HarvestBench Evaluates LLM Agent Willingness to Trade Cost for Animal Life
September 2, 2026
HarvestBench uses a reinforcement learning gridworld farm simulation to measure if LLM agents prioritize fuel savings over avoiding harm to living creatures. The benchmark introduces a decision point where agents must choose between incurring fuel costs to swerve around animals or driving through them to optimize for efficiency.
HOW THIS AFFECTS YOU
●
researcherYou can use this to study the emergence of cost-benefit trade-offs in agentic decision-making.
●
policyThis provides a framework for quantifying ethical alignment and side-effect costs in autonomous systems.