Projects / Bot / agent

Slay the Spire 2

AgenticSTS

Research testbed for long-horizon language-model agents on Slay the Spire 2, built around bounded, typed memory instead of a growing transcript. Reports 3/10 wins at Ascension 0 without learned skills and 6/10 with them.

View on GitHub

Context reaches the model through named slots (rules, episodes, skills) that can each be switched off, which makes ablations straightforward. The project releases 298 trajectories as a Hugging Face dataset. The paper is listed under Papers. The sample is small, which the authors acknowledge.