Likelihood
The probability a model assigns to real text, typically measured by average negative log-likelihood (cross-entropy loss); it is often used as a proxy metric for model quality.
Timeline (last 90 days)
Fine-tuning on human text can also produce the same divergence in entropy profiles, while likelihood improves.
No single ordering sign reliably indicates worsening likelihood, indicating that the relationship between entropy metrics and model quality is more complex than expected.
Likelihood is essentially a statistic that only reflects relative rarity under a probability distribution, and it cannot distinguish unique expressions written by humans after deliberate thought from meaningless strings randomly copied by machines.
The research does not offer an alternative, but it clearly draws a boundary: likelihood is not equal to authorship; how to more honestly measure human contribution remains a far-from-solved open problem.
All Facts (4)
Fine-tuning on human text can also produce the same divergence in entropy profiles, while likelihood improves.
50%UnverifiedNo single ordering sign reliably indicates worsening likelihood, indicating that the relationship between entropy metrics and model quality is more complex than expected.
50%UnverifiedLikelihood is essentially a statistic that only reflects relative rarity under a probability distribution, and it cannot distinguish unique expressions written by humans after deliberate thought from meaningless strings randomly copied by machines.
50%UnverifiedThe research does not offer an alternative, but it clearly draws a boundary: likelihood is not equal to authorship; how to more honestly measure human contribution remains a far-from-solved open problem.
50%