Mechanistic control of large language models as simulated participants via linear representation

Large Language Models (LLMs) have garnered significant attention within psychology due to their potential to simulate human participants. However, recent studies often rely on simple prompt engineering that elicits behavioral patterns without directly manipulating internal model states. We introduce a novel simulation method, hypothesizing that high-level psychological traits correspond to specific directional vectors within the LLM’s activation space.

We extract activation-space vectors corresponding to 18 clinically relevant cognitive schemas, known as early maladaptive schemas (EMS) in schema therapy. Our experiments demonstrate that projection onto these vectors is associated with externally evaluated EMS expression in LLM outputs. Linear manipulation of these vectors further induced EMS expression.

Our work offers a mechanistically informed complement to prompt-based simulation and represents a step toward building more stable, controllable, and interpretable simulated participants. This work was supported in part by the grants from National Science and Technology Major Project (No. 2023ZD0121104), and the Anhui Natural Science Foundation (No. 2508085ZD006). School of Computer Science and Technology, University of Science and Technology of China, Hefei, Anhui, China Ruikang Zhang, Tong Xu, Derong Xu, Sirui Zhao & En-Hong Chen Department of Psychology, University of Science and Technology of China, Hefei, Anhui, China Correspondence to Tong Xu or En-Hong Chen.

The authors declare no competing interests. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material.

You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. Zhang, R., Xu, T., Xu, D. et al. Mechanistic control of large language models as simulated participants via linear representation. npj Artif.