Disrupting convergence in large language model generation

Abstract Efforts to reduce homogenization in large language model (LLM) outputs increasingly use prompting interventions, but these interventions often change both the information supplied to models and the instructions governing its use. We experimentally separate this behavioral guidance from a second potential source of variation: the information available during generation. Across 16,200 generations spanning three preregistered studies and two cross-model robustness replications, supplying heterogeneous external cues substantially increased semantic diversity and generally did so more reliably than explicit instructions to produce different responses. Crucially, this effect did not increase monotonically with stronger instructions to use the information. Merely presenting varied cues produced substantial diversification, whereas instructing models to disregard the same information largely eliminated the effect. When cue identities and cue-use instructions were held constant, differentiation instructions increased collective diversity relative to the other experimental conditions and increased departure from models’ default response spaces, whereas convergence instructions reduced such departure. These findings distinguish two sources of variation often conflated in prompt design: informational inputs can supply variation, while accompanying instructions can suppress their influence or further shape the resulting response distribution. Preserving diversity may therefore require designing information environments, not simply better prompts.