NeurIPS 2025, Position Paper Track · Paper · OpenReview · Project page and demo · BibTeX
Chance Jiajie Li*, Jiayi Wu*, Zhenze Mo, Ao Qu, Yuhan Tang, Kaiya Ivy Zhao, Yulu Gan, Jie Fan, Jiangbo Yu, Jinhua Zhao, Paul Pu Liang, Luis Alonso, Kent Larson
MIT Media Lab, MIT EECS, MIT IDSS, MIT CEE, MIT DUSP, Northeastern University, Brown University, McGill University · *equal contribution
Position. Today's LLM social simulations are black boxes: demographics in, behavior out. Matching behavior is weak evidence of fidelity, because the same behavior can come from different mechanisms and an agent can give the right answer for the wrong reason. The paradigm has to make the move psychology made once, from behaviorism to cognitivism.
What we ask for. Reasoning fidelity, in three properties: auditable causality (you can inspect how a stance was formed), grounding in a real individual (heterogeneity preserved rather than averaged away), and consistency under counterfactuals (beliefs revise predictably when the world changes). The paper introduces GenMinds, a modeling paradigm for cognitively grounded agents, and RECAP, a framework for evaluating reasoning fidelity.
Companion work. HugAgent (EMNLP 2026) is the benchmark that tests these properties against real people.
docs/ holds the project page (GitHub Pages), including the interactive belief-graph demo.
@inproceedings{li2025simulating,
title = {Simulating Society Requires Simulating Thought},
author = {Li, Chance Jiajie and Wu, Jiayi and Mo, Zhenze and Qu, Ao and Tang, Yuhan and
Zhao, Kaiya Ivy and Gan, Yulu and Fan, Jie and Yu, Jiangbo and Zhao, Jinhua and
Liang, Paul Pu and Alonso, Luis and Larson, Kent},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Position Paper Track},
year = {2025}
}