Researchers propose a new approach to scaling up the reasoning capabilities of large language models by using multiple agents that work together. This method, called Multi-Agent Systems (MAS), breaks down complex reasoning tasks into smaller, manageable chunks across different agents. The team created a dataset with 500 high-quality examples of multi-agent collaborative reasoning and fine-tuned several open-source models to perform better in MAS settings. They also developed an adaptive scaling strategy that dynamically guides the reasoning process. --- Why it matters: This work matters because it addresses a significant scalability bottleneck in large language models, enabling them to reason more efficiently and effectively in complex tasks. Engineers can apply this approach to improve the performance of their own models in multi-agent collaborative settings. Source: https://arxiv.org/abs/2504.09772