Learning Who Talks to Whom: Role-Aware Communication Topology Scheduling (arxiv.org)

A new framework called ACTS has been proposed to enhance communication within multi-agent systems powered by large language models (LLMs). Traditional approaches often face challenges in dynamically adapting communication schedules, either by relying on static graphs or failing to leverage individual agent roles. ACTS addresses these shortcomings through a role-aware adaptive communication topology scheduling mechanism. It features a dual-attention system that employs a Role-conditioned Relational Attention module and a Topological Attention Mixer, which work together to optimize partner selection for communication and ensure effective credit assignment. The significance of ACTS lies in its ability to improve collaborative reasoning efficiency while reducing overhead costs associated with token usage. Comprehensive evaluations show that ACTS consistently outperforms state-of-the-art methods across nine diverse benchmarks, achieving an average accuracy increase of up to 6.19 percentage points and a reduction in token consumption by over 15%. This advancement not only enhances the performance of LLMs in complex problem-solving scenarios but also sheds light on the importance of role-based collaboration in AI systems, marking a significant step forward in multi-agent communication strategies within the AI/ML community.