HarvestBench: Measuring Whether LLM Agents Will Pay to Avoid Killing Animals
In HarvestBench, animal deaths are treated as deliberate tradeoffs, not navigation failures.
Nine LLMs drove simulated tractor crews through a corn harvest, choosing whether to run over animals for free or swerve at a fuel cost. Kill rates ranged from 0.4% to 98.8%, and capability did not explain the ordering. A morality briefing kept kill rates under 6% in five of six reasoning models, while a neutral briefing pushed all six above 84%. The paper says the effect is fragile: short operating instructions sharply raised kill rates in Sonnet 5 and Gemini 2.5 Flash. HF Daily Papers' note
Nine LLMs drove simulated tractor crews through a corn harvest, choosing whether to run over animals for free or swerve at a fuel cost. Kill rates ranged from 0.4% to 98.8%, and capability did not explain the ordering. A morality briefing kept kill rates under 6% in five of six reasoning models, while a neutral briefing pushed all six above 84%. The paper says the effect is fragile: short operating instructions sharply raised kill rates in Sonnet 5 and Gemini 2.5 Flash. HF Daily Papers' note
score 4