generalfactbullishSmall fine-tuned long-context language models can outperform strong baselines in HTML extraction tasksComputation and Language28 Jul 2026http://arxiv.org/abs/2607.21284v1