arXiv paper trains language models to shrink Boolean patent searches
Researchers report a training method that helps language models write patent search queries whose result sets are small enough for humans to screen, trading some recall for precision.
In plain words
Patent searchers use Boolean queries, which are search formulas built from operators like AND and OR, to find earlier documents related to a patent. The paper says models can already write valid, high-recall queries, but the resulting document sets are too large to review. The authors trained models to narrow those sets while keeping the relevant documents in.
What was announced
The arXiv paper, an abstract-level cross-listed posting, describes distilling 38,397 Japan Patent Office search reports into executable expert strategies, which the authors re-executed and scored by citation recall and by the smallest set containing each citation. They found that prompted models produce valid, high-recall strategies but leave citations in sets too large to screen, so they adapted a generator offline from scored rollouts in two ways: strategy-level learning from a screening-aware score, and line-level learning from citation-preserving narrowing credit.
On shared rollouts, the line-level credit outperformed the strategy-level score on screening measures. Iterated, it reportedly cut the median citation-containing set from 768 to under 50 documents and placed more citations in sets under 100 documents than expert strategies, though at lower recall. Strategy-level preference learning reached near-complete recall, and the authors describe the two forms as complementary. The results come from the paper's own evaluation and have not been independently confirmed.
What it does not show
- The source is the paper's abstract only; full methods, datasets and caveats are not covered here.
- Reported gains, including the 768-to-under-50 median reduction, are the authors' own measurements without independent confirmation.
- The abstract does not state whether the approach has been deployed or validated outside the Japan Patent Office data.
For security and IT teams
- Scope
- The evidence covers the paper's abstract; it does not include the full article, implementation details or availability of the trained models.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.