LLMs Crack a Byzantine Music Notation Cipher

Researchers discovered that large language models like Claude and GPT-4 can decode a peculiar cipher based on the Byzantine music notation Unicode block. This cipher resembles a Caesar cipher, but with an offset of 118784. The models can decode this cipher directly without chain-of-thought, achieving even higher success rates than with regular Caesar ciphers. Researchers hypothesize this is due to a linear relationship between addition in a specific Unicode range and addition in token space, allowing the models to learn a shift cipher based on this relationship. This phenomenon suggests the existence of yet-ununderstood mechanisms within LLMs.