LLM-Generated Commonsense Axioms for NLI: A Reference-Free Evaluation

A recent study published on arXiv (2507.15100) explores the capability of Large Language Models (LLMs) to produce factual commonsense axioms for Natural Language Inference (NLI) and assesses their effectiveness using the SNLI and ANLI benchmarks. Utilizing Llama-3.1-70B and gpt-oss-120b, the research presents a novel reference-free approach through an LLM-as-Judge framework, as traditional factuality metrics are inadequate for commonsense axioms that do not have clear textual references. Results indicate a significant disparity between the models: gpt-oss-120b predominantly generates accurate axioms, while Llama tends to produce more inaccuracies. Additionally, the study examines three prompting strategies, emphasizing the potential and challenges of leveraging LLMs for commonsense knowledge in NLI tasks.