Where does an AI model learn to interpret emotions, intentions, or everyday moral decisions? A Georgia Tech team used the open Olmo ecosystem to trace which types of texts influence these abilities and test whether those data actually have a causal effect.
A magnifying glass over training data
Glenn Matlin and Chandreyi Chakraborty studied a language model’s social reasoning: its ability to understand what people believe, feel, or are trying to do in a given situation.
To investigate the origin of this ability, they used influence functions, a technique that estimates how much each training document contributed to a specific model response. That document could be an article, a web post, a technical page, or a literary text.
The idea is simple, even if the analysis is complex: by adding up the influence of many documents, researchers can identify which content categories appear to support a specific ability.
It’s not enough to know what a model answers. It also matters to understand which parts of its data helped produce that answer.
The study was only feasible with a fully open model. That’s why the team chose Olmo 3, developed by Ai2. In addition to its weights, the project publishes its training data, model checkpoints, and evaluation tools.
Olmo’s open ecosystem
The research combined five components from the project:
- Olmo 3, the language model being analyzed.
- Dolma 3, the dataset used to train it.
- WebOrganizer, which classifies documents by topic and format.
- OlmoEval, with tests for measuring different capabilities.
- OLMES, a shared standard for scoring evaluations.
Dolma 3 contains approximately 1.26 billion documents, far too many to review one by one. The team selected a stratified sample of around 5.68 million, distributed across the dataset’s 576 categories.
They then calculated how much each document influenced Olmo 3’s answers and grouped the results by category. The process was repeated across four tests:
- SocialIQA, focused on reasoning about social situations.
- ARC-Challenge, aimed at general reasoning.
- The social science subsets of MMLU, for factual knowledge.
- The STEM subsets of MMLU, also focused on factual knowledge.
This gave each benchmark its own profile: a kind of map showing which classes of documents had the strongest relationship with the model’s answers.
Literature and dialogue matter in social reasoning
The most interesting result was not a simple divide between “social” and “scientific” content. Social science knowledge looked more similar to STEM knowledge and general reasoning than to the social reasoning evaluated by SocialIQA.
SocialIQA was the outlier. The model’s answers were more closely related to narrative and interpersonal texts, such as literature, social life, customer service, and question-and-answer conversations.
The other tests, by contrast, were more influenced by expository material, such as technical documentation and scientific texts. According to the researchers, categories with more dialogue and emotional language influenced social reasoning more than any of the other evaluations.
Does this mean that simply adding novels is enough to create a more empathetic model? Not so fast. The study does not offer an automatic recipe. What it shows is that certain types of language could provide useful structures for interpreting relationships, intentions, and emotions.
In addition, texts rich in dialogue and interpersonal relationships had more influence on general reasoning tests than on factual knowledge tests. This suggests that this material could contribute to reasoning itself, not just to answering questions about social situations.
A causal test with Olmo 3
Finding a correlation does not prove that a dataset is responsible for an ability. That’s why Matlin and Chakraborty conducted a second test.
The team made Olmo 3 forget the most influential documents in Dolma 3’s literature category, which was the category most closely related to its performance in social reasoning. They then compared the result with removing documents selected at random from the same category.
Olmo 3’s SocialIQA score dropped more when the documents identified as influential were removed. That result strengthens the possibility that some texts contributed causally to the ability, although it does not turn the finding into a complete explanation of the model’s behavior.
This distinction matters. Influence functions provide an estimate of data impact, but interpreting a language model remains an uncertain task. A document’s influence can depend on the rest of the corpus, the model’s state, and how the ability is evaluated.
Why it matters that the model is open
Until now, tracing an ability through a complete training dataset was mainly a task reserved for large labs with internal access to their data, models, and tools.
In this case, an external team was able to work with a substantial part of the same scientific ecosystem because Ai2 published the necessary artifacts. That openness makes it possible to repeat experiments, question results, and explore auditing methods that would be difficult to apply to closed models.
For AI safety researchers, this represents more than access to the model’s weights. It also makes it possible to study the data, compare checkpoints, review evaluations, and connect observed behaviors to specific training decisions.
In the long term, this type of analysis could help determine which data should be enriched, reweighted, or removed during training. It does not offer a formula such as “more literature produces better social reasoning,” but it does provide a more informed way to investigate what each type of content contributes.
AI does not learn its abilities from one identifiable source. It builds them through millions of interactions among data, architecture, training, and evaluation. Access to that process does not eliminate the mystery, but it allows us to turn it into a scientific question that other researchers can test.