Intel / article

SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire

Source: arxiv.org
Author: Scott C. Lowe, Sander W. Keemink
Date: Published Oct 12, 2026

Context

SpireBench tests whether strong immediate decisions add up to a successful long-term strategy. It puts language models through complete Slay the Spire runs, where players fight with cards, improve their deck and choose routes. The strongest models won the first act about as often as expert players, then fell behind in later acts.

The authors report that models tended to preserve health rather than spend it to grow stronger. For example, they rested at 31% of the rest sites they entered with at least 70% health, against 2% for the study's reference population, which upgraded a card instead. Among the models' own runs, upgrading rather than resting was associated with clearing Act 2, a correlation that does not show cause. Separately, what distinguished strong from weak models was fight quality: the top four took 8.2 damage per Act 1 hallway fight against 14.0 for the bottom tier. The study tested 20 models on one character and ten shared game seeds; human reference runs used different seeds.

Key takeaways · AI-distilled

The authors report the gap to expert players widens as a run goes on: the strongest models win the first act about as often as experts, then fall behind.
Per the paper, what separates strong models from weak ones is how well they play individual fights, while all models misjudge risk by preserving health instead of spending it to grow stronger.
A consistent renaming of every game term showed that part of the models' performance rests on prior knowledge of Slay the Spire, according to the authors.
Memories the models wrote across runs did not improve later play; the authors describe those notes as often wrong and overly cautious.

Read the source arxiv.org