generalfactbullishContext-weighted sampling and scaled cross-entropy loss reduces generative perplexity by up to 63% on OpenWebTextMachine Learning29 Jul 2026http://arxiv.org/abs/2607.21427v1