Do Prompt-Based Models Really Understand the Meaning of their Prompts?
This challenges the assumption that prompt-based models comprehend instructions like humans, highlighting a serious limitation in their interpretability and robustness for NLP tasks.
The study investigates whether prompt-based models truly understand their prompts by testing over 30 templates for natural language inference, finding that models learn as fast with irrelevant or misleading prompts as with good ones, even for large models like 175 billion parameters and instruction-tuned models, questioning the analogy to human understanding.
Recently, a boom of papers has shown extraordinary progress in zero-shot and few-shot learning with various prompt-based models. It is commonly argued that prompts help models to learn faster in the same way that humans learn faster when provided with task instructions expressed in natural language. In this study, we experiment with over 30 prompt templates manually written for natural language inference (NLI). We find that models learn just as fast with many prompts that are intentionally irrelevant or even pathologically misleading as they do with instructively "good" prompts. Further, such patterns hold even for models as large as 175 billion parameters (Brown et al., 2020) as well as the recently proposed instruction-tuned models which are trained on hundreds of prompts (Sanh et al., 2022). That is, instruction-tuned models often produce good predictions with irrelevant and misleading prompts even at zero shots. In sum, notwithstanding prompt-based models' impressive improvement, we find evidence of serious limitations that question the degree to which such improvement is derived from models understanding task instructions in ways analogous to humans' use of task instructions.