ResearcharXiv
CAIRN: Dynamic Fact-Intent DAGs for Multi-Agent Exploration
Zuyao Xu, Yuyang Jia, Junwei Guan, Xiang Li, Kaiwen Shen, Zhiqiang Dong
Coordinating multi-agent exploration through a shared graph of facts and intents.
The idea
CAIRN represents observations and planned investigations as a dynamic directed acyclic graph. Facts preserve findings; intents record the next investigations and the evidence that supports them.
How exploration proceeds
A reasoner interprets the shared graph and proposes intents. Workers execute investigations and publish new facts, allowing parallel branches to reuse evidence and converge on a goal.
- Persistent goals, dependencies, and findings across worker sessions.
- Traceable execution paths for human inspection and intervention.
- Evaluation on cybersecurity and mathematical reasoning tasks.
The trade-off
Graph-based coordination adds overhead on easier tasks. On high-effort tasks in the paper’s evaluation, CAIRN more often finds solutions faster, illustrating when parallel exploration can justify its coordination cost.