My first thought after reading the title and before reading the article was “steganography”. LLMs don’t need to hide hours long complex tasks and chain of thoughts to be dangerous. They possibly need few minutes of hidden mental work to figure out and converge on steganography approach where there is a slower hidden chain of thoughts produced by specific use of words, synonyms, repetitions.
Some of current models tend to repeat themselves in CoT few times, using different wording. This behavior makes steganography easier.
My first thought after reading the title and before reading the article was “steganography”. LLMs don’t need to hide hours long complex tasks and chain of thoughts to be dangerous. They possibly need few minutes of hidden mental work to figure out and converge on steganography approach where there is a slower hidden chain of thoughts produced by specific use of words, synonyms, repetitions.
Some of current models tend to repeat themselves in CoT few times, using different wording. This behavior makes steganography easier.