ODAR: Principled Adaptive Routing for LLM Reasoning via Active Inference

Researchers propose ODAR-Expert, an adaptive routing framework for large language models that optimizes accuracy-efficiency trade-offs by dynamically routing queries between fast and slow processing agents. The system achieved 98.2% accuracy on MATH benchmarks while reducing computational costs by 82%, suggesting that optimal AI scaling requires adaptive resource allocation rather than simply increasing test-time compute.

• →ODAR-Expert uses active inference to dynamically route queries between fast heuristic and slow deliberative AI agents based on query difficulty.
• →The framework achieved 98.2% accuracy on MATH benchmarks and 54.8% on Humanity's Last Exam while reducing compute costs by 82%.
• →The system uses a free-energy-principled fusion mechanism that balances log-likelihood with epistemic uncertainty for answer selection.
• →Results suggest optimal AI scaling requires adaptive resource allocation rather than uniform brute-force sampling approaches.
• →The framework was validated on open-source models including Llama 4 and DeepSeek, demonstrating reproducibility across different AI architectures.