新基准 Taste-Bench 揭示 AI 智能体长期决策品味差，最佳模型仅 59.7%
原标题：AI agents have bad taste in long-term decisions and a new benchmark proves it
AI 摘要
微软与香港城市大学的研究者提出 Taste-Bench，用「决策分叉」方法评测 AI 智能体在长任务中做长期决策的能力（即「品味」）。测试 14 个模型，最好的 GPT-5.6 Sol 仅 59.7%，接近随机猜测；当关键线索只在后续工作中出现时，准确率跌至 21%。研究者还通过蒸馏教师模型判断力，使学生模型在未见任务上从 30% 提升到 47.9%，并在 SWE-bench Pro 上把任务成功率从 14.6% 提升到 33.7%。
正文节选
AI agents have bad taste in long-term decisions and a new benchmark proves it Researchers from Microsoft and City University of Hong Kong built a method to test how well AI agents make decisions during long-running tasks. The best model barely hit 60 percent. AI agents now take on multi-step tasks. They might write a research paper or build out a software project. Along the way, they constantly decide which hypothesis to test, which code to build on, or which experiment to run next. Those choice