Learning by Surprise: Surplexity for Mitigating Model Collapse in Generative AI

NetSciX 2026 Conference Submission195 Authors

Keywords: LLM, Model Collapse, Linguistic Diversity, Synthetic Data, Artificial Intelligence

TL;DR: We propose a new characterization of model collapse in LLM autophagy based on next token probabilities, building a mitigation strategy and a network representation of the phenomenon

Submission Number: 195