Show HN: XML, Markdown, or JSON: Which gives LLMs the most reliable boundaries? (systima.ai)

A recent study called the Delimiter Hypothesis investigates how different prompt delimiters—XML, Markdown, and JSON—affect large language models (LLMs) in terms of understanding boundaries. The research involved testing four models (including GPT-5.2 and Claude Opus 4.6) across three formats and ten tasks, totaling 600 model calls. The findings reveal that while formatting generally had minimal impact on model performance, Markdown significantly underperformed, particularly with the MiniMax M2.5 model, which displayed a serious vulnerability to prompt injection when using Markdown. This study is vital for the AI/ML community as it addresses a critical aspect of prompt engineering, especially in regulated environments. The ability of models to accurately differentiate between instructions, context, and constraints is essential to prevent security risks like prompt injections. With historical recommendations around delimiter formats based largely on intuition rather than data, this benchmark provides evidence-based insights, emphasizing that robust delimiter formats like XML and JSON ensure better boundary comprehension and system integrity. Developers and engineers can utilize these findings to enhance the security and reliability of LLM-based applications.