Shared Lexical Task Representations Explain Behavioral Variability In LLMs

One of the most common complaints about large language models (LLMs) is their prompt sensitivity---that is, the fact that their ability to perform a task or provide a correct answer to a question can depend unpredictably on the way the question is posed. We investigate this variation by comparing two very different but commonly-used styles of prompting: instruction-based prompts, which describe the task in natural language, and example-based prompts, which provide in-context few-shot demonstration pairs to illustrate the task. We find that, despite large variation in performance as a function of the prompt, the model engages some common underlying mechanisms across different prompts of a task. Specifically, we identify task-specific attention heads whose outputs literally describe the task --- which we dub lexical task heads---and show that these heads are shared across prompting styles and trigger subsequent answer production. We further find that behavioral variation between prompts can be explained by the degree to which these heads are activated, and that failures are at least sometimes due to competing task representations that dilute the signal of the target task. Our results together present an increasingly clear picture of how LLMs' internal representations can explain behavior that otherwise seems idiosyncratic to users and developers.

Large language models often exhibit high sensitivity to phrasing, meaning that minor variations in how a prompt is worded can unpredictably alter the accuracy of their outputs. To investigate this behavior, we examined the internal mechanisms of these models when given two distinct prompting styles: direct natural language instructions and example-based demonstrations. We discovered that despite the differences in prompt formatting, the model utilizes some shared underlying components—which we call lexical task heads—to represent the task. Furthermore, we found that performance inconsistencies are tied to how strongly these specific components are activated; a poorly constructed prompt can trigger competing internal representations that dilute the necessary signal and lead to errors. By demonstrating that language models rely on shared internal structures to process diverse prompts, this research explains seemingly erratic system behavior and offers insight for building more reliable models.