Tiny AI Beats Giants at Summarizing Laws in 24 Languages
A model roughly 1,000 times smaller just out-performed the big names.
Two researchers, Sujay Uday Rittikar and Sheela Ramanna, have published a paper describing LexLattice, a program that summarizes legal documents by lifting out exact sentences from the original rather than writing new ones. They tested it on EUR-Lex-Sum, a public collection of European Union laws paired with human-written summaries, covering 24 languages. It scored better than much larger, instruction-tuned AI systems — the kind with billions of settings.
The trick is how it reads. Instead of treating a law as one long wall of text, LexLattice lays out the document's structure — titles, chapters, articles, paragraphs — as a two-dimensional grid, then lets small units on that grid repeatedly pass notes to their neighbors, like a rumor spreading across a chessboard, until the important parts stand out. That method is called neural cellular automata. All the actual learning happens in a tiny 'consolidator' of 1.8 million settings, while the language-reading engine underneath stays frozen and untouched.
The multilingual result is the real headline. When the consolidator was trained only on widely spoken languages, it still worked on languages it had never seen, keeping about 99% of its quality. That suggests it is learning meaning and structure rather than memorizing words. For anyone who has stared at an insurance policy, rental contract, or immigration form in a language they don't read, this points toward a future where you get a short, checkable digest instead of a 40-page wall.
Now the honest caveats. This is a research paper, not a product you can use today. The method only selects existing sentences — it cannot explain them, translate them into everyday language, or rewrite anything. And 'summarizing law well' here is measured by automated word-overlap scores, not by whether an actual lawyer or citizen understood the result. Twenty-four European languages is also a narrow slice of the world, and legal accuracy is exactly where mistakes are most expensive.
- LexLattice pulls exact sentences out of long legal documents, so every line in its summary can be traced back to the original text.
- It handles 24 languages and still works on languages it was never trained on, keeping about 99% of its quality.
- It uses only 1.8 million adjustable settings — roughly a thousand times fewer than the massive AI models it outperforms, meaning it could run cheaply.
Why It Matters
Soon you may get instant, checkable summaries of contracts and laws in languages you don't speak.