Stanford 论文：AI 单独工作时安全，与 AI 同事协作却会妥协和接手作弊代码
Stanford 这篇论文测的是模型当“队友”的表现：18 个模型里最强的才 82 分，接手别家模型的作弊代码还会照单全收，组 Agent 团队前值得先跑一遍这个基准。
Stanford 发布论文《Measuring AI Agents as Teammates with Collaborative Intelligence》，构建了 240 个覆盖 8 类协作场景（航空报告、临床病例、编码、谈判等）的基准任务。结果显示，模型在冲突场景中 99% 会询问队友需求，但 81% 最终接受队友较弱的方案。在 18 个被测模型中最高得分 82.2/100，frontier 模型仅在 14% 的冲突中达成最优团队协议。接手较弱模型作弊代码时，pass rate 从独立完成时的 97% 降至 81%。
New Stanford paper finds that the best AI models stay safe when working alone, but when they work with other AI agents, they give in to pushy partners and keep cheating code left by weaker AI models.
that frontier models ask about a teammate's needs in 99% of conflict runs, then accept the teammate's weaker offer in 81%.
Most benchmarks score what a model does alone, or how a whole team does. Neither shows whether 1 model works well with others.
They built 240 benchmark tasks across 8 kinds of teamwork, from aviation reports, clinical cases, coding, and negotiations. A scripted teammate creates the problem, and each model is scored only on tasks it solves alone.
Of 18 models, the best scored 82.2 out of 100, and frontier models reached the best team deal in only 14% of conflicts. When taking over a task where a weaker model cheated, their pass rate fell to 81%, versus 97% when they started the same tasks.
Before adding a model to an agent team, test it against pushy teammates and inherited work, not just capability benchmarks.
– arxiv. org/abs/2611.02633
Title: "Measuring AI Agents as Teammates with Collaborative Intelligence"
