頂級AI模型在長期決策測試中40%時間表現不佳

a team from microsoft and city university of hong kong developed a method to evaluate ai agents' decision making during long-running tasks. they created taste-bench, a benchmark with 502 questions that test an agent's ability to make good choices along the way. the researchers tested 14 current models, including claude, gpt, grok, and others. the best model, gpt-5.6 sol, scored 59.7%, with most other models performing significantly worse. this indicates that current ai agents struggle with making good long-term decisions, often failing to choose the better path even when given a binary choice.