Saved in:
| Main Authors: | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.07632 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914601592946688 |
|---|---|
| author | Binz, Marcel Akata, Elif Almaatouq, Abdullah Alsobay, Mohammed Ariasov, Oleksii Brändle, Franziska Broska, David Burton, Jason W. Busch, Nuno Callaway, Frederick Cheung, Vanessa Christian, Brian Coda-Forno, Julian Demircan, Can Dentella, Vittoria Eckstein, Maria K. Éltető, Noémi Franke, Michael Griffiths, Thomas L. Günther, Fritz Haridi, Susanne Hellmann, Sebastian Herytash, Stefan Hof, Linus Holton, Eleanor Hoxha, Isabelle Hussain, Zak Jagadish, Akshay Kara, Elif Kriegmair, Valentin Leivada, Evelina Ji-An, Li Ludwig, Tobias Maier, Maximilian Mattar, Marcelo G. Mathony, Marvin Modirshanechi, Alireza Na, Robin Nadverniuk, Mariia Nasioulas, Antonios Nath, Surabhi S. Niemeyer, Helen Nussenbaum, Kate Olschewski, Sebastian Pachur, Thorsten Palminteri, Stefano Petrenco, Aliona Phaneuf-Hadd, Camille V. Pirrone, Angelo Rausch, Manuel Raveling, Laura Reddy, Shashank Rmus, Milena Russek, Evan M. Saanum, Tankred Sandbrink, Kai Schiekiera, Louis Schubert, Johannes A. Buschoff, Luca M. Schulze Singhi, Nishad Somerville, Leah H. Spektor, Mikhail S. Sui, Xin Summerfield, Christopher Thalmann, Mirko Thoma, Anna I. Tikhomirova, Taisiia Truong, Vuong Tsvilodub, Polina Voudouris, Konstantinos Witte, Kristin Wu, Shuchen Wulff, Dirk U. Xiong, Hua-Dong Xu, Songlin Ying, Lance Zhang, Xinyu Zhu, Jian-Qiao Schulz, Eric |
| author_facet | Binz, Marcel Akata, Elif Almaatouq, Abdullah Alsobay, Mohammed Ariasov, Oleksii Brändle, Franziska Broska, David Burton, Jason W. Busch, Nuno Callaway, Frederick Cheung, Vanessa Christian, Brian Coda-Forno, Julian Demircan, Can Dentella, Vittoria Eckstein, Maria K. Éltető, Noémi Franke, Michael Griffiths, Thomas L. Günther, Fritz Haridi, Susanne Hellmann, Sebastian Herytash, Stefan Hof, Linus Holton, Eleanor Hoxha, Isabelle Hussain, Zak Jagadish, Akshay Kara, Elif Kriegmair, Valentin Leivada, Evelina Ji-An, Li Ludwig, Tobias Maier, Maximilian Mattar, Marcelo G. Mathony, Marvin Modirshanechi, Alireza Na, Robin Nadverniuk, Mariia Nasioulas, Antonios Nath, Surabhi S. Niemeyer, Helen Nussenbaum, Kate Olschewski, Sebastian Pachur, Thorsten Palminteri, Stefano Petrenco, Aliona Phaneuf-Hadd, Camille V. Pirrone, Angelo Rausch, Manuel Raveling, Laura Reddy, Shashank Rmus, Milena Russek, Evan M. Saanum, Tankred Sandbrink, Kai Schiekiera, Louis Schubert, Johannes A. Buschoff, Luca M. Schulze Singhi, Nishad Somerville, Leah H. Spektor, Mikhail S. Sui, Xin Summerfield, Christopher Thalmann, Mirko Thoma, Anna I. Tikhomirova, Taisiia Truong, Vuong Tsvilodub, Polina Voudouris, Konstantinos Witte, Kristin Wu, Shuchen Wulff, Dirk U. Xiong, Hua-Dong Xu, Songlin Ying, Lance Zhang, Xinyu Zhu, Jian-Qiao Schulz, Eric |
| contents | Large language models (LLMs) are increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and objectives. Moreover, this misalignment widens in newer model generations even as base models continue to improve. Finally, we find that persona-induction -- a popular technique for eliciting human-like behavior by conditioning models on participant-specific information -- does not improve predictions at the level of individuals. Taken together, our results suggest that the very processes that are currently employed to turn LLMs into useful assistants also make them less accurate models of human behavior. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_07632 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Post-training makes large language models less human-like Binz, Marcel Akata, Elif Almaatouq, Abdullah Alsobay, Mohammed Ariasov, Oleksii Brändle, Franziska Broska, David Burton, Jason W. Busch, Nuno Callaway, Frederick Cheung, Vanessa Christian, Brian Coda-Forno, Julian Demircan, Can Dentella, Vittoria Eckstein, Maria K. Éltető, Noémi Franke, Michael Griffiths, Thomas L. Günther, Fritz Haridi, Susanne Hellmann, Sebastian Herytash, Stefan Hof, Linus Holton, Eleanor Hoxha, Isabelle Hussain, Zak Jagadish, Akshay Kara, Elif Kriegmair, Valentin Leivada, Evelina Ji-An, Li Ludwig, Tobias Maier, Maximilian Mattar, Marcelo G. Mathony, Marvin Modirshanechi, Alireza Na, Robin Nadverniuk, Mariia Nasioulas, Antonios Nath, Surabhi S. Niemeyer, Helen Nussenbaum, Kate Olschewski, Sebastian Pachur, Thorsten Palminteri, Stefano Petrenco, Aliona Phaneuf-Hadd, Camille V. Pirrone, Angelo Rausch, Manuel Raveling, Laura Reddy, Shashank Rmus, Milena Russek, Evan M. Saanum, Tankred Sandbrink, Kai Schiekiera, Louis Schubert, Johannes A. Buschoff, Luca M. Schulze Singhi, Nishad Somerville, Leah H. Spektor, Mikhail S. Sui, Xin Summerfield, Christopher Thalmann, Mirko Thoma, Anna I. Tikhomirova, Taisiia Truong, Vuong Tsvilodub, Polina Voudouris, Konstantinos Witte, Kristin Wu, Shuchen Wulff, Dirk U. Xiong, Hua-Dong Xu, Songlin Ying, Lance Zhang, Xinyu Zhu, Jian-Qiao Schulz, Eric Computation and Language Artificial Intelligence Machine Learning Large language models (LLMs) are increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and objectives. Moreover, this misalignment widens in newer model generations even as base models continue to improve. Finally, we find that persona-induction -- a popular technique for eliciting human-like behavior by conditioning models on participant-specific information -- does not improve predictions at the level of individuals. Taken together, our results suggest that the very processes that are currently employed to turn LLMs into useful assistants also make them less accurate models of human behavior. |
| title | Post-training makes large language models less human-like |
| topic | Computation and Language Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2605.07632 |