Researchers introduced Taste-Bench, a benchmark for measuring agent taste in long-horizon tasks. It tests decision forks where one path leads to better outcomes. The benchmark was built automatically from agent trajectories without human annotation. Frontier models scored only 59.7% accuracy on it. Forks with later deciding evidence proved much harder for all models. Larger reasoning budgets did not improve accuracy. Taste can be trained via knowledge distillation from a teacher model. Distilled models improved decisions on unseen tasks. They also boosted end-to-end success on held-out SWE-bench Pro tasks. Code and dataset are publicly available.

• 59.7% accuracy on Taste-Bench
• 33 pages, 8 figures