The Power of Test-Time Training for Approximate Sampling

The paper introduces a formalization of test-time training (TTT) for efficient sampling from complex probability distributions, particularly in the context of generative AI and LLMs. It establishes a quadratic lower bound on the query complexity for sampling from a target distribution when using an oracle for approximate density estimates, confirming the optimality of existing random walk methods. The findings suggest that by constraining the class of distributions, practitioners can develop more effective sampling algorithms, enhancing the adaptability of models to specific tasks during inference.