PUBLISHED · NPJ ARTIFICIAL INTELLIGENCE · 2026
PsychAdapter: adapting LLMs to reflect traits, personality, and mental health
Huy Vu, Huy Anh Nguyen, Adithya V. Ganesan, Swanie Juhng, Oscar N. E. Kjell, João Sedoc, Margaret L. Kern, Ryan L. Boyd, Lyle Ungar, H. Andrew Schwartz, Johannes C. Eichstaedt
Finding:
PsychAdapter is an architectural modification to auto-regressive transformers that conditions generation on continuous psychological trait scores rather than on prompt text. A learned projection matrix at each transformer layer expands a low-dimensional trait vector, five values for the Big Five, to the width of the model's hidden state, so the psychological input influences generation at every layer. The added parameters are negligible: 0.07% of GPT-2 Large, 0.004% of LLaMA-3.
Expert raters with psychology training matched generated text to its intended Big Five level with 87.3% accuracy, rising to 91.0% when generation was prompted with a short lead-in. For depression and life satisfaction, expert accuracy was 96.7%. The approach transferred across Gemma-2B, GPT-2 Large and LLaMA3-8B with comparable results, and produced coherent combinations of traits rather than only single-dimension extremes.
Relevance:
Co-authored by Ryan Boyd of the University of Texas at Dallas, academic advisor to Receptiviti. The relevance to the measurement layer is structural rather than incidental. PsychAdapter demonstrates that psychological state can enter a language model as a typed, continuous input at every layer, in the same form in which Receptiviti's API returns it, without consuming context window or depending on prompt phrasing. The trait scores it conditions on are derived from empirical language-to-trait models built in the same research tradition as LIWC. A system that can accept a psychological state vector as input needs a validated source for that vector.