## Intelligent Routing to Reduce Energy Consumption of Reasoning Models A recent study focuses on optimizing the energy consumption of large reasoning models (LRMs). These models have heterogeneous inference energy costs, depending on the specific model used and the intensity of reasoning. The main goal is to identify the critical operating point where both auxiliary and baseline energy waste are avoided. This regime is characterized by a balance between the mean energy provisioning and stochastic fluctuations. Excessive baseline energy supply leads to persistent waste, while insufficient supply induces a continuous reliance on auxiliary energy. ## Strategies for Energy Efficiency The research highlights the importance of variance-aware routing and dispatch policies based on training-compute and inference-compute scaling laws for LRMs. This approach provides a theoretical basis for developing energy-aware model routing policies. In summary, efficient energy management in large reasoning models requires careful consideration of variability and dynamic routing that adapts to the specific needs of each model and task.