• Designed probabilistic controllers that trigger specialized LLMs only when marginal reasoning gains exceed cost thresholds.
• Demonstrated up to 30% reduction in inference cost on multi-hop reasoning benchmarks without hurting accuracy.