Reward-Guided Autoregressive Graph Generation for Efficient Multi-Agent Communication Topology Design
RGA-Designer cuts multi-agent communication tokens by 20.5% while matching ARG-Designer’s task accuracy.
The paper targets LLM-based multi-agent systems, where richer agent communication can drive up token use. It extends ARG-Designer by adding a reward model that scores both task correctness and structural compactness. That reward then guides fine-tuning of the graph generator toward sparser communication topologies. The full version of the extended abstract was accepted as an ICONIP 2026 poster. ArXiv · AI/CL/LG's note
The paper targets LLM-based multi-agent systems, where richer agent communication can drive up token use. It extends ARG-Designer by adding a reward model that scores both task correctness and structural compactness. That reward then guides fine-tuning of the graph generator toward sparser communication topologies. The full version of the extended abstract was accepted as an ICONIP 2026 poster. ArXiv · AI/CL/LG's note
score 4