CoolingResearch
Taste-Bench tests whether agents pick the better next step at decision forks
A paper, "The Tasteful Agent" (arXiv 2609.25804), releases Taste-Bench: 502 items, 390 engineering and 112 research, each a frozen agent trajectory stopped at a fork with two candidate next steps. Labels come from the later work's outcome, using SWE-bench runs and METR transcripts. The dataset is on Hugging Face.
- HEAT
- OUTLETS
- 1
- FIRST SEEN
- LAST SIGNAL
Primary source
Hugging Facehuggingface.co/datasets/wenbopan/taste-benchCoverage · 2 articles, oldest first
Sunday, October 18, 2026
Hugging FacePrimary
Taste-Bench tests whether agents pick the better next step at decision forks