🤖 AI Summary
This work addresses the insufficient logical rigor and frequent premise-conclusion disconnections in natural language proof generation by large language models (LLMs), particularly within chain-of-thought (CoT) reasoning. To this end, we propose a subgoal-decomposition and contrastive stepwise decoding framework. Our key contribution is the first integration of negative reasoning paths into stepwise decoding, coupled with fine-grained, subgoal-driven proof planning. We further enhance the framework via lightweight model fine-tuning, multi-step entailment modeling (leveraging EntailmentBank), and contrastive learning to strengthen logical coherence and structural integrity. Experimental results on the EntailmentBank benchmark demonstrate substantial improvements over strong CoT baselines: our method achieves significant gains in Proof Accuracy and Step Consistency—key metrics reflecting both global correctness and local logical soundness—thereby advancing the reliability and fidelity of LLM-generated formal proofs.
📝 Abstract
Logical reasoning remains a pivotal component within the realm of artificial intelligence. The recent evolution of large language models (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 language model, 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 language models.
