Paper: https://arxiv.org/pdf/2503.01935
Github: https://github.com/MultiagentBench/MARBLE
It’s a comprehensive framework for evaluation multi-agent system. They score the final answer + the process and how well it was executed by the team
- Seems promising, they have a set of problems (Reseach, minecraft, …) but also can be extended to other problems and evaluators -> Can be used as the backbone for any other benchmark!!
- They provide a standard for representing teams and communication between them -> quite usefull instead of guessing
