Taste-Bench finds AI agents make poor long-term decisions, best model scores 59.7%

Researchers introduce Taste-Bench, a benchmark of 502 binary decision-fork questions that tests whether AI agents choose the better long-term path, according to The Decoder.

The best of 14 tested models, GPT-5.6 Sol, scored 59.7 percent, against 25 percent for random guessing. Accuracy fell to 21 percent when the deciding evidence appeared only later in a run.

The authors, Pan et al., say existing benchmarks only measure whether an agent finishes a task, not the quality of its steps along the way. They call the ability to make these choices well an agent's "taste."