EMAS: Evidence-Guided Multi-Agent System Evolution
EMAS (Evolving Multi-Agent System) introduces a novel approach for designing automated multi-agent systems by utilizing runtime experiences to adjust system topology and prompts without the need to modify LLM parameters. In contrast to conventional methods that optimize in a preliminary design phase and remain static thereafter, EMAS gathers insights from samples to create reusable updates aimed at enhancing accuracy or minimizing token expenses. This technique transforms execution traces into structured diagnoses that outline necessary revisions. A candidate revision is proposed only when a diagnosis is repeatedly observed across samples, and it is implemented only if validation against the existing system satisfies the acceptance criteria. EMAS was tested on four benchmarks and two LLMs, achieving the highest task-weighted overall accuracy for both models and excelling or tying in six out of eight scenarios. The research can be found on arXiv with the identifier 2608.07196.
Key facts
- EMAS stands for Evolving Multi-Agent System.
- It revises MAS topology and prompts without updating LLM parameters.
- It uses traces to generate structured diagnoses.
- Candidate revisions are generated only when the same diagnosis recurs.
- Revisions are applied only if validation meets acceptance criteria.
- Evaluated on four benchmarks and two LLMs.
- Achieved highest task-weighted overall accuracy for both backbones.
- Best or tied in six of eight settings.
Entities
Institutions
- arXiv