What they did

The authors build on ARG-Designer, which represents multi-agent communication topology design as autoregressive graph generation. Because its original training objective does not explicitly reward efficiency, it may generate unnecessarily dense communication graphs.

Their RGA-Designer trains a reward model to score both task correctness and structural compactness, then uses that feedback to fine-tune the pretrained graph generator in a process inspired by reinforcement learning from human feedback.

Key findings

• RGA-Designer reduced token consumption by an average of 20.5% relative to ARG-Designer.
• The generated topologies retained task accuracy at the level reported for ARG-Designer.
• Explicitly rewarding compactness produced more communication-efficient graphs than relying on the autoregressive training objective alone.

Why it matters

Communication between LLM agents can dominate the cost of multi-agent reasoning. This work suggests that topology selection can be optimized directly for both performance and resource use, rather than treating communication overhead as a fixed consequence of adding more agents.

Caveats

The abstract does not specify the evaluation tasks, model families, absolute token counts, statistical variation, or computational cost of training the reward model. It therefore remains unclear how consistently the savings transfer across domains and whether optimizing structural compactness introduces failure modes on harder or unfamiliar tasks.