This study investigates the resilience of large language models (LLMs) to handle noisy instructions, which are common in real-world interactions and system integrations. The research focuses on five types of disruptions: Automatic Speech Recognition (ASR) errors, Optical Character Recognition (OCR) errors, grammatical mistakes, typographical errors, and distractive content. The study uses a "re-pass" strategy to evaluate how well LLMs can correct these errors before processing the instructions. Key findings include: 1. **Performance Impact**: LLMs show varying degrees of resistance to different types of noise. While some models perform better with certain types of errors, their overall performance is significantly affected by the presence of noise. 2. **ASR and OCR Errors**: These errors are particularly challenging for LLMs, leading to a significant decline in performance. The models struggle to handle ASR and OCR errors due to their lack of exposure during pre-training and fine-tuning. 3. **Grammatical and Typographical Errors**: Models exhibit more resilience to grammatical mistakes, which are less disruptive compared to other types of errors. Typographical errors, however, severely impact performance. 4. **Distractive Content**: Both cooperative and non-cooperative distractions lead to performance declines, with non-cooperative distractions having a more significant impact. 5. **"Re-pass" Strategy**: The study evaluates the effectiveness of the "re-pass" strategy, which involves using LLMs to correct noisy instructions before processing them. While ChatGPT demonstrates superior performance in correcting errors, other models struggle, and the strategy introduces new challenges in real-world applications. The research highlights the need for further development to enhance LLMs' resilience to noisy instructions, especially in scenarios involving ASR and OCR. The findings also underscore the importance of addressing the limitations of current models in handling various types of noise and improving their ability to filter out irrelevant content.This study investigates the resilience of large language models (LLMs) to handle noisy instructions, which are common in real-world interactions and system integrations. The research focuses on five types of disruptions: Automatic Speech Recognition (ASR) errors, Optical Character Recognition (OCR) errors, grammatical mistakes, typographical errors, and distractive content. The study uses a "re-pass" strategy to evaluate how well LLMs can correct these errors before processing the instructions. Key findings include: 1. **Performance Impact**: LLMs show varying degrees of resistance to different types of noise. While some models perform better with certain types of errors, their overall performance is significantly affected by the presence of noise. 2. **ASR and OCR Errors**: These errors are particularly challenging for LLMs, leading to a significant decline in performance. The models struggle to handle ASR and OCR errors due to their lack of exposure during pre-training and fine-tuning. 3. **Grammatical and Typographical Errors**: Models exhibit more resilience to grammatical mistakes, which are less disruptive compared to other types of errors. Typographical errors, however, severely impact performance. 4. **Distractive Content**: Both cooperative and non-cooperative distractions lead to performance declines, with non-cooperative distractions having a more significant impact. 5. **"Re-pass" Strategy**: The study evaluates the effectiveness of the "re-pass" strategy, which involves using LLMs to correct noisy instructions before processing them. While ChatGPT demonstrates superior performance in correcting errors, other models struggle, and the strategy introduces new challenges in real-world applications. The research highlights the need for further development to enhance LLMs' resilience to noisy instructions, especially in scenarios involving ASR and OCR. The findings also underscore the importance of addressing the limitations of current models in handling various types of noise and improving their ability to filter out irrelevant content.