Building AI safety benchmark environments on themes of universal human values
Proposes multi-objective, multi-agent gridworld benchmarks based on universal human values and homeostasis to test whether RL and LLM agents balance objectives and avoid misaligned maximisation.