“Optimizing a large language model's test-time compute can be more effective than scaling up its parameters, allowing smaller models with advanced inference algorithms to offer Pareto-optimal trade-offs in cost and performance.”