后训练使大语言模型更不像人类
计算与语言
2026-05-27 v2 人工智能
机器学习
摘要
大语言模型（LLMs）日益被用作人类参与者的替代品，但仍不清楚哪些模型最能捕捉人类行为及其原因。为此，我们引入了 Psych-201，这是一个新型数据集，使我们能够大规模测量行为对齐程度。我们发现，后训练——将基础模型转化为有用助手的阶段——在不同模型家族、规模和目标之间， consistently 降低了与人类行为的对齐程度。此外，尽管基础模型持续改进，但这种不对齐在更新的模型生成中进一步扩大。最后，我们发现，persona-induction——一种通过以参与者特定信息为条件来引导模型表现出人类化行为的流行技术——并未在个体层面提高预测精度。综上所述，我们的结果表明，当前用于将 LLMs 转化为有用助手的过程，同时也使其更不像人类行为的模型。
引用
@article{arxiv.2605.07632,
title = {Post-training makes large language models less human-like},
author = {Marcel Binz and Elif Akata and Abdullah Almaatouq and Mohammed Alsobay and Oleksii Ariasov and Franziska Brändle and David Broska and Jason W. Burton and Nuno Busch and Frederick Callaway and Vanessa Cheung and Brian Christian and Julian Coda-Forno and Can Demircan and Vittoria Dentella and Maria K. Eckstein and Noémi Éltető and Michael Franke and Thomas L. Griffiths and Fritz Günther and Susanne Haridi and Sebastian Hellmann and Stefan Herytash and Linus Hof and Eleanor Holton and Isabelle Hoxha and Zak Hussain and Akshay Jagadish and Elif Kara and Valentin Kriegmair and Evelina Leivada and Li Ji-An and Tobias Ludwig and Maximilian Maier and Marcelo G. Mattar and Marvin Mathony and Alireza Modirshanechi and Robin Na and Mariia Nadverniuk and Antonios Nasioulas and Surabhi S. Nath and Helen Niemeyer and Kate Nussenbaum and Sebastian Olschewski and Thorsten Pachur and Stefano Palminteri and Aliona Petrenco and Camille V. Phaneuf-Hadd and Angelo Pirrone and Manuel Rausch and Laura Raveling and Shashank Reddy and Milena Rmus and Evan M. Russek and Tankred Saanum and Kai Sandbrink and Louis Schiekiera and Johannes A. Schubert and Luca M. Schulze Buschoff and Nishad Singhi and Leah H. Somerville and Mikhail S. Spektor and Xin Sui and Christopher Summerfield and Mirko Thalmann and Anna I. Thoma and Taisiia Tikhomirova and Vuong Truong and Polina Tsvilodub and Konstantinos Voudouris and Kristin Witte and Shuchen Wu and Dirk U. Wulff and Hua-Dong Xiong and Songlin Xu and Lance Ying and Xinyu Zhang and Jian-Qiao Zhu and Eric Schulz},
journal= {arXiv preprint arXiv:2605.07632},
year = {2026}
}