Active Inference as the Test-Time Scaling Law for Physical AI Agents

The paper introduces a novel test-time scaling law for physical AI agents based on active inference, allowing for effective reasoning and generalization in unforeseen scenarios. This law dynamically updates the agent's policy via a soft Bayesian inference process that minimizes prediction errors, enabling learning beyond the training distribution. Simulation results indicate that this approach significantly outperforms traditional methods like Q-learning and Bayesian reinforcement learning in autonomous driving tasks, enhancing inference efficiency by over 36%.