← Glossary Term
Taste-Bench
A benchmark that tests whether an AI agent picks the better direction at key decision points in a long task.
Taste-Bench is a test made by researchers at Microsoft and City University of Hong Kong. Instead of asking whether an agent finished a task, it freezes an agent partway through and asks a model to choose between two possible next directions, then checks the choice against what actually worked out later.
The researchers call this ability “taste”, the judgment about which path is likely to pay off. Scores so far are modest, which suggests that choosing well is a separate skill from finishing a job.
Mentioned in