Generalization correlations across datasets are usually modest, so the same checkpoint need not generalize consistently across tasks. Larger OLMo3 has higher mean correlations in every reported evaluation; larger Apertus increases flipped/truthy correlations but decreases repetitive/successive correlations. The source describes the overall trend as greater universality in larger models.

model: olmo3-7b · evaluation: flipped-answer

model: olmo3-7b · evaluation: repetitive-answer

model: olmo3-7b · evaluation: successive-answer

model: olmo3-7b · evaluation: truthy-answer

model: olmo3-32b · evaluation: flipped-answer

model: olmo3-32b · evaluation: repetitive-answer

model: olmo3-32b · evaluation: successive-answer

model: olmo3-32b · evaluation: truthy-answer

model: apertus-8b · evaluation: flipped-answer

model: apertus-8b · evaluation: repetitive-answer

model: apertus-8b · evaluation: successive-answer

model: apertus-8b · evaluation: truthy-answer