centric approaches where group structures emerge implicitly, GoAgent explicitly treats collaborative groups as the atomic units of system construction. The method first uses an LLM to enumerate task-relevant candidate groups, then autoregressively selects and connects these groups to build the final communication graph, jointly capturing both intra-group cohesion and inter-group coordination.
To address the critical issues of communication redundancy and noise propagation, the researchers incorporated a conditional information bottleneck objective. This CIB mechanism compresses inter-group communication, preserving task-relevant signals while filtering out redundant historical noise. Extensive testing across six benchmarks demonstrated GoAgent's superior performance, achieving state-of-the-art results with 93.84% average accuracy while simultaneously reducing token consumption by approximately 17%. This dual achievement of higher accuracy with lower computational cost represents a significant advancement in making multi-agent systems more efficient and effective for complex problem-solving tasks.
- Explicitly models agent groups as atomic units rather than individual nodes, improving coordination
- Achieves 93.84% average accuracy across six benchmarks while reducing token use by ~17%
- Uses conditional information bottleneck to filter redundant communication and historical noise
Why It Matters
Enables more efficient, scalable multi-agent systems for complex tasks like software development, research, and enterprise workflows.