One of the paper's authors announced its acceptance to NeurIPS 2026 and explained XTC, short for Exclude Top Choices. When at least two possible next tokens clear a probability threshold, the method sometimes removes all but the least likely of those tokens, then rescales the remaining probabilities, the author says. They report more diverse, less repetitive creative output across model families, with human raters preferring XTC for creativity and no loss in fluency. The author also reports far less harm to instruction following than temperature at the same diversity, and says XTC was already built into llama.cpp, ExLlamaV2 and text-generation-webui.