How a Georgia Tech team used the open Olmo stack to trace social reasoning

Literature and interpersonal writing appear to support a different capability from knowing social-science facts. In a study described by Ai2, researchers estimated how sampled training documents influenced Olmo 3's benchmark answers. Narrative and dialogue-rich material mattered especially for SocialIQA, which tests reasoning about social situations; factual-knowledge tests depended more on explanatory writing. The account describes the comparison.

The researchers also made the model unlearn influential literature documents. SocialIQA performance fell more than after unlearning randomly selected literature documents, strengthening the evidence that the identified material supported that ability. This gives researchers a way to test training-data choices, although it does not establish that simply adding more literature improves social reasoning.