Where Does Social Reasoning Come From? Capability Provenance in Language Models
For researchers studying capability provenance in language models, this work provides a method to attribute reasoning capabilities to specific pretraining data regions, though it is incremental as it applies existing attribution methods to a new domain (reasoning vs. knowledge).
Using training-data attribution, the authors map which pretraining corpus regions support social-reasoning versus STEM-reasoning in OLMo3-7B, finding that social and STEM reasoning draw on distinct corpus regions, with sharper contrast at the reasoning level than knowledge level. Targeted unlearning of high-attribution topic bins degrades the aligned benchmark more than random baselines.
We use training-data attribution as an interpretable tool for capability discovery, mapping which regions of the pretraining corpus support social-reasoning versus STEM-reasoning in OLMo3-7B. Training-data attribution measures how strongly each training document influences a model's predictions on a benchmark, but document-level scores are too noisy to identify which corpus regions support which capabilities, and prior work has emphasized factual knowledge rather than reasoning. We compute gradient-based attribution (TrackStar via Bergson) over a working set drawn from the de-duplicated Dolma3 mix, aggregate influence across WebOrganizer's 24-format x 24-topic taxonomy (576 bins), and contrast benchmark pairs in a 2x2 design that varies domain (social vs. STEM) and capability type (reasoning vs. knowledge): SocialIQA and MMLU Social Sciences against ARC-Challenge and MMLU STEM. Social and STEM reasoning draw on qualitatively distinct corpus regions, and the contrast is sharper at the reasoning level than at the knowledge level. Targeted machine unlearning provides partial causal validation: forgetting high-attribution topic bins (e.g., Literature for SocialIQA) degrades the aligned benchmark more than within-bin random baselines, and we open-source all code, sampling manifests, the bin-level influence matrix, and unlearning checkpoints.