One could very well argue that a Markov Chain as seen in Diffusion models is nothing more than following a graph from a point sampled from the prior back to the

One could very well argue that a Markov Chain as seen in Diffusion models is nothing more than following a graph from a point sampled from the prior back to the original distribution. The edges in a said graph are generated either via a scoring model which predicts an estimate of the gradient directly, or by generating a conditional pdf and sampling from it. This is no different to tree search, not really anyway, it's just that ANNs are used to handle the intractability of expanding an infinite and continuous space or providing the probabilities within a MCTS - which is what AlphaGo and AlphaZero do. One could argue that LMs / Causal Transformers are also doing treesearch, indeed one of our evaluation metrics, perplexity, is a measurement of the model's uncertainty in the transition p(token_i | token_{i-1} ... token_{0}).