Introducing Operadic Consistency: A New Signal for Detecting Reasoning Failures in LLMs

Researchers have developed a new diagnostic tool known as operadic consistency (OC) to detect reasoning failures in large language models (LLMs) without the need for ground-truth labels. This method evaluates a model's responses to compositional queries, ensuring that direct answers align with those generated from decomposed queries. In a study involving twelve instruction-tuned LLMs across four multi-hop question-answering datasets, OC demonstrated a strong correlation with accuracy, with Pearson correlation coefficients ranging from 0.86 to 0.94. OC outperformed other confidence baselines like chain-of-thought self-consistency (CoT-SC) in several datasets, providing valuable insights and improvements in selective prediction accuracy. The findings indicate that OC offers a robust alternative for assessing LLM performance in complex reasoning tasks.