Large language models (LLMs) have shown impressive abilities in various tasks, including complex reasoning. Chain-of-thought prompting, where the model is provided with intermediate reasoning steps, has further enhanced this capability. But what happens when these reasoning steps are flawed? New research explores how LLMs handle noisy rationales—reasoning chains containing irrelevant or inaccurate information—and introduces a novel method to help them reason more robustly. Imagine an LLM trying to solve a math problem. It's given a few examples with reasoning steps to learn from, but some of these steps include irrelevant facts or incorrect calculations. Surprisingly, LLMs are easily thrown off by these noisy rationales, significantly impacting their performance. This vulnerability has been highlighted by the creation of the NoRa (Noisy Rationales) dataset, a benchmark designed to test LLMs against such imperfect reasoning chains. NoRa covers mathematical, symbolic, and commonsense reasoning tasks, revealing that even state-of-the-art models struggle with noisy rationales, often performing worse than they would without any examples at all. This vulnerability stems from LLMs' tendency to mimic the provided examples, even when flawed. Existing methods like self-correction and self-consistency offer limited help. Self-correction, where the model tries to fix its own mistakes, often fails without external feedback. Self-consistency, which involves generating multiple answers and choosing the most frequent one, doesn’t address the underlying problem of noisy rationales. To overcome this, researchers have developed a new method called Contrastive Denoising with noisy Chain-of-thought (CD-CoT). This technique leverages a simple yet powerful idea: LLMs can learn to identify noise by comparing a noisy rationale with a clean one. CD-CoT uses a four-step process. First, it rephrases the noisy examples by contrasting them with a clean example. Second, it selects the best rephrased candidates. Third, it explores different reasoning paths based on these refined examples. Finally, it votes on the most frequent answer. Experiments show that CD-CoT significantly improves LLM performance on the NoRa dataset across various noise levels and different LLMs, demonstrating its effectiveness in mitigating the negative impact of noisy rationales. This research highlights a critical challenge in LLM reasoning and provides a practical solution for building more robust and reliable AI systems. The implications are significant, especially as LLMs become increasingly integrated into applications where accurate reasoning is paramount. Future research will focus on creating self-supervised variants of CD-CoT and exploring knowledge-enhanced denoising to further improve robustness and automate the process.