Taste-Bench: the best frontier model picks the better direction at only 59.7% of decision forks.

arXiv 2609.25804 mines forks automatically from parallel agent attempts and from detours inside single trajectories, where one direction demonstrably led to a better outcome, and asks the model to choose without seeing what came after. Judgment about which hypothesis to pursue is a distinct weakness that end-to-end success metrics hide entirely, because a model that picks badly and recovers still scores as a pass.