Approaches to Slay the Spire AI: rules, search, RL and LLMs

Rules & heuristics

Hand-written priorities and conditions: which cards to take, which events to accept, when to rest.

Works for

Fast, explainable, and needs no data. Bottled AI wins 20–52% depending on character with rules plus combat search.

Watch out for

Every interaction has to be written by hand, and a list tuned for one deck plays badly with another.

Projects (5)

Trying sequences of actions in a model of the game and keeping the best outcome, from exhaustive hand orderings to MCTS.

Works for

Combat. Every strong bot in this index searches during fights.

Watch out for

Needs a fast, accurate simulator. Knowing the real RNG makes search look stronger than it is.

Projects (12)

Policies trained from game outcomes, usually inside a simulator.

Works for

Out-of-combat choices on top of combat search, where it beats hand-written heuristics.

Watch out for

Sparse, delayed reward over 50-plus floors. Learning combat directly has mostly failed so far.

Projects (8)

Papers and write-ups (1)

Models trained to copy decisions from human runs or from a stronger bot.

Works for

Bootstrapping a policy, distilling a slow search into a fast network, and pick or damage advisors.

Watch out for

Human data inherits human mistakes; bot data caps the student at the teacher.

Projects (5)

Papers and write-ups (1)

Language models given the game state as text and asked for the next action, sometimes with tools, memory or fine-tuning.

Works for

Quick to start, readable reasoning, works across both games with the same bridge.

Watch out for

Slow and expensive per decision, weak at long-horizon planning without help, and results vary between models.

Projects (11)

Papers and write-ups (7)