Uncertainty-Guided Optimization on Large Language Model Search Trees

원문 초록

Tree search algorithms such as greedy and beam search are the standard when it comes to finding sequences of maximum likelihood in the decoding processes of large language models (LLMs). However, they are myopic since they do not take the complete root-to-leaf path into account. Moreover, they are agnostic to prior knowledge available about the process: For example, it does not consider that the objective being maximized is a probability and thereby has specific properties like being bound in the unit interval. Taking a probabilistic approach, we define prior beliefs over LLMs' transition probabilities and obtain posterior beliefs over the most promising paths in each iteration. These beliefs are useful for defining a sample-based, non-myopic acquisition function that allows for a more data-efficient exploration scheme than standard search algorithms on LLMs. Crucially, unlike expensive simulation-based non-myopic methods like the Monte Carlo tree search, our method only requires samples from the beliefs. Our formulation thus views LLM decoding as Bayesian optimization on trees. We discuss how to select the prior and the acquisition function, and demonstrate in experiments with various LLMs that our method achieves higher efficiency than recent baselines: Our method achieves the same or a higher likelihood while expanding fewer nodes.