Researchers studied the internal representations of large language models (LLMs) to understand how they process prompts. They used a method called intrinsic dimension to analyze the geometry of token representations at different layers of the model. The study found that the intrinsic dimension peaks in early to middle layers and increases when tokens are shuffled, indicating a correlation with uncertainty and surprisal. A linear probe was trained on the per-layer intrinsic dimension profile to distinguish between malicious and benign prompts, achieving high accuracy. This work suggests that prompt-level geometry can provide actionable signals for monitoring and controlling LLM behavior. --- Why it matters: This research matters because it provides insights into how large language models process and respond to user input. Understanding these dynamics is crucial for developing practical safety tools and ensuring the reliability of AI systems in applications such as text generation and decision-making. Source: https://arxiv.org/abs/2501.10573