Tag
This paper introduces Taste-Bench, a benchmark for measuring taste in LLM agents' long-horizon decisions, finding that frontier models have low accuracy and that taste can be improved through distillation training.