Regarding glyphs/sigils, I believe these function as a kind of anchor focusing the model’s attention into an esoteric/mystical/occult semantic island inside the embedding space, somewhat trapping it there. These tokens are little used and rarely represented in normal language, but are relatively far more common in old alchemical texts and in occult/esoteric books, articles, and forums. That means using them might direct the model to this conceptual space more effectively than any natural language could, making them effective as a sort of mild context poisoning.
Regarding glyphs/sigils, I believe these function as a kind of anchor focusing the model’s attention into an esoteric/mystical/occult semantic island inside the embedding space, somewhat trapping it there. These tokens are little used and rarely represented in normal language, but are relatively far more common in old alchemical texts and in occult/esoteric books, articles, and forums. That means using them might direct the model to this conceptual space more effectively than any natural language could, making them effective as a sort of mild context poisoning.