A stripped-down AI pipeline outperforms elaborate multi-step systems when translating medieval Latin manuscripts.
Researchers tested a range of image-to-translation pipelines on handwritten medieval Latin texts — a domain where scribal shorthand, ligatures, and degraded parchment expose weaknesses that clean-text benchmarks never reveal. Using the CATMuS Latin dataset as a benchmark, they found that domain-specific OCR models cut character error rates by up to 4.3 times compared to general-purpose Vision Language Models, despite having far fewer parameters. The team also released the Interpres-Parallel-Corpus (IPC), a dataset of 1,383 aligned manuscript image lines with transcriptions and expert translations — the first of its kind for medieval Latin.
The counterintuitive finding is the headline: the simplest pipeline — a specialized OCR model feeding directly into a large language model — beat every more elaborate variant. Adding retrieval-augmented generation or post-OCR correction actually made things worse, introducing what the researchers call prompt saturation and error propagation. That is a useful corrective for anyone who assumes that stacking more AI components always improves results.
The findings land at a moment when labs are racing to expand AI into low-resource and historical domains, often by scaling up general models rather than refining specialized ones. This paper suggests that in messy, niche domains, a purpose-built small model doing one job well is still hard to beat.