How Surprising Is Historical Italian to Language Models? Tokenization and Comprehension Taxes

arXiv·low signal

This arXiv paper (2026-06-25) decomposes the difficulty LLMs have with historical language into distinct dimensions — tokenization cost, predictive surprisal, semantic robustness, and pretraining exposure — using historical Italian as a case study, and proposes a simple mitigation. It quantifies a 'tokenization tax' and 'comprehension tax' rather than treating historical difficulty as a single barrier. The diagnostic framework is relevant for builders working on digital-library, OCR, and low-exposure-domain LLM workflows.