On 16 July 2026, researchers posted an arXiv preprint describing PAT, a RAG-based system that feeds whole-document context to an LLM to translate U.S. English into Latin American Spanish at the document level, not sentence by sentence [S1]. The six translations they evaluated show the approach can push LLMs toward genuine reformulation and reshape how a text is organised and argued to fit Spanish-language conventions. But the results were inconsistent, and the study is small enough that whether this works at scale is still an open question.

The sentence-by-sentence problem

Translation tools, from computer-assisted translation (CAT) software to machine translation (MT) engines, overwhelmingly treat translation as a sentence-by-sentence act [S1]. Each sentence goes in, each sentence comes out. The problem: languages differ in more than vocabulary. Spanish-language discourse follows different conventions for how arguments are structured and how language is used, and those differences show up at the document level, not the sentence level [S1]. A sentence-perfect translation can still read wrong to a native speaker, because the seams between sentences never get adjusted.

This is not a new observation. A 2024 arXiv preprint on adapting LLMs for document-level machine translation noted the same gap [P2], and a 2023 paper on contextual refinement of translations explored how LLMs might use broader context for post-editing [P4]. The field has been circling this problem for years.

How PAT works

PAT (Pragmatic Auto-Translator) uses retrieval-augmented generation, the same technique that lets an LLM pull in relevant documents before answering a question [S1]. Here is what is different: PAT pairs user-configured specifications with context pulled from a comparable corpus of authentic longform texts in U.S. English and Latin American Spanish [S1]. It then passes retrieved examples at the paragraph, section, and document level to an LLM for whole-document generation [S1].

The goal is not a finished translation. It is a draft for professional verification [S1]. A human translator still reviews the output. PAT just tries to give them a better starting point, one that accounts for how the whole document flows rather than how each sentence reads in isolation.

What the evaluation found

The evaluation covered six automatic translations of essays on generative AI, across three projects, working from U.S. English into LATAM and Mexican Spanish [S1]. Two trained evaluators assessed the output using a customised MQM typology (MQM is a framework for classifying translation errors by severity and type) [S1].

Results were mixed. A limited prompt, one without the corpus context or user specifications, produced no meaningful reformulation [S1]. The LLM just translated sentence by sentence, as expected. But when PAT supplied specifications and corpus-informed context, the translations at times showed substantial reformulation [S1]. The target texts were reshaped to fit Spanish-language conventions, where the structure and style of argument differ from English [S1].

The key qualifier: the reformulation was "not always to effect" [S1]. Sometimes the system reorganised the text and the result was better. Sometimes it reorganised and the result was not. The authors conclude that LLMs can be moved toward reformulation and away from the sentence-by-sentence approach, but more work is needed to improve the effectiveness of those reformulations [S1].

PAT's mixed results point to a familiar tension: more context does not automatically produce better output. The question is when it helps and when it just adds noise.

What it means

The core finding is modest but real. LLMs are not locked into sentence-by-sentence translation. Given the right context, retrieved from a corpus of authentic texts in the target language, they can be pushed to restructure a document to fit how that language actually works. That matters because the gap between sentence-level accuracy and document-level fluency is where most machine translation still falls flat. A translation can be technically correct and still feel wrong to a native reader, because the argument flows the way an English speaker would build it, not the way a Spanish speaker would.

PAT's approach, using RAG to pull in document-level examples before generation, is a different bet than fine-tuning a model on translation pairs. A 2025 paper explored fine-tuning LLMs for document-level translation refinement using two intermediate translations [P5]. PAT instead keeps the model frozen and changes what it sees. That is cheaper to build and easier to adjust, because swapping the corpus or the user specifications does not require retraining.

What it means for business

For a two-person translation agency handling English-to-Spanish content, the practical question is whether PAT-style drafts save review time. The evidence here is thin: six translations, two evaluators, one subject domain. But the direction is clear. If a RAG-based system can produce drafts that already account for document-level conventions, a professional translator spends less time fixing structure and more time refining voice.

The cost picture is familiar to anyone running RAG in production. You need a curated corpus of authentic longform texts in both languages, which is not trivial to build. You need storage and retrieval infrastructure. And you need an LLM with a context window large enough to hold examples at every level, from paragraph to full document, plus the source text. Inference costs, the cost of actually running the model, go up with longer contexts. Whether the time saved in post-editing offsets the infrastructure cost is a calculation each operator would need to run on their own volume.

For a content team publishing in both U.S. English and Latin American Spanish, the appeal is consistency. A CAT tool translating sentence by sentence will produce different structural choices for each sentence, with no eye on the whole. PAT's approach, in theory, gives the LLM enough context to make consistent document-level decisions.

What we don't know yet

The study is small. Six translations, two evaluators, one domain (essays on generative AI), one language pair (U.S. English into LATAM and Mexican Spanish) [S1]. Whether the reformulation effect holds across other language pairs, other domains, or longer documents is unknown. The authors note this themselves.

The evaluation used a customised MQM typology, not a standard one [S1], which makes it hard to compare results against other systems. The authors are also evaluating their own system, which creates potential for confirmation bias.

The preprint has not been peer-reviewed [S1]. Findings are preliminary and may change.

The next concrete signal to watch: whether the authors release the corpus, the specifications, or the evaluation data. Without those, the results cannot be independently reproduced. A follow-up study testing PAT across multiple language pairs and domains would answer whether this approach generalises or is specific to English-to-Spanish longform essays.

If you found this useful, subscribe for more plain-English decoding of AI research as it lands.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.