AI agents have bad taste in long-term decisions and a new benchmark proves it

Why it matters

Microsoft and City University of Hong Kong researchers have introduced Taste-Bench, a benchmark that measures whether AI agents can choose well at long-term decision points during extended tasks. The researchers frame this judgment as an agent's "taste": the ability to pick the better path when a poor choice only reveals its cost much later. Existing benchmarks check whether a task gets finished, not the quality of intermediate decisions, and human grading of such choices would be costly and hard to scale. Taste-Bench instead uses hindsight, building questions from recorded agent runs that branched at a fork, then hiding later work and asking a model to pick between the two directions. The team evaluated models from several major families and found that even the strongest performers struggled, especially when the deciding evidence only appears after substantial further work. Notably, the researchers also demonstrated that this judgment can be taught by distilling a teacher model's knowledge into a student, which improved both benchmark accuracy and downstream task success. The benchmark and code are publicly available.