Multi-agent systems unlock better test-time reasoning for AI models, new research shows

A new approach to improving AI reasoning divides complex thinking tasks among multiple agents rather than forcing a single model to handle everything at once. The method addresses a fundamental bottleneck in current test-time scaling strategies, where extending reasoning chains eventually leads to slower inference and instability.

The core problem#

Test-time scaling — the practice of dedicating more computational resources to a model's thinking process — has helped large language models tackle harder reasoning tasks. But pushing single models to think through increasingly long chains of logic creates headaches: context management breaks down, computation gets expensive, and the results become unreliable.

A multi-agent approach#

Instead of asking one model to generate an exhaustive reasoning trace, researchers propose decomposing the problem into separate, focused reasoning contexts distributed across multiple agents in a multi-agent system (MAS). This structural shift offers several advantages: agents work within manageable context windows, reasoning remains interpretable and stable, and the system can coordinate across different thinking paths.

Training and testing the idea#

The team created M500, a dataset of 500 high-quality multi-turn collaborative reasoning examples, generated using DeepSeek-R1. They then fine-tuned open-source models on these collaborative patterns, producing adapted versions that better understand how to reason together. An additional innovation introduces a "CEO" agent that dynamically orchestrates the reasoning process and decides how deep the collaboration should go.

Two fine-tuned models — Qwen2.5-32B-MAS and Phi4-14B-MAS — substantially outperformed their base versions when tested within the AgentVerse framework, suggesting the approach is both practical and effective.