Drift and selection in LLM text ecosystems
Develops an exactly solvable n-gram-based recursive framework analyzing drift and selection effects on public text ecosystems.
Key Findings
Methodology
This paper constructs a mathematically transparent framework based on variable-order n-gram models, describing recursive text generation via conditional probability distributions. It employs fixed point analysis to characterize long-term stable states, distinguishing between drift—caused by finite sampling leading to rare word loss—and selection—driven by filtering mechanisms. The model analyzes both unsmoothed and smoothed scenarios, revealing how different filtering strategies influence the support and depth of the resulting text distribution. The framework leverages the geometry of de Bruijn graphs to describe fixed points, providing a rigorous understanding of how shallow and deep structures emerge and persist in recursive ecosystems.
Key Results
- In the unsmoothed case, the recursive process converges to fixed points characterized by uniform distributions over cycles in the de Bruijn graph, representing support erosion and shallow equilibrium states. When normative selection (rewarding quality or novelty) is introduced, the ecosystem maintains deeper structures, with the KL divergence between the corpus r-gram distribution and the induced continuation law bounded by L log2 s bits. Experiments on Conan Doyle texts confirm that pure drift leads to loss of high-order support, while normative filtering sustains complex, non-shallow distributions, aligning with theoretical predictions.
- Simulation results demonstrate that without smoothing, high-order n-gram support diminishes over generations, but with normative filtering, the system stabilizes with persistent deep structure. The geometric interpretation via circulation polytopes provides a precise description of fixed points, supporting the understanding of how recursive publication influences content diversity and structure. These findings have implications for designing training corpora and filtering strategies in large language models.
- The fixed point analysis reveals that support erosion and deep structure preservation are governed by the interplay of sampling noise and filtering, with the latter capable of maintaining complex, multi-periodic distributions. The results highlight the importance of normative filtering in preventing collapse into shallow, repetitive content, offering a theoretical foundation for optimizing recursive content generation systems.
Significance
This work advances the understanding of how recursive text generation and filtering shape the evolution of public language ecosystems. By providing a rigorous mathematical framework, it clarifies the conditions under which deep, meaningful structures can be preserved or lost. This insight is crucial for developing AI systems that generate diverse, high-quality content without succumbing to collapse or homogenization. The analysis bridges information theory, graph theory, and language modeling, offering tools to design better curation and filtering strategies. As AI-generated content proliferates, understanding these dynamics helps mitigate risks of content degradation, supports ethical AI deployment, and guides the development of robust, controllable language models. The framework also opens avenues for future research into multi-agent ecosystems, multi-modal data, and adaptive filtering mechanisms.
Technical Contribution
The paper introduces a fixed point analysis of variable-order n-gram models in recursive generation, utilizing circulation polytopes on de Bruijn graphs to describe stable distributions. It provides explicit formulas for fixed points in the infinite corpus limit, characterizing support erosion and deep structure retention. The framework distinguishes between descriptive and normative filtering, establishing bounds on the divergence from shallow equilibria. This approach offers a novel theoretical lens to analyze how filtering strategies influence the support and complexity of the public language distribution, bridging graph theory, information theory, and probabilistic modeling. It also generalizes classical Wright-Fisher drift to language ecosystems, providing a rigorous foundation for understanding content evolution.
Novelty
This is the first comprehensive mathematical analysis of recursive text ecosystems using circulation polytopes and fixed point theory. It uniquely combines graph-theoretic geometric descriptions with information-theoretic bounds to explain support erosion and deep structure maintenance. Unlike prior work focusing on single-generation or empirical observations, this study offers a formal, exact characterization of long-term equilibrium states under different filtering regimes, providing new insights into the dynamics of AI-generated content and its impact on language diversity. The distinction between descriptive and normative filters, with explicit bounds on divergence, represents a significant innovation in understanding content support and structure in recursive ecosystems.
Limitations
- The model assumes no smoothing and idealized conditions, which may not fully capture real-world language models that incorporate smoothing, back-off, and neural architectures. These factors could alter the support dynamics and fixed points.
- The analysis primarily applies in the infinite corpus limit; finite data scenarios may exhibit deviations, especially with limited sample sizes or non-stationary distributions.
- The framework does not explicitly model multi-modal data, multi-agent interactions, or complex filtering policies beyond simple normative rules, limiting immediate applicability to real-world systems.
Future Work
Future research will extend this framework to neural language models, incorporating smoothing and neural approximations. Exploring multi-agent ecosystems, multi-modal data, and adaptive filtering strategies will deepen understanding of content evolution. Developing algorithms to optimize normative filters for maintaining deep structure while ensuring diversity and factual correctness is another promising direction. Additionally, empirical validation on large-scale datasets and real-world AI systems will be crucial to refine and operationalize these theoretical insights, ultimately guiding the design of more robust, controllable, and ethically aligned AI content ecosystems.
AI Executive Summary
Deep Dive
Plain Language Accessible to non-experts
想象一个工厂每天都在生产不同的玩具。工厂用各种材料和工具制造出许多不同的玩具，但随着时间推移，工厂会逐渐淘汰一些稀有的材料，只留下常用的材料。这就像内容的漂移，稀有的词逐渐消失。而工厂还会根据市场需求，优先生产受欢迎的玩具，过滤掉那些不受欢迎的。这就像内容的选择机制，决定哪些信息可以进入公共记录。如果只追求市场偏好，工厂的玩具会变得越来越单一，没有新意。但如果工厂鼓励创新，保持多样性，就能生产出丰富多彩的玩具。这就像深层结构的保持，确保内容丰富多样。这个模型帮助我们理解，AI生成的内容也是这样，它们受到“漂移”和“选择”的影响，决定未来会出现什么样的内容。
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 first is drift: unfiltered reuse progressively removes rare forms, and in the infinite-corpus limit we characterise the stable distributions exactly. The second is selection: publication, ranking and verification filter what enters the record, and the outcome depends on what is selected. When publication merely reflects the statistical status quo, the corpus converges to a shallow state in which further lookahead brings no benefit. When publication is normative -- rewarding quality, correctness or novelty -- deeper structure persists, and we establish an optimal upper bound on the resulting divergence from shallow equilibria. The framework therefore identifies when recursive publication compresses public text and when selective filtering sustains richer structure, with implications for the design of AI training corpora.