HOST: Have you ever taken a sensible-looking route and only later realised it was a detour?
EXPERT: That is the problem behind Taste-Bench. It pauses an AI agent at a real choice and asks which next step is better.
HOST: What sort of choice?
EXPERT: In a coding task, should the agent run a small diagnostic test or rewrite a big section immediately? The later task record tells us which route worked
HOST: How hard is the test?
EXPERT: The authors made 502 questions. Their best tested model reached 59.7 percent on the benchmark’s scoring rule.
HOST: Does that score mean the model has good judgment everywhere?
EXPERT: No. These examples mainly come from software and research tasks, and the labels rely on recorded outcomes.
HOST: What can teams do with the idea?
EXPERT: Save key decisions and later outcomes. Then review where an agent made a costly detour before giving it more autonomy.