Aron Vallinder, Edward Hughes · 2024 · arXiv; AAMAS 2025 extended abstract (per listing)
Across generations of a donor game, whether cooperative norms evolve depends heavily on the base model: Claude 3.5 Sonnet populations built far more cooperation than Gemini 1.5 Flash or GPT-4o, and only Claude used costly punishment effectively. Outcomes were sensitive to the random seed.