Study Reveals the Root Cause of LLM Prompt Sensitivity: Shared Lexical Task Representations

Introduction: The Prompt Sensitivity Problem in LLMs

One of the most criticized issues with large language models (LLMs) is so-called "prompt sensitivity" — for the same task or question, the model may produce drastically different answers due to subtle differences in how the question is phrased. This unpredictable behavioral variability has long troubled researchers and developers. Recently, a paper published on arXiv (arXiv:2604.22027v1) offers a entirely new explanatory perspective for this phenomenon: Shared Lexical Task Representations are the key mechanism behind LLM behavioral variability.

Core Findings: A Deep Comparison of Instruction-Based and Example-Based Prompts

The research team chose two prompting styles that are fundamentally different in practice yet both widely used as their entry point: instruction-based prompts, which describe task requirements in natural language, and example-based prompts, which let the model "grasp" the task intent through input-output demonstrations.

Despite the vast surface-level differences between these two prompting approaches, the researchers discovered a surprising phenomenon: the model's internal representations when processing both types of prompts exhibit a highly shared structure at the lexical level. In other words, whether users "tell" the model what to do or "show" the model what to do, the model tends to map tasks into a similar lexicalized representation space at a fundamental level.

The significance of this finding lies in the fact that it is precisely this shared representation that gives rise to both consistency and subtle deviations in model behavior across different prompting methods. When the lexical representations triggered by different prompts shift, the model's outputs produce the puzzling behavioral variability we observe.

Technical Analysis: Why the Lexical Layer Is Key

From a technical perspective, this research makes contributions in several key areas:

1. Revealing the "Bottleneck Layer" of Task Understanding

The study shows that an LLM's understanding of tasks is not uniformly distributed across all network layers. Instead, critical task encoding occurs at the lexical layer — the representation space closely associated with token embeddings and output projections. This means the prompt sensitivity problem may not stem from the model "failing to understand" the task, but rather from differences in where different prompts map to within the lexical representation space.

2. Unifying the Understanding of Two Mainstream Prompting Paradigms

Previously, instruction-based prompting and example-based prompting (i.e., few-shot learning) were typically studied as two independent mechanisms. This study is the first to demonstrate from an internal representation perspective that both share the same task representation foundation within the model, laying the groundwork for a unified theoretical framework for prompt engineering.

3. Pointing the Way Toward Mitigating Prompt Sensitivity

Since the root cause of behavioral variability lies in shifts in lexical-layer task representations, future optimization strategies can be more targeted — for example, by aligning representations across different prompting methods in the lexical space, or by enhancing the robustness of lexical-layer representations during training, fundamentally reducing the model's prompt sensitivity.

Industry Impact and Future Outlook

Prompt sensitivity is not merely an academic topic — it is a core barrier preventing LLMs from being deployed in high-reliability scenarios such as healthcare, law, and finance. If a model's output produces inconsistent results due to minor variations in question wording, users cannot establish trust in the system.

This research provides actionable theoretical foundations for solving this problem. In the future, researchers can leverage the "shared lexical task representations" framework to develop more robust prompting strategies, and even implement targeted optimizations at the model architecture and training pipeline levels.

Notably, as our understanding of LLM internal mechanisms continues to deepen, "interpretability" and "reliability" are evolving from independent research directions toward convergence. Understanding why a model behaves inconsistently is the first step toward making it stable. This study undoubtedly represents an important stride along that path.