AIARD
FrançaisEnglish

Research / T3

Patent and scientific literature intelligence

Millions of solved problems, written in a language machines must learn.

Training and evaluating language models specialised in patents and scientific articles: retrieving the documents that answer a problem, across domains and across patent families.

Generic text embeddings read patents poorly. The team has published specialised patent embedding models and benchmarks that measure how well a system finds relevant patents inside and outside the problem’s own technical domain.

These resources power PatentSpark and the patent context of our other demonstrators.

Related projects

People

Demonstrators

TRIZBench

DemonstratorMembers

From a problem statement to the few contradictions that really matter.

Thesis of Nathan Witkowicz, conducted with funding from ArcelorMittal

PatentSpark

DemonstratorMembers

Find, in patents, the solutions and contradictions related to your problem.

PhD thesis of Iliass Ayaou (2025)

PIXSEB

TransferredMembers

Collect contradictions across competitors and prioritise the resolution effort.

Maturation project led by Achille Souili as a postdoctoral researcher (PhD thesis 2015)

Related publications

All publications