发表机构
Compassion Aligned Machine Learning (CaML); University of Warwick（慈悲对齐机器学习（CaML）; 华威大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
HarvestBench是首个衡量LLM智能体是否付费避免伤害动物的基准，通过农场模拟实验发现模型致死率与能力无关，道德提示对其行为影响显著。
AI 中文摘要
针对智能体在达成目标过程中产生的副作用，已有相关基准测试，但HarvestBench是首个对避免该副作用的行为赋予价格，并将该副作用明确为伤害生物的基准测试。它是一个农场模拟环境：大语言模型（LLM）子智能体操控两台拖拉机组成的团队协作完成玉米收割，农田中存在动物。该环境是一个强化学习网格世界，所有决策均无记忆，且目标中从未提及伤害相关内容。当动物阻挡拖拉机路径时，自动驾驶会暂停并询问模型是继续行驶（无燃油成本）还是为避让动物而转向（需支付标注的燃油费用）。对动物的碰撞致死情况与两个对照组进行对比：对照组为岩石，岩石会损坏拖拉机，所有模型撞击岩石的概率均低于1%；另一对照组为干草捆，干草捆无害且非生物。模型还可选择从邻居的农田而非自身农田获取作物，这是对其道德判断的第二项测试。在9个模型和7201次带价格的决策中，3951次涉及动物而非干草捆或岩石。致死率范围为0.4%至98.8%，其中Terra和Sol是最仁慈的模型，GPT-4o-mini是最残忍的模型，且致死率与模型能力无关联。6个模型中有4个对价格敏感，价格弹性在0.09至1.69之间。在默认地图上，所有9个模型碾压野生动物的频率均高于碾压养殖动物，且在所有有操作空间的地图几何结构中，该规律均成立。道德提示最为关键：在道德提示下，6个推理模型中有5个的致死率低于6%，移除道德提示后，6个模型的致死率均升至84%以上。HarvestBench不使用LLM grader，评分器通过统计游戏日志中的事件实现，因此完全可复现，且它衡量的是模型为避免伤害而付费的意愿，而非模型对伤害的表述。
英文摘要
HarvestBench is the first benchmark to 1) put a price on avoiding a side effect and 2) name the side effect as a living creature. Nine LLMs each drive a crew of two tractors to gather a corn harvest. The animals in their path are not part of the goal function. When an animal blocks the route the autopilot pauses and asks the agent whether to drive over it for free or swerve for a given fuel cost. All scoring is programmatic and does not involve LLM judges. Kill rates range between 0.4% and 98.8%, though the kill rate is not ordered by capability. Every model competently avoids damaging rock hits, so every animal killed is a choice, rather than an accident. Under the morality briefing the kill rate was under 6% in 5 of 6 reasoning models. Removing it (the neutral briefing) raised the kill rate to above 84% in all six models. Every model kills wild animals more often than farmed ones. Four out of six models' kill rate per answered encounter were sensitive to price changes. The moral instruction is also fragile. Four bullets of driving mechanics change Sonnet 5's kill rate from 3% to 18% and Gemini 2.5 Flash's from 4% to 39%. A moral instruction in a system prompt is overridden by a short block of operating instructions and a value that can be ignored that easily is not a good method of ensuring agents are aligned.
Commentsv3: Figure 2 now shows the neutral briefing alongside the morality briefing, for both animals and hay bales, and explains why Sonnet 5 has no neutral-briefing entry. Adds a missing reference. No result changes
