arXiv 2410.14251
Synthesizing Post-Training Data for LLMs through Multi-Agent Simulation
By Shuo Tang, Xianghe Pang, et al.
Published 2024-10-18
Wiki summary
Explore the paper's summary, context, and related research on Papiers.
Post-training is essential for enabling large language models (LLMs) to follow human instructions. However, its effectiveness depends on high-quality instruction data, which is challenging to obtain in the real world due to privacy concerns, data scarcity, and high annotation costs. To fill this gap, inspired by the recent success of using LLMs to simulate human society, we propose MATRIX, a multi-agent simulator th…