LLM-based multi-agent systems (MAS) achieve strong reasoning performance but consume many tokens. The prior ARG-Designer generates communication topologies autoregressively but does not encourage sparsity in its training objective. This paper introduces RGA-Designer, which adds a reward model that jointly scores task correctness and structural compactness. The graph generator is fine-tuned using from that reward model. Across six benchmarks, the method reduces token consumption by an average of 20.5% while maintaining the same task accuracy as ARG-Designer.
Reward-guided graph generation cuts multi-agent communication costs without hurting accuracy
RGA-Designer learns to favor compact communication topologies while maintaining the task performance of an earlier autoregressive approach.
Academic
Poomphob Suwannapichat · Boonyarit Changaival · Caesar Wu · Pascal Bouvry
University of Luxembourg · King Mongkut’s University of Technology Thonburi
Research Digest··2 min read
Thread:Agent Harness Optimization
In one line
RGA-Designer reduces token consumption by 20.5% while preserving task accuracy in multi-agent communication topologies.
Before you read this
Two ideas this paper builds onmulti-agent architectureA system where multiple AI agents collaborate, each handling specialized roles, communicating and coordinating to complete complex tasks.99 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.239 papers →
Read it as
Suwannapichat et al. add reward-guided fine-tuning to a model that generates communication graphs for LLM-based multi-agent systems.
The authors build on ARG-Designer, which represents multi-agent communication topology design as autoregressive graph generation.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code