Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbations

Research reveals that Large Language Models show varying vulnerabilities to different types of Chain-of-Thought reasoning perturbations, with math errors causing 50-60% accuracy loss in small models while unit conversion issues remain challenging even for the largest models. The study tested 13 models across parameter ranges from 3B to 1.5T parameters, finding that scaling provides protection against some perturbations but limited defense against dimensional reasoning tasks.

• →Math error perturbations cause the most severe degradation in small models (50-60% accuracy loss) but show strong scaling benefits.
• →Unit conversion issues remain challenging across all model sizes, causing 20-30% accuracy loss even in the largest models.
• →Extra steps in reasoning chains have minimal impact on accuracy (0-6% loss) regardless of model scale.
• →Model scaling follows power-law patterns for robustness improvements, but offers limited protection against dimensional reasoning challenges.
• →The findings highlight critical vulnerabilities for deploying LLMs in multi-stage reasoning applications and financial analysis pipelines.