general

LLMs Play Civ V, Develop Distinct Personalities

Researchers demonstrate that large language models playing Civilization V develop unique strategic personalities and decision-making patterns, revealing

Someone got LLMs to play full Civilization V games and found they develop totally different personalities based on the model.

They ran 1,408 games using a hybrid setup where the LLM sets strategy and Civ V’s built-in AI executes it. Here’s what happened:

OSS-120B went full warmonger:

  • +31.5% more Domination victories
  • -23% fewer Cultural wins

GLM-4.6 played more balanced between military and culture.

Both models weirdly preferred the Order ideology (~24% more) over Freedom, basically choosing communist-style government every time.

The breakthrough is that pure LLM approaches couldn’t even survive a full game before, but this hybrid method gets ~97.5% survival rate with models as small as 20B parameters.

Cost per game: ~$0.86 using OpenRouter pricing, burning through ~53k input tokens and 1.5k output tokens per turn.

Full writeup comparing the playstyles: Check the post with replay GIFs showing how different the strategies actually look in practice.