This paper investigates an under-explored challenge in large language models (LLMs): chain-of-thought prompting with noisy rationales—irrelevant or inaccurate reasoning steps—despite advancements in in-context learning. We construct the NoRa dataset, specifically designed to evaluate LLMs’ robustness to noisy rationales, based on which we reveal a widespread vulnerability among LLMs to such noise, with limited efficacy from existing reasoning methods. To combat this, we propose the contrastive denoising with noisy chain-of-thought (CD-CoT) method to enhance denoising-reasoning capabilities by contrasting noisy rationales with only one clean rationale, thereby advancing the robustness of LLMs in reasoning.