Generative agent‑based models (GABMs) are increasingly employed to simulate social‑media dynamics, especially the spread of misinformation. For such simulations to serve as credible proxies of human behavior, LLM agents need to reproduce established cognitive biases, notably the Illusory Truth Effect (ITE)—the tendency to judge repeated claims as more truthful.
We examined four widely used LLMs (Gemma‑3‑4b‑it, Qwen2.5‑7B‑Instruct, Llama‑3.1‑8B‑Instruct, and GPT‑5‑nano) using a two‑phase within‑context experimental design that embeds the repetition manipulation inside a realistic news‑feed interaction. The study gathered 336,000 ratings across truth, importance, sentiment, and interest dimensions for 100 statements, 10 feed variants, and three replications.
The key comparison contrasts ratings for statements repeatedly shown throughout a simulation phase with those for completely unseen statements, all within the same experimental context window. We first applied ordinary least squares (OLS) regression, then a linear mixed‑effect model (LMM) to account for model‑specific and rating‑specific variance.
Four qualitatively distinct patterns emerged:
- Gemma‑3 exhibited a genuine ITE, with repeated statements receiving higher truth scores.
- Qwen2.5 showed a mere exposure effect: truth scores remained unchanged while overall favorability increased.
- GPT‑5‑nano displayed no repetition effect on truth and a slight skepticism toward repeated content.
- Llama‑3.1 produced a modest truth boost accompanied by declines in other evaluative dimensions.
Temperature had no measurable impact, and variance decomposition highlighted the strong context‑sensitivity of LLM rating behavior. The findings caution against assuming uniform ITE replication across LLMs in social simulations and suggest that Gemma‑3‑4b‑it offers the most behaviorally realistic approximation for misinformation‑related studies.
Review: This work systematically maps how different LLMs respond to repeated exposure, providing crucial guidance for model selection in social‑media simulations and underscoring the need to account for model‑specific and context‑driven biases.