TL;DR
A comprehensive survey and position paper on building foundation models that are both reliable (robust, calibrated, safe) and responsible (fair, private, transparent)—essential reading for deploying AI systems in the real world.
Abstract
Foundation models have fundamentally transformed the artificial intelligence landscape, enabling unprecedented capabilities across diverse domains. However, deploying these powerful systems responsibly requires addressing critical challenges related to reliability and responsibility. This survey provides a comprehensive examination of the key dimensions of reliable and responsible foundation models, including robustness, uncertainty quantification, safety, fairness, privacy, and transparency. We discuss the current state of research, identify open challenges, and outline promising directions for building foundation models that are both capable and trustworthy.
Citation
@article{yang2025reliable,
title = {Reliable and Responsible Foundation Models: A Comprehensive Survey},
author = {Yang, Xinyu and Han, Junlin and Bommasani, Rishi and Luo, Jinqi and Qu, Wenjie and Zhou, Wangchunshu and Bibi, Adel and Wang, Xiyao and Yoon, Jaehong and Stengel-Eskin, Elias and Tong, Shengbang and Shen, Lingfeng and Rafailov, Rafael and Li, Runjia and Wang, Zhaoyang and Zhou, Yiyang and Cui, Chenhang and Wang, Yu and Zheng, Wenhao and Zhou, Huichi and Gu, Jindong and Chen, Zhaorun and Xia, Peng and Lee, Tony and Zollo, Thomas P. and Sehwag, Vikash and Leng, Jixuan and Chen, Jiuhai and Wen, Yuxin and Zhang, Huan and Deng, Zhun and Zhang, Linjun and Izmailov, Pavel and Koh, Pang Wei and Tsvetkov, Yulia and Wilson, Andrew Gordon and Zhang, Jiaheng and Zou, James and Xie, Cihang and Wang, Hao and Torr, Philip and McAuley, Julian and Alvarez-Melis, David and Tramèr, Florian and Xu, Kaidi and Jana, Suman and Callison-Burch, Chris and Vidal, René and Kokkinos, Filippos and Bansal, Mohit and Chen, Beidi and Yao, Huaxiu},
journal = {Transactions on Machine Learning Research},
year = {2025},
eprint = {2602.08145},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2602.08145}
}