Words Matter: How Prompt Framing Shapes Strategic Behaviour in Large Language Models

Abstract

This study investigates how linguistic framing influences the strategic behaviour of large language models in repeated interactions. Four models (Mistral, Qwen3, Llama3.2, Llama3.3) were embedded as autonomous agents in a simulation of a 25-round iterated prisoner's dilemma under three prompt conditions: neutral, positively biased, and hunger-framed. Although payoff structures remained constant, linguistic variation produced substantial behavioural divergence. A one-way analysis of variance showed significant prompt effects in 13 out of 16 model pairings (adjusted p < 0.05). Positively biased prompts increased cooperation by 4–9 percentage points, while survival-framed prompts increased cooperation up to 80 percentage points. While Qwen3 and Llama3.3 were highly sensitive to framing, Llama3.2 showed minimal responsiveness. Several models exhibited emergent strategies such as conditional cooperation and end-game defection. These findings indicate that subtle linguistic cues can systematically modulate cooperative behaviour in large language model agents.