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

A new study traces the origins of social reasoning in language models by linking benchmark performance to specific regions of the pretraining corpus, using a technique called training-data attribution on the OLMo3-7B model [1]. The research, submitted to arXiv in 2026, applies gradient-based attribution to measure how strongly each training document influences the model's predictions [1]. Because document-level scores are noisy, the team aggregated influence across WebOrganizer's taxonomy of 24 formats by 24 topics, creating 576 bins [1]. They then contrasted benchmark pairs in a 2x2 design that varies domain — social versus STEM — and capability type — reasoning versus knowledge — using SocialIQA, MMLU Social Sciences, ARC-Challenge, and MMLU STEM [1]. The analysis found that social and STEM reasoning draw on qualitatively distinct corpus regions, and the contrast is sharper at the reasoning level than at the knowledge level [1]. To validate the findings, the researchers used targeted machine unlearning. Forgetting high-attribution topic bins — for example, Literature for SocialIQA — degraded performance on the aligned benchmark more than within-bin random baselines [1]. The authors have open-sourced all code, sampling manifests, the bin-level influence matrix, and unlearning checkpoints [1]. The study addresses a gap in prior work, which has emphasized factual knowledge rather than reasoning [1]. Understanding which training data drives specific capabilities is relevant to broader concerns about algorithmic bias, which can emerge from how data is coded, collected, selected, or used to train an algorithm [5]. Bias in AI systems has been documented in areas such as criminal justice, healthcare, and hiring, often compounding existing social inequalities [5]. The European Union's Artificial Intelligence Act, adopted in 2024, represents one regulatory response to such risks [5]. Artificial intelligence research has traditionally pursued goals including learning, reasoning, knowledge representation, and natural language processing [4]. Since the 2020s, generative AI has become widely available, and funding and interest surged after 2012 when graphics processing units accelerated neural networks and deep learning outperformed previous techniques [4]. The new attribution work contributes to the interpretability branch of AI research by providing a method to map capabilities back to their training-data origins [1].