Tag
This paper studies small language models (SLMs) as closed-loop policies for graph algorithm execution, evaluating both step accuracy and rollout reliability across multiple graph procedures. The results show a gap between local decision quality and global execution reliability, especially for weighted algorithms.