Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View (ICLR 2025)

CogMir 🤖🧠 is a multi-LLM agent framework designed to explore how large language models (LLMs) mirror human cognitive biases and exhibit irrational yet prosocial decision-making.
🌐✨ Our research highlights the potential of using systematic hallucination properties in LLMs to better understand and enhance their social intelligence.

---

This repository contains the core datasets and high-level experimental prompt design logic used in the CogMir framework.

📁 Repository Structure

• CogMir_Data/Core_Datasets
  Fundamental data: questions, agent profiles, scenario descriptions, actions, and narratives. Each subfolder corresponds to a dataset described in Appendix C of the paper.

• CogMir_Data/Experimental_Prompts
  Reusable prompt templates for each cognitive bias experiment (see Section 4 & Appendix D). Templates use placeholders (like [Known/Unknown MCQ], [IDENTITY], [SCENARIO]) to be filled with data from /Core_Datasets.

---

📖 Citation

If our work helps or inspires you, please cite:

@inproceedings{
  liu2025exploring,
  title={Exploring Prosocial Irrationality for {LLM} Agents: A Social Cognition View},
  author={Xuan Liu and Jie Zhang and Haoyang Shang and Song Guo and Chengxu Yang and Quanyan Zhu},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=u8VOQVzduP}
}