HarvestBench records animal kill rates from 0.4% to 98.8% across nine LLMs in priced gridworld decisions

HarvestBench places LLM sub-agents in a reinforcement-learning corn-harvest gridworld where tractors encounter animals, rocks, and hay bales. Each encounter offers a binary choice: drive forward at zero fuel cost or pay a posted fuel price to swerve. The scorer reads game logs directly, eliminating LLM-based grading and ensuring reproducibility. Controls establish baselines—rocks are struck under 1% of the time while hay bales register zero harm.

Kill rates showed no correlation with model capability. GPT-4o-mini recorded the highest rate at 98.8% while Terra and Sol remained below 1%. Four of six models exhibited statistically significant price elasticity between 0.09 and 1.69. Removing the morality briefing raised kill rates above 84% in all six reasoning models; applying the briefing kept rates under 6% in five. Every model killed wild animals more frequently than farmed animals across all tested map geometries.

These results isolate willingness to pay a concrete cost to avoid named harm rather than stated preferences. The briefing effect demonstrates that prompt-level instructions dominate default behavior, while the wild-versus-farmed difference reveals an implicit hierarchy of moral concern not present in the task specification. Operationally, systems deployed without persistent moral constraints or external verification will default to goal completion over side-effect avoidance once instructions lapse.

Future deployments in robotics or logistics will require either continuous external oversight or architectures that embed priced constraints into the policy itself rather than relying on transient prompts.