Are LLMs Rigorous Logical Reasoner? Empowering Natural Language Proof Generation with Contrastive Stepwise Decoding (2311.06736v2)

Abstract: Logical reasoning remains a pivotal component within the realm of artificial intelligence. The recent evolution of LLMs has marked significant progress in this domain. The adoption of strategies like chain-of-thought (CoT) has enhanced the performance of LLMs across diverse reasoning tasks. Nonetheless, logical reasoning that involves proof planning, specifically those that necessitate the validation of explanation accuracy, continues to present stumbling blocks. In this study, we first evaluate the efficacy of LLMs with advanced CoT strategies concerning such tasks. Our analysis reveals that LLMs still struggle to navigate complex reasoning chains, which demand the meticulous linkage of premises to derive a cogent conclusion. To address this issue, we finetune a smaller-scale LLM, equipping it to decompose proof objectives into more manageable subgoals. We also introduce contrastive decoding to stepwise proof generation, making use of negative reasoning paths to strengthen the model's capacity for logical deduction. Experiments on EntailmentBank underscore the success of our method in augmenting the proof planning abilities of LLMs.