Explore our comprehensive analysis of undetected glitch tokens across multiple large language models (LLMs) on our dedicated webpage. Here, you will find detailed statistics on the nature of these tokens, presented in a clear, tabular format. 

For instance, Llama2 predominantly shows glitches classified as words (85.82%), whereas Mistral has a notable percentage of special characters (18.47%). Similarly, models like Gemma and Qwen reveal diverse distributions with substantial proportions of letter-character and character types, respectively. Our data provides insights into the variable nature of undetected glitch tokens, highlighting the challenges and complexities involved in enhancing the robustness of LLMs.