Eluna is a production-deployed agentic system that uses graph-guided, multi-agent frameworks to encode warehouse Standard Operating Procedures as directed acyclic graphs for reliable execution. Developed by Ning Liu and team, it employs asymmetric episodic distillation, where a strong teacher model improved through error memories trains a smaller student on corrected trajectories. Fine-tuned models matched or exceeded their teacher, beat larger baselines, and achieved 94% expert agreement on ticket processing.
No score is assigned. Sources and their independence are shown in the citation chain below.