This paper introduces a mathematical framework to analyze the recursive process where AI-generated text re-enters and shapes the public record from which LLMs learn. It distinguishes between "drift," which removes rare forms through unfiltered reuse, and "selection," which filters content based on criteria like quality, showing normative selection preserves deeper linguistic structures.
The paper proposes a bilayer SIR/SIRS framework to model synthetic data contamination and model collapse within the AI ecosystem. This phenomenological mean-field model treats data corpora and AI models as interacting populations, deriving a basic reproduction number to analyze cross-contamination.
This study explores the existence of task-specific neurons in large language models, focusing on mathematical reasoning and code generation. It introduces an activation-based selectivity metric for neuron pruning, which consistently outperforms random pruning in reducing computational cost and preserving task accuracy, while preventing performance collapse.
O Reinforcement Learning é crucial para escalar modelos de linguagem, mas algoritmos existentes sofrem de instabilidade e colapso do modelo. Para resolver isso e permitir o escalonamento bem-sucedido, propõe-se o algoritmo Group Sequence Policy Optimization (GSPO).