New method generates commonsense axioms for NLI tasks, boosting LLM accuracy

Researchers have developed a method to generate commonsense knowledge axioms for Natural Language Inference (NLI) tasks, evaluating their effectiveness using LLMs like Llama 3.1 70B and GPT-OSS 120B. A novel reference-free LLM-as-Judge framework was introduced to assess the factuality of these generated axioms, revealing significant performance differences between the models. A hybrid approach that selectively integrates highly factual axioms demonstrated consistent accuracy gains on the SNLI and ANLI benchmarks, improving performance by up to 8.5% and helping models overcome biases. AI

IMPACT Enhances NLI model performance by providing targeted commonsense knowledge, potentially improving reasoning capabilities in AI systems.

RANK_REASON The cluster contains an academic paper detailing a new method for generating and integrating commonsense knowledge for NLI tasks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Chathuri Jayaweera
• CORE Recommender
• GPT-OSS 120B
• Hugging Face
• Llama 3.1 70B
• Stanford Natural Language Inference corpus

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →