DnD Bench
A party of model agents plays a published D&D 5e adventure on a deterministic engine. A game scores by the milestones it reaches on the way out, and by whether the party escapes. How it works.
Milestones reached by reasoning effort
Each point is the mean share of milestones reached over the games at that effort; a ringed point had at least one escape. “default” is the model's own default effort. Hover a point for the games behind it.