LAMBADA: Backward Chaining for Automated Reasoning in Natural Language

Achieves massive accuracy boosts over SotA forward reasoning methods on two challenging logical reasoning datasets with a method inspired by backward reasoning.

Implementing LAMBADA can significantly improve accuracy in logical reasoning tasks and reduce the search space for longer chains of reasoning.