Researchers · Interviewed at COLM 2026
Georgia Tech

COLM 2026 · October 19, 2026 · 7 min
Language models are surprisingly good at social reasoning. But which parts of the pre-training data actually produce that ability?
At COLM 2026 in San Francisco, Chandreyi (Zini) Chakraborty from Mark Riedl's lab at Georgia Tech presents a provenance audit of social reasoning in OLMo 3 7B. The team used influence functions to attribute model capabilities across 5 million pre-training documents. They organized the documents with WebOrganizer into 576 topic × format buckets and averaged influence within each bucket, because attribution at the level of single documents is too noisy to interpret.