Epistemic Diversity Mitigates Knowledge Collapse in Language Models

The study explores how diversity among AI language models can mitigate knowledge collapse, a reduction in the breadth of ideas. It finds that an optimal level of diversity is needed, as too few or too many diverse models can lead to performance decay.

Details

The study explores how diversity among AI language models can mitigate 'knowledge collapse' - a reduction in the breadth of ideas expressed by AI systems. Prior work has demonstrated 'single-model collapse', where an AI model's performance decays when trained on its own output. Inspired by ecological diversity, the researchers built ecosystems of language models trained on their collective output and evaluated the effect of diversity on model performance. They segmented the training data across multiple models and tracked performance over 10 self-training iterations. The results show that increased epistemic diversity can mitigate collapse, but only up to an optimal level. Too few diverse models fail to express the full true distribution, leading to rapid performance decay. Yet distributing the data across too many models reduces each model's approximation capacity on the true distribution, also leading to poor performance. The findings suggest the need to monitor diversity across AI systems and develop policies that incentivize more domain- and community-specific models, rather than a monoculture of large language models.