Research, theses, conference talks and long-form posts, newest first. Each takeaway summarizes the source's own abstract or text.
Gives a Slay the Spire 2 agent typed, bounded memory instead of a growing transcript. Ascension 0 wins go from 3/10 to 6/10 with a learned skill layer, and the authors note the sample is small.
Measures the entropy of map paths across 20,000 runs. Winning runs take higher-entropy (riskier) paths, and stronger players change how much risk they take late in a run.
Twelve-game benchmark for language-model agents with a shared game interface and fine-tuning data. The project's repository lists Slay the Spire among its games.
Releases a dataset of Slay the Spire card pairs labeled as positive, negative or neutral synergies. Models spot non-synergies well but miss positive and especially negative ones, mostly on timing and rule errors.
A pixel-only agent builds a graph of game states and actions from experience to plan over long horizons without an API. Slay the Spire is one of its evaluation games.
Swaps an old rules-based skeleton's heuristics for a small language model. Per the coverage, the bot reached the second of three bosses, and long-term planning was its weak point.
Published at FDG 2024. Language models play a simplified version of Slay the Spire working only from card text, compared against other agent types.
Language-model agents play Wordle and Slay the Spire below the average human, but their results correlate strongly with the difficulty human players report, which makes them useful for testing difficulty.
Write-up of an unfinished reinforcement-learning attempt on spirecomm and decapitate-the-spire, with practical notes on action masking and action-space design.
Bachelor thesis (Malmö University) training a neural network to predict successful map paths.
Works through the 2020 slice (18 million runs, 90 GB of JSON) of a 75-million-run dump Mega Crit shared. Of those 18 million runs, 9% were wins, a population baseline for any bot.
GDC talk on how Mega Crit balanced the game with metrics from every player decision. Background on why so much run data exists.