Where Does Social Reasoning Come From? Capability Provenance in Language Models

The paper introduces a methodology for training-data attribution to analyze the sources of social and STEM reasoning capabilities in the OLMo3-7B model. It employs gradient-based attribution to identify distinct corpus regions that support these reasoning types, utilizing a detailed taxonomy of 576 bins, and demonstrates that social reasoning relies on different data compared to STEM reasoning. This work is significant for practitioners as it provides insights into model interpretability and capability provenance, aiding in the development of more targeted training strategies and enhancing understanding of model behavior.