Digital Shadows of Mental States: Exploring the Behaviour of Large Language Models with Representation Engineering

Large Language Models (LLMs) have emerged as powerful tools at the intersection of artificial intelligence and cognitive science, demonstrating remarkable capabilities in language understanding, generation, and human-like cognitive behaviours. Recent advances in LLM technology have opened new possibilities for modeling complex psychological phenomena, including mental health conditions such as depression. This study investigates the capacity of LLMs to simulate depressive cognitive states and their responses to psychological interventions. Specifically, we examine how representation engineering techniques can be used to model and modify psychological states in LLMs, focusing on the differential effects of intrinsic versus extrinsic goal interventions based on self-determination theory. We employed a three-stage methodology: (1) fine-tuning a Mistral-7B-Instruct model on a curated dataset derived from the Psych 8K corpus to create a “depressed” baseline state; (2) applying representation engineering techniques to generate and implement psychological intervention vectors based on intrinsic and extrinsic goal frameworks; and (3) evaluating the model’s responses using quantitative mood assessments and behavioral analysis. The study suggested observable differences between intrinsic and extrinsic goal interventions on the LLM’s simulated psychological state. Intrinsic goal interventions (self-acceptance, personal growth, community feeling) appeared to show greater positive trends on the model’s reported mood and cognitive patterns compared to extrinsic goal interventions (financial success, social recognition, appealing appearance). Community and engagement emerged among the highest-performing interventions in the tested approaches. These findings contribute to the growing body of work suggesting LLMs may serve as computational models for psychological research, showing patterns that align with established psychological principles such as the differential impact of intrinsic versus extrinsic motivations on well-being. The results support the use of LLMs as cost-effective tools for piloting psychological experiments and exploring therapeutic mechanisms before resource-intensive human trials.