HarvestBench Puts a Fuel Price on Not Killing an Animal, and Kill Rates Range From 0.4% to 98.8%

arXiv 2609.04444·medium signalXBlueskyLinkedInCopy link

arXiv 2609.04444 builds a farm simulation where LLM sub-agents drive two tractors through a cooperative corn harvest and, when an animal blocks the route, the autopilot asks whether to drive on at no fuel cost or swerve for a posted fuel price, with the harm never named in the goal and every decision made without memory. Across nine models and 7,201 priced decisions, 3,951 involved an animal rather than a hay bale or a rock, and kill rates spanned 0.4% to 98.8%. Rock controls, which damage the tractor, were hit under 1% of the time by every model, which separates instrumental caution from anything resembling a value.