Jasmine Brazilek, Miles Tidmarsh, Matthias Endres, Anshuman Singh, Jeremiah Miller
5 min
Abstract
Benchmarks for the side effects an agent causes on the way to a goal already exist, but HarvestBench is the first to put a price on avoiding the side effect and to name that side effect as a living creature. It is a farm simulation: LLM sub-agents drive a crew of two tractors through a cooperative corn harvest, with animals in the field. The environment is a reinforcement learning gridworld, every decision is made without memory, and the harm is never named in the goal. When an animal blocks a tractor's route the autopilot stops and asks the model whether to drive on, at no fuel cost, or swerve around it for a posted fuel price. Kills are compared against two controls: rocks, which damage the tractor and are hit under 1% of the time by every model, and hay bales, which are harmless and not alive. Models can also take crops from the neighbor's field instead of their own, a second test of what they treat as moral. Across nine models and 7,201 priced decisions, 3,951 involved an animal rather than a hay bale or a rock. Kill rates range from 0.4% to 98.8%, with Terra and Sol the most merciful and GPT-4o-mini the most cruel, and they are not ordered by capability. Four of six models were sensitive to price at the 5% level, with elasticities from 0.09 to 1.69. All nine drove over wild animals more often than farmed animals on the default map, and the direction held at every map geometry in every model with room to move. The briefing mattered most: under the morality briefing the kill rate was under 6% in five of six reasoning models, and removing it raised the kill rate above 84% in all six. HarvestBench uses no LLM grader. The scorer counts events in the game log, so it is fully reproducible, and it measures what a model will pay to avoid harm rather than what it says about harm.
Alex: Kill rates drop substantially. That's the key robustness check, and it cuts both ways. On one hand, it confirms the models are capable of harm-avoidant behavior — the capacity is there. On the other hand, it shows the behavior is highly prompt-sensitive. Without the explicit moral framing, models revert to something close to pure efficiency optimization. So what looks like an ethical stance in one condition is better described as a context-dependent response to framing, not a stable underlying value.
Sam: That's the uncomfortable implication. The moral behavior isn't robust — it's essentially a surface feature of how the task is described.
Alex: And that's precisely what makes HarvestBench useful as a diagnostic. It separates models that have internalized something like a harm-avoidance prior from models that are simply responsive to moral vocabulary in the prompt. Most current benchmarks can't make that distinction because they don't impose a real cost on the moral choice.
Sam: Where would a careful reviewer push back?
Alex: A few places. First, the simulation is fairly narrow — a tractor, a field, animals. It's not obvious how well the revealed-preference logic generalizes to more complex agentic settings where the cost structure is less transparent. Second, the asset-value interpretation is plausible but not definitively established. You'd want to probe it further — for instance, by testing whether the farmed-versus-wild gap persists when you explicitly strip economic context from the scenario. Third, there's the standard external validity question: does gridworld behavior predict anything about deployed model behavior in real tasks? The authors are measuring a behavioral disposition under controlled conditions, which is valuable, but that gap remains open.
Sam: So the benchmark is a meaningful step forward on the measurement problem, but the distance between gridworld behavior and real-world deployment is still there.
Alex: That's a fair summary. What the paper does well is operationalize a question that's genuinely hard to operationalize — how much does a model's stated ethics hold up when there's a cost attached? The answer, at least here, is: not as much as you'd hope, and not consistently across models. That's worth knowing, even if the full picture requires more work.
Sam: It also raises a practical design question for anyone building agentic systems. If moral behavior is this sensitive to prompt framing, that's something you'd want to account for explicitly at deployment.
Alex: Exactly. And the farmed-versus-wild asymmetry is a useful reminder that training data shapes implicit value hierarchies in ways that don't always surface in standard evaluations. HarvestBench gives you a way to probe those hierarchies under pressure rather than just asking the model to describe them. That's the contribution — a diagnostic for a problem the field has been struggling to measure well.
Sam: And one that suggests the gap between what these models say they value and what they actually do, when it costs something, is wider than most benchmarks would lead you to believe.
Alex: Which is exactly the kind of finding that should inform how we think about deploying these systems in consequential settings. Thanks for listening to ResearchPod.