ODAR: Principled Adaptive Routing for LLM Reasoning via Active Inference

A research team led by Siyuan Ma has introduced ODAR-Expert, a groundbreaking framework that fundamentally rethinks how large language models allocate computational resources during reasoning tasks. The system addresses a critical inefficiency in current approaches: most LLM reasoning methods rely on uniform brute-force sampling (like fixed best-of-N or self-consistency) that wastes computation on easy problems while potentially under-investing in difficult ones. ODAR-Expert instead implements principled adaptive routing, using a difficulty estimator grounded in amortized active inference to dynamically decide whether to send a query to a heuristic Fast Agent or a more deliberative Slow Agent. This represents a paradigm shift from simply scaling test-time compute to optimizing resource allocation based on problem complexity.

The technical innovation extends beyond routing to answer selection, where ODAR-Expert introduces a free-energy-principled, risk-sensitive fusion mechanism. Rather than using ad hoc voting over heterogeneous candidates, the system selects answers by minimizing a variational free energy objective that balances log-likelihood with epistemic uncertainty (varentropy). Extensive evaluation across 23 benchmarks demonstrates remarkable performance, including 98.2% accuracy on MATH and 54.8% on Humanity's Last Exam (HLE), while consistently improving the compute-accuracy frontier. Crucially, the researchers validated reproducibility on a fully open-source stack (Llama 4 + DeepSeek), where ODAR surpassed homogeneous sampling strategies while reducing computational costs by 82%. This suggests that thinking-optimal scaling requires adaptive resource allocation with free-energy-based decision-making rather than simply increasing test-time compute.