From scaling laws to agentic grounding: a survey of emerging reliability risks in large language models

Abstract

As the artificial intelligence industry exhausts high-quality, human-generated datasets, the shift toward recursive training on synthetic outputs has introduced significant systemic instabilities. This paper analyzes the compounding risks of model collapse and functional incoherence as failure modes that can result in the irreversible erosion of data diversity and model stability. The results of this analysis are an urgent call to both industry and the research community to recognize that current scaling laws are insufficient for resolving these defects, as increased compute may actually broaden the surface area for incoherent execution. To address these vulnerabilities, a transition toward agentic architectures is proposed that implements non-machine learning constraints as a way to ground model outputs. By integrating structural anchors like state-based persistence and parallel consensus loops, these frameworks provide a deterministic foundation for maintaining logical consistency in high-consequence environments. This research suggests that the path to sustainable, trustworthy AI lies in moving beyond purely probabilistic scaling toward verifiable, context-dependent reliability.