Resilience of Large Language Models for Noisy Instructions

TL;DR

This study investigates the resilience of LLMs against five common types of disruptions including 1) ASR errors, 2) OCR errors, 3) grammatical mistakes, 4) typographical errors, and 5) distractive content, and evaluates a "re-pass"strategy, designed to purify the instructions of noise before the LLMs process them.

Abstract

As the rapidly advancing domain of natural language processing (NLP), large language models (LLMs) have emerged as powerful tools for interpreting human commands and generating text across various tasks. Nonetheless, the resilience of LLMs to handle text containing inherent errors, stemming from human interactions and collaborative systems, has not been thoroughly explored. Our study investigates the resilience of LLMs against five common types of disruptions including 1) ASR (Automatic Speech Recognition) errors, 2) OCR (Optical Character Recognition) errors, 3) grammatical mistakes, 4) typographical errors, and 5) distractive content. We aim to investigate how these models react by deliberately embedding these errors into instructions. Our findings reveal that while some LLMs show a degree of resistance to certain types of noise, their overall performance significantly suffers. This emphasizes the importance of further investigation into enhancing model resilience. In response to the observed decline in performance, our study also evaluates a "re-pass" strategy, designed to purify the instructions of noise before the LLMs process them. Our analysis indicates that correcting noisy instructions, particularly for open-source LLMs, presents significant challenges.

Figures (6)

• Figure 1: Our analysis scrutinized 500 inputs from real users, focusing on three distinct types of noise. The findings reveal that more than 40% of the inputs to the model are affected by noise.
• Figure 2: Our study identifies and assesses the impact of five distinct categories of textual disruptions on the ChatGPT-3.5 model's effectiveness. We noted a reduction in accuracy between 2.5% to 8.2% across the MMLU dataset, a phenomenon directly linked to these varied types of noisy instructions.
• Figure 3: Evaluation of the performance of three Large Language Models (LLMs) using the adapted MMLU dataset, emphasizing different error ratios, as measured by Word Error Rate (WER). The x-axis represents the WER values for Automatic Speech Recognition (ASR) and Optical Character Recognition (OCR), indicated within brackets. The performance declines with noisy instructions.
• Figure 4: Evaluation of the performance of three Large Language Models (LLMs). The x-axis represents the WER values for grammatical mistakes and typographical errors, indicated within brackets.
• Figure 5: The performance of LLMs with both cooperative and non-cooperative distactive content. Both lead to performance declines while non-cooperative distractions have a more disruptive impact.