One theory that I’ve not seen mentioned here is that there’s been a lot of work based around multiple LLMs in communication. Of these were used in the RL loop we could see similar degradatory effects as those that have recently been in the news with regards to image generation models.