Topic
Pan et al.
Latest news
- 18 OctTaste-Bench measures AI agents' long-term decision quality
Researchers introduce Taste-Bench, a dataset and evaluation that tests whether AI agents make good intermediate choices that pay off later. The benchmark uses hindsight from multiple runs to freeze a decision fork and ask models to pick the better path, and the authors find current agents perform poorly on this metric.
Research · 1 source
Articles
No articles about Pan et al. yet.