API
< Back to all clusters
[TECHNOLOGY] · Italy, New Zealand · 5 sources

started · updated

Researchers identify method to prevent AI model collapse

An international study conducted by CNR-ISTI, the University of Pisa, the Scuola Normale Superiore, and Victoria University of Wellington has identified a potential solution to prevent AI model collapse. This phenomenon, often called digital autophagy, occurs when generative AI models are trained on content produced by other AI systems, leading to a progressive loss of variety, originality, and coherence.

By selecting less predictable texts during the training process, researchers aim to preserve the quality and diversity of linguistic models. The study, published in ACM Transactions on Intelligent Systems and Technology, analyzed several generative models, including Meta’s Llama, to observe how different forms of collapse manifest, such as restricted vocabularies and repetitive patterns.

Entities

Ann Francis · CNR-ISTI · Cardiff University · LLaMA · Meta · National Museum Wales · Peter Hughes · University of Pisa · Victoria University of Wellington · Wales