Learning to Reduce: Optimal Representations of Structured Data in Prompting Large Language Models | Synapse