Georgia Tech Traces OLMo Abilities Back to Training Data

Georgia Tech researchers used the fully open OLMo ecosystem and influence functions to trace model performance in reasoning and knowledge tasks to specific training text types. They found dialogue-rich, interpersonal texts boost reasoning far more than knowledge-focused texts.

2026-08-22 ~ 2026-08-22 · 3 related posts

• Georgia Tech Traces LLM Reasoning to Training Data via Open OLMo — allen_ai · 2026-08-22
• Georgia Tech Study: Dialogue-heavy text boosts reasoning more than facts — allen_ai · 2026-08-22
• GeorgiaTech traces OLMo capabilities back to specific training data types — BlancheMinerva · 2026-08-22