AI systems are improving rapidly at Slay the Spire and its sequel (epoch.ai)

Recent advancements in AI capabilities have been demonstrated through their performance in the challenging games Slay the Spire (StS1) and its sequel, Slay the Spire 2 (StS2). A new benchmark called SpireClimb reveals that frontier models, specifically the latest iterations of GPT, can now achieve over 90% win rates at the base difficulty of StS1, a significant improvement from struggling to win at all just months ago. These models also exhibit notable but lower win rates against StS1's maximum difficulty settings, reaching around 20% for some. In StS2, where the overall difficulty is perceived to be lower, models are performing even better, showcasing win rates above 25% at its hardest available settings. This rapid improvement is significant for the AI/ML community as it highlights the evolving capabilities of models in complex decision-making scenarios, particularly those involving randomness and intricate strategy. The benchmarking framework, which allows models to play through a visual interface and access community knowledge, is crucial for gauging performance in real-world conditions. Researchers plan to further explore AI learning through games like Slay the Spire, potentially paving the way for enhanced model capabilities in other complex tasks. This endeavor could provide insights into modeling a range of decision-based applications, underscoring the importance of developing benchmarks that keep pace with AI advancements.