Drift and Selection in LLM Text Ecosystems
Summary: arXiv:2604.08554v1 Announce Type: cross
Abstract: The public text record — the material from which both people and AI systems now learn — is increasingly shaped by its own outputs. Generated text enters the public record, later agents learn from it, and the cycle repeats. Here we develop an exactly solvable mathematical framework for this recursive process, based on variable-order n-gram agents, and separate two forces acting on the public corpus.
The interaction between drift and selection in large language models (LLMs) is a complex phenomenon that has profound implications for the quality and evolution of AI-generated text. As AI systems increasingly rely on the public text corpus for training, understanding how these systems influence and are influenced by the data they generate is essential.
The Forces at Play
There are two primary forces that shape the evolution of public text: drift and selection. Each plays a distinct role in determining the characteristics of the text available for learning and generation.
- Drift: This refers to the unfiltered reuse of text, which progressively eliminates rare forms and contributes to homogenization within the corpus. In an infinite-corpus scenario, we can characterize the stable distributions that emerge as a result of this drift, resulting in a predictable yet limited diversity of outcomes.
- Selection: Unlike drift, selection involves the processes of publication, ranking, and verification that filter what becomes part of the public record. The outcome of selection is highly dependent on the criteria used; when publication reflects the existing statistical status quo, the corpus tends to converge to a shallow state. Conversely, when publication emphasizes quality, correctness, or novelty, it fosters the persistence of deeper structural elements within the text.
The Impact of Recursive Publication
The mathematical framework we propose allows us to explore the implications of recursive publication on the public text corpus. The outcome of this recursive process can lead to two distinct scenarios:
- In cases where publication merely replicates the existing patterns, the corpus is likely to stagnate, resulting in a shallow state where further exploration or deeper lookahead provides little additional benefit.
- In contrast, when publication is normative and actively promotes high-quality contributions, the corpus can sustain a richer structure, allowing for greater diversity and complexity in the generated text. Our findings establish an optimal upper bound on the divergence from these shallow equilibria, indicating a threshold beyond which richer structures can be maintained.
Implications for AI Training Corpora
This framework identifies critical points at which recursive publication compresses the quality of public text and highlights scenarios where selective filtering can enhance the richness of the corpus. The implications of this research extend to the design and curation of AI training corpora, emphasizing the importance of balancing drift and selection to ensure the ongoing evolution of high-quality language models.
Ultimately, as AI systems continue to interact with and learn from the public text record, fostering environments that prioritize quality and novelty will be essential to avoid stagnation and maintain a dynamic and diverse linguistic ecosystem.