Definition
Plain language
A simpler, configurable set of dungeon-game tasks built on top of NetHack for testing AI agents.
As stated in the literature
An RL sandbox built on the NetHack Learning Environment that exposes smaller, customizable task instances, used as a mid-difficulty long-horizon agent benchmark.
Why it matters: It gives a dial to tune task difficulty, making it easier to isolate and test specific agent skills between toy problems and the full, brutal game.
For example, a researcher can use MiniHack to set up a small dungeon room where an agent just has to find a key and open a door, without the full game's complexity.
Heard on the show
“MiniHack more than triples.”Episode 192 — A 32B Open Model Matched Frontier Systems By Learning to Take Notes