HarvestBench Evaluates LLM Agent Moral Costs and Side Effects
September 7, 2026
HarvestBench uses a farm simulation gridworld to measure if LLM agents will incur fuel costs to avoid harming animals. The benchmark compares agent decisions to controls like rocks and hay bales to quantify the trade-off between goal efficiency and moral side effects.
HOW THIS AFFECTS YOU
●
researcherYou can use this benchmark to study the alignment of agentic decision-making with unintended environmental side effects.
●
policyThis provides a quantitative way to measure the ethical risks of autonomous agents in real-world simulations.