Explanation Selection Using Unlabeled Data for In-Context Learning

Recent work has addressed textual reasoning tasks by prompting large language models with explanations via the chain-of-thought paradigm. However, subtly different explanations can yield widely varying downstream task accuracy, so explanations that have not been "tuned" for a task, such as off-the-shelf explanations written by nonexperts, may lead to mediocre performance. This paper tackles the problem of how to optimize explanation-infused prompts in a black-box fashion. We first generate sets of candidate explanations for each example in the prompt using a leave-one-out scheme. We then use a two-stage framework where we first evaluate explanations for each in-context example in isolation according to proxy metrics. Finally, we search over sets of explanations to find a set which yields high performance against a silver-labeled development set, drawing inspiration from recent work on bootstrapping language models on unlabeled data. Across four textual reasoning tasks spanning question answering, mathematical reasoning, and natural language inference, results show that our proxy metrics correlate with ground truth accuracy and our overall method can effectively improve prompts over crowdworker annotations and naive search strategies.