EMNLP accepts study on memory and creativity in language diffusion models, author says
The author reports that as the model enters generalization, its token recovery rate after perturbation improves on unseen test examples and worsens on training examples.
TLDR
A study author announced acceptance to EMNLP’s Main Track for work exploring how memories and novel attractors—which the author describes as creativity—coexist in language diffusion models. The author reports a shift as the model generalizes: its recovery rate for tokens, or units of text, after perturbation improves on unseen test examples but worsens on training examples.
Combined views
—
2 Sources, first seen ago
