20个大模型挑战《杀戮尖塔》A20，仅GPT-5.6 Sol通关一次

Canada's Vector Institute and Radboud University in the Netherlands released SpireBench on arXiv, testing 20 large language models on Slay the Spire at A20 difficulty, with only GPT-5.6 Sol clearing the Heart once across 200 runs. The study found the AIs play too conservatively, fighting about 0.9–1.4 elites in Act 1 versus roughly 2.6 for top streamers and resting at campfires 57%–75% of the time versus about 9% for experts. The models also rarely skip card rewards, leaving their decks bloated.