AI models see every letter better with character-level tokenization

Research shows subword tokenization limits LLMs' character-level understanding: ChatGPT splits "LMU München" into "LM", "U" and "München", never seeing individual letters. Character-level approaches improve tasks like spelling words backward.

• Subword tokenization speeds up LLMs but hides individual letters
• Example: "LMU München" becomes three chunks: "LM", "U", "München"
• Character-level models perform better on symbol-level tasks