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

In 30 Sekunden

Was passiert ist

Georgia Tech researchers used the open Olmo 3 model stack to trace which training data shapes a language model's social reasoning abilities.

Warum es zählt

Matters for AI safety researchers and engineers studying model behavior, especially those without access to proprietary lab resources and training data.

Achtung

Results show correlations, not simple recipes. Dialogue-rich data influenced social reasoning, but removing it caused only modest score drops, leaving causation unclear.