Enhancing LLM Robustness to Perturbed Instructions: An Empirical Study

2025-04-03

AI Summary

Large language models (LLMs) exhibit significant sensitivity to character- and word-level perturbations in task instructions; existing work primarily addresses data-level robustness, neglecting instruction-level resilience. Method: We present the first systematic investigation and enhancement of LLMs’ stability and generalization under perturbed instructions. Our approach introduces a lightweight, annotation-free self-denoising strategy that requires no additional supervision and is compatible with both frozen-weight inference and fine-tuning paradigms. Contribution/Results: Evaluated across diverse models—including Llama-3 and Flan-T5—and benchmarks—CoLA, QNLI, SST-2—the method achieves substantial average gains in downstream accuracy. It demonstrates strong robustness against spelling errors, synonym substitutions, and word-order permutations. Our work establishes a new paradigm for instruction robustness research and delivers a practical, scalable solution for improving LLM reliability in real-world instruction variations.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Large Language Models (LLMs) are highly vulnerable to input perturbations, as even a small prompt change may result in a substantially different output. Existing methods to enhance LLM robustness are primarily focused on perturbed data samples, whereas improving resiliency to perturbations of task-level instructions has remained relatively underexplored. In this work, we focus on character- and word-level edits of task-specific instructions, which substantially degrade downstream performance. We experiment with a variety of techniques to enhance the robustness of LLMs, including self-denoising and representation alignment, testing different models (Llama 3 and Flan-T5), datasets (CoLA, QNLI, SST-2) and instructions (both task-oriented and role-oriented). We find that, on average, self-denoising -- whether performed by a frozen LLM or a fine-tuned model -- achieves substantially higher performance gains than alternative strategies, including more complex baselines such as ensembling and supervised methods.

Problem

• Enhancing LLM robustness to perturbed instructions
• Addressing performance drop from instruction-level perturbations
• Comparing self-denoising with other robustness techniques

Innovation

• Self-denoising enhances LLM robustness
• Representation alignment improves instruction resilience
• Character-word edits test perturbation impact