The document proposes a framework for learning to rank user queries to detect search tasks from query logs. It reduces the query similarity learning (QSL) problem to a learning to rank problem and uses gradient boosted regression trees and LambdaMART to learn a query similarity function. Experiments on an AOL query log dataset show these learning to rank techniques more effectively detect search tasks compared to baselines like logistic regression and decision trees. Future work will apply the search task discovery framework in a streaming setting to detect tasks in real time.