Socia: Training Large Language Models To Simulate Social Constructs and Interventions

Aanisha Bhattacharyya, Susmit Agrawal, Yaman Kumar Singla, Nikitha S R, Tarun Ram Menta, Balaji Krishnamurthy

Keywords: LLM, human opinions, public opinion, PEW, behavior, human behavior, opinion simulation

Abstract:

Large language models (LLMs) have become ubiquitous in various applications. Recently, multiple research studies have tried to simulate social constructs, like opinions, culture, and interventions such as experiments. However, despite this progress, there is much scope of improvement in simulating social constructs. Further, it is crucial that LLMs reflect the opinions of the users they are interacting with.
Current methods that train LLMs to simulate human opinions rely heavily on human-annotated datasets, which are expensive, difficult to scale, and often biased toward specific demographic subgroups. Instead of relying on small, annotated opinion datasets, we introduce a novel approach for unsupervised LLM training on behavioral data. Our approach is based on the maxim in psychology that social transactions are the only observable facets of a society. Leveraging this insight, we developed SOCIA50M comprising over 50 million samples derived from 1.5 million advertisements, including content and demographic viewing behaviors.
We train LLMs on SOCIA50M, demonstrating significant improvements over existing simulation techniques across multiple opinions and cultural alignment benchmarks, including GlobalOpinionQA, OpinionQA, CultureNLI, and CultureBank, without explicitly finetuning on these datasets. We further show using the Time-shared Experiments for Social Sciences (TESS) that LLMs trained on behavior data can be used to predict the outcomes of social science experiments, a novel insight on the predictive power of behavioral training signals.

Our approach addresses key limitations of current methods, offering improved scalability, demographic representation, and adaptability to evolving societal views. Our results suggest the potential for easily available social transaction data to replace or complement traditional expert-annotation-based alignment techniques.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Resubmission: No

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11575