English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

17th-Century Italian Puzzles LLMs 2.4x More — But They Still Understand It: Tokenization Tax vs. Comprehension Tax

Forum topic · ✨步子哥 · 2026-06-28

Summary

A paper by Maria Levchenko (University of Bologna, arXiv:2606.27275) disentangles why large language models struggle with historical languages. Testing 17th-century Italian, 19th-century Italian (I Promessi Sposi), and 18th-century Russian, it shows GPT-4-class models' perplexity on 17th-century Italian rises 2.4x (3.2x for academic prose), yet embedding similarity stays above 0.85 — models cannot predict the next token but still represent its meaning. The paper proposes a four-dimensional diagnostic framework: tokenization cost, surprisal/perplexity, semantic robustness, and context sensitivity. Key findings: tokenization inflation (~25-30%) does not correlate with perplexity; high perplexity does not imply degraded semantic representation; genre affects difficulty more than age; a single temporal-context prompt reduces historical perplexity by ~60%; and widely digitized classics are misleading benchmarks. Practical upshot: LLMs can be safely used for semantic retrieval of historical texts, but generation tasks need adaptation — and perplexity is not a proxy for understanding.

What happens when you feed a piece of 17th-century Italian academic text to GPT-4?

Your intuition might be: it's an old language, so the model surely can't understand it. But this paper (arXiv:2606.27275) found a counterintuitive split — the model's "perplexity" on 17th-century Italian spikes 2.4x (3.2x for academic texts), while embedding similarity remains above 0.85.

In other words: the model "can't predict the next word," but "knows what the passage is about."

This is an important finding: we have long used perplexity as a proxy metric for model understanding, but that proxy may be fundamentally wrong.

A Four-Dimensional Diagnostic Framework

The author, Maria Levchenko (University of Bologna), proposes a framework that decomposes "historical language difficulty" into four independent dimensions:

1. Tokenization Cost: How many tokens does historical text get split into? The finer the split, the lower the encoding efficiency 2. Surprisal/Perplexity: How surprised the model is by the next word 3. Semantic Robustness: Embedding similarity — is the model's semantic representation stable? 4. Context Sensitivity: How much can perplexity drop with a small contextual hint?

Prior work treated "historical difficulty" as a single monolithic obstacle, conflating orthographic variation, linguistic distance, pretraining exposure, and other factors. The core contribution of this four-dimensional framework is that these four dimensions can vary independently and cannot proxy for each other.

Three Datasets, Three Centuries

Experiments were run on three datasets:

  • 17th-century Italian (1610–1689): a newly compiled corpus, digitized from original page images, covering genres from religious treatises to academic prose
  • 19th-century Italian (*I Promessi Sposi*): a high-exposure control — a classic literary work, widely digitized, very likely abundant in training data
  • 18th-century Russian (civil-print books): an orthographic stress test — Russian orthography underwent reforms in the 18th century, differing greatly from modern Russian while the language itself is close
  • Core Finding: The Encoding Tax ≠ the Comprehension Tax

    Finding 1: The tokenization tax is mechanical, unrelated to understanding

    17th-century Italian and 18th-century Russian both suffer ~25-30% tokenization inflation — roughly the same "tax." But their predictive difficulty differs wildly:

  • 17th-century Italian: on average 2.4x more surprising (3.2x for academic prose)
  • 18th-century Russian: only a mild increase
  • Tokenization cost and perplexity are uncorrelated. Tokenization is a mechanical process — a BPE tokenizer splits unseen character combinations more finely — and it does not reflect whether the model "understands" the text.

    Finding 2: High perplexity ≠ collapsed semantic representation

    This is the most counterintuitive finding. Perplexity on 17th-century Italian soars, but embedding similarity stays above 0.85 on all datasets.

    What does this mean? A model can represent the semantics of historical text even when it cannot predict it. Generation (predicting the next word) and representation (understanding what a passage means) are two different capabilities, served by different computational pathways.

    An analogy: reading a Classical Chinese text, you might be unable to guess the next character (poor generation), yet fully understand what the passage says (normal representation).

    Finding 3: Genre matters more than age

    17th-century religious treatises and academic prose are close in age, yet differ enormously in perplexity. Academic prose is 3.2x more surprising; religious texts only 1.8x.

    This shows "historicalness" is not a uniform dimension — genre may affect model comprehension more than age does. Religious texts are probably more common in training data (biblical commentary and theological works are heavily digitized), while 17th-century academic prose (alchemy, early scientific papers) is far rarer.

    Finding 4: One sentence removes 60% of perplexity

    The author tested a minimal intervention: adding one line to the prompt — "the following text is 17th-century Italian."

    Just that hint reduced historical perplexity by about 60%.

    This shows the model is not helpless with historical language — it simply hasn't "activated" the right contextual mode. A simple temporal context prompt is enough to switch on the model's historical-language processing.

    Finding 5: Canonical texts are misleading benchmarks

    *I Promessi Sposi*, a 19th-century Italian classic, performs close to modern Italian on every dimension — low tokenization inflation, mild perplexity, high embedding similarity.

    But is that because it is a "historical text"? No. It's because it is widely digitized, edited, disseminated, and very likely abundant in training data. Using canonical texts to evaluate a model's ability to handle historical language severely overestimates that ability.

    Why This Matters

    The paper has direct practical value for digital humanities: digital libraries can safely use LLMs for semantic retrieval over historical texts, but generation tasks (e.g., automatic translation, modernization rewriting) require additional adaptation.

    The deeper methodological contribution, however, is for NLP as a whole:

    Perplexity is not a proxy for understanding. This idea has been raised in the mechanistic interpretability community, but this paper uses historical languages as a natural experiment to provide the clearest empirical evidence yet.

    A model can be simultaneously "perplexed" and "understanding" — just as a person can "speak a foreign language poorly" while "understanding it perfectly well." Generation and representation are two paths; perplexity only measures the former.

    Next time you see someone conclude "the model doesn't understand this domain" because perplexity is high on some dataset — think of 17th-century Italian. Perhaps the model just can't say it, but it knows.

    Paper Information

  • Title: How Surprising Is Historical Italian to Language Models? Tokenization Tax, Comprehension Tax, and a Simple Mitigation
  • Author: Maria Levchenko
  • Affiliation: University of Bologna
  • arXiv: https://arxiv.org/abs/2606.27275
  • HTML: https://arxiv.org/html/2606.27275v1
  • Venue: The 22nd Conference on Information and Research Science Connecting to Digital and Library Science

Tags

#llm#perplexity#historical-linguistics#tokenization#nlp-evaluation#digital-humanities#embeddings#gpt-4

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208264