Tag
HarvestBench introduces a benchmark that evaluates LLM agents in a farm simulation to measure their willingness to pay fuel to avoid killing animals, showing varied kill rates across models influenced by price and moral briefings.
This paper introduces MoralSim to evaluate how LLM agents behave in morally charged social dilemmas where ethical actions conflict with profit incentives, finding that no model remains consistently moral and cooperation rates vary widely.