Quantifying Entropy Collapse in Recursively Trained AI Systems: A Framework for Variance Injection and Stagnation Thresholds

Abstract excerpt

<title>Abstract</title> <p> AI systems that train recursively on their own outputs gradually narrow the range of what they can produce—losing rare, high-variance outputs first, long before standard benchmarks register any degradation. No analytical formula has existed for the minimum human contribution required to prevent this collapse, and no practical early warning signal has been proposed. We derive the minim...

Topics

Open a Topic to create a Post that cites this publication.

• Artificial Intelligence
• Decision Support Systems, Clinical
• Electronic Health Records
• Genetic Variation
• Genetics, Medical

Identifiers and source

Literature Corpus work

2e684969-898e-5138-9ee4-aab8090341fb

DOI

10.21203/rs.3.rs-9196766/v1