Think or Step-by-Step? UnZIPping the Black Box in Zero-Shot Prompts

2025-02-05

AI Summary

This work investigates the causal mechanisms underlying zero-shot prompting (e.g., “Let’s think step-by-step”) in large language models (LLMs), focusing on the differential contributions of key lexical components (e.g., “think” vs. “step-by-step”). To overcome limitations of existing interpretability methods—such as reliance on open-source models and high computational cost—we propose ZIP, a lightweight, word-level importance scoring metric enabling the first cross-model (open- and closed-source) quantification of token-level prompt efficacy. ZIP is grounded in systematic input perturbation and sensitivity analysis, validated across four LLMs, seven prompt variants, multiple reasoning tasks, and human judgment benchmarks. Results demonstrate that word importance is jointly dependent on both model architecture and task semantics; notably, importance rankings from closed-source models align more closely with human intuition. This work establishes an efficient, generalizable, and interpretable paradigm for prompt engineering and model diagnostics.

Technology Category

• Natural Language Processing: Prompt Engineering / Prompting
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Zero-shot prompting techniques have significantly improved the performance of Large Language Models (LLMs). However, we lack a clear understanding of why zero-shot prompts are so effective. For example, in the prompt"Let's think step-by-step,"is"think"or"step-by-step"more crucial to its success? Existing interpretability methods, such as gradient-based and attention-based approaches, are computationally intensive and restricted to open-source models. We introduce the ZIP score (Zero-shot Importance of Perturbation score), a versatile metric applicable to both open and closed-source models, based on systematic input word perturbations. Our experiments across four recent LLMs, seven widely-used prompts, and several tasks, reveal interesting patterns in word importance. For instance, while both 'step-by-step' and 'think' show high ZIP scores, which one is more influential depends on the model and task. We validate our method using controlled experiments and compare our results with human judgments, finding that proprietary models align more closely with human intuition regarding word significance. These findings enhance our understanding of LLM behavior and contribute to developing more effective zero-shot prompts and improved model analysis.

Problem

• Understanding zero-shot prompt effectiveness
• Developing ZIP score for word importance
• Comparing model and human word significance

Innovation

• ZIP score metric
• systematic input perturbations
• word importance analysis