I don't think this paper goes deep enough, they don't consider for example structured diversity seeding for LLMs, which entails prompting the model to combine d

I don't think this paper goes deep enough, they don't consider for example structured diversity seeding for LLMs, which entails prompting the model to combine diverse entities, knowledge and skills selected at random from a list. Random combinations of conditioning context can lead the model to hyper-diversity. The Skill-Mix paper shows how this can work. https://arxiv.org/abs/2310.17567 https://arxiv.org/abs/2310.17567 LLMs have been trained on everything, when they are also prompted by diverse people with their own tasks and styles it won't lead to knowledge collapse. Two diverse worlds meeting each other in chat LLM interactions - the training corpus and people. But if you remove the human and just use fixed prompts, and then also narrow the training set with various methods, it might collapse.