Recursive Training Loops in LLMs: How training data properties modulate distribution shift in generated data?

Large language models (LLMs) are increasingly contributing to the creation of content on the Internet. This creates a feedback loop as subsequent generations of models will be trained on this generated, synthetic data. This phenomenon is receiving increasing interest, in particular because previous studies have shown that it may lead to distribution shift - models misrepresent and forget the true underlying distributions of human data they are expected to approximate (e.g. resulting in a drastic loss of quality). In this study, we study the impact of human data properties on distribution shift dynamics in iterated training loops. We first confirm that the distribution shift dynamics greatly vary depending on the human data by comparing four datasets (two based on Twitter and two on Reddit). We then test whether data quality may influence the rate of this shift. We find that it does on the twitter, but not on the Reddit datasets. We then focus on a Reddit dataset and conduct a more exhaustive evaluation of a large set of dataset properties. This experiment associated lexical diversity with larger, and semantic diversity with smaller detrimental shifts, suggesting that incorporating text with high lexical (but limited semantic) diversity could exacerbate the degradation of generated text. We then focus on the evolution of political bias, and find that the type of shift observed (bias reduction, amplification or inversion) depends on the political lean of the human (true) distribution. Overall, our work extends the existing literature on the consequences of recursive fine-tuning by showing that this phenomenon is highly dependent on features of the human data on which training occurs. This suggests that different parts of internet (e.g. GitHub, Reddit) may undergo different types of shift depending on their properties.