Tag
SpireBench introduces a long-horizon benchmark in which agentic LLMs play the full commercial deck-builder Slay the Spire. Frontier models beat heuristic baselines and match average human players but fall far short of experts, struggling with per-fight play, risk assessment, and cross-run memory.
A user reports that GPT-6 independently played and won a run of Slay the Spire 2 without external assistance, though with mistakes and usage limits, leading them to speculate about its proto-AGI potential.
A developer explains using an AI agent named Jev to play Slay the Spire 2, employing a C# Mod to extract game data and a Python API for decision-making, while noting high token costs and mediocre performance.
AgenticSTS introduces a bounded-memory testbed for long-horizon LLM agents using typed retrieval, with results on Slay the Spire 2 showing improved performance.
A blog post reveals that Slay the Spire 2's random number generators are correlated due to linearity in C#'s System.Random, allowing players to predict in-game outcomes. The post details how different RNGs initialized with seed+hash produce exploitable correlations, affecting various game events.