Imagine trying to learn a new skill from a slightly inaccurate instruction manual. Frustrating, right? That's the challenge facing Large Language Models (LLMs) when they use "in-context learning" (ICL). ICL is like giving an LLM a few examples to learn from, instead of explicitly reprogramming it. It's incredibly powerful, but new research reveals a surprising vulnerability: noisy data. While LLMs have shown some resilience to errors in text classification, this research demonstrates that even small inaccuracies in training examples can significantly harm performance on text *generation* tasks. Think of it like this: if you're learning to write poems from examples with grammatical errors, your own poems are likely to suffer. The researchers explored this phenomenon by testing LLMs on various tasks, including question answering, reading comprehension, and code generation. They introduced different types of "noise" into the examples and observed a consistent drop in performance, especially when using more sophisticated example selection methods. So, what's the solution? The team developed a clever technique called "Local Perplexity Ranking" (LPR). LPR works by analyzing the "perplexity" of the LLM—essentially, how surprised it is by the data. By comparing the perplexity of an example to its "neighbors" in the dataset (similar examples), LPR can identify and replace noisy examples with cleaner ones. This local comparison helps filter out the bad examples while preserving the benefits of sophisticated selection strategies. The results are impressive. LPR significantly boosted the performance of LLMs on noisy datasets, sometimes by as much as 18%. This research highlights a critical challenge in ICL and offers a promising solution. As LLMs become more prevalent, ensuring the quality of their training data will be crucial for reliable performance. LPR offers a practical and effective way to do just that, paving the way for more robust and reliable AI systems.