Are there LLMs that don't do next token prediction? If not, isn't the associated misalignment just a baseline for all LLMs and the other flavors are layered on top of it, not alternatives to it?
Are there LLMs that don't do next token prediction? If not, isn't the associated misalignment just a baseline for all LLMs and the other flavors are layered on top of it, not alternatives to it?