Test whether a language model has generalized a concept by evaluating withheld combinations of concepts rather than only exact examples absent from training.

Soon you can unlock how this was done.

Behind this: the method · what actually resulted.