But we also know from other research that LLMs don't actually do mechanistic translations. Even when they are asked to and say that they did, they're basically

But we also know from other research that LLMs don't actually do mechanistic translations. Even when they are asked to and say that they did, they're basically rewriting the code from their training data