We propose an inference-time intervention framework based on Optimal Transport that generalizes previous methods and allows interpretable control of both LLMs and Diffusion models.