Polhem 1.0 finds more of the prior art examiners cite
Polhem 1.0 is Stilta’s own embedding model for prior art search. We trained it on millions of citations made by patent examiners. On 3,969 patent applications held out from training, it finds 72.6% of the prior art examiners cited in its top 1,000 results.
- 72.6%
- of examiner-cited prior art found in the top 1,000 results
- 98%
- of cited patents found in the top 10,000 results
- +19.3 pts
- over the untrained base model, at 345M parameters
The problem
Every prior art search starts with retrieval. Before anyone, human or agent, can read and compare references, a system has to pull the right candidates out of more than 170 million patent documents. If a key reference never makes that first list, no amount of careful analysis later can bring it back.
General-purpose embedding models are trained to match text that reads alike. Patent search asks for something narrower: matching an invention to the earlier documents that disclose the same technical idea, often in different words and written for a different purpose.
To get the best possible results for our users, we don’t rely on one model for everything. We use leading LLMs where they’re strongest, and we train our own models where that makes the difference. Retrieval for prior art search is one of those places.
How we trained Polhem 1.0
Patent examiners leave a public record of their judgment: for each application, the earlier documents they cited as relevant prior art. Across the patent system, that adds up to millions of expert-made pairs of a new application and the art that matters for it.
We trained Polhem 1.0 on this public patent data, starting from PaECTER, an open patent embedding model [3], and teaching it to place an application close to the documents examiners cited against it. The result is a compact model with 345M parameters.
To measure it fairly, we held 3,969 patent applications out of training entirely. For each one, we ran the search and counted how many of the examiner’s citations appeared in the results.
Results
We compared Polhem 1.0 with the untrained PaECTER model it was built on, and with a range of open embedding models, from 345M to 8B parameters, on three benchmarks: our own holdout and two public research datasets, DAPFAM [1] and perf200 [2]. We ran every model ourselves, with the same setup, in October 2026. Polhem 1.0 ranks first on all three. Closed embedding models and late-interaction models were not part of the comparison, as they are not viable at the scale of a full patent corpus.
Stilta holdout
On the holdout, Polhem 1.0 finds 72.6% of the examiner-cited prior art in its top 1,000 results (Recall@1000). That is 19.3 percentage points more than untrained PaECTER, which is what our training adds, and 11.5 points more than the strongest open model we tested, Nemotron 3 8B, which has more than 20 times as many parameters. Deeper in the ranking, 98% of cited patents appear within the top 10,000 results.
DAPFAM and perf200
DAPFAM pairs 1,247 query patent families with the families they cite, among 45,336 targets [1]. perf200 takes 200 European patent applications and the documents examiners cited in their search reports, weighted by citation category (X, Y and A), among 47,636 candidates [2]. On both we report nDCG, which rewards a model for placing relevant documents near the top of the list.
On DAPFAM, Polhem 1.0 scores 0.399 nDCG@100, against 0.368 for the best open models (Nemotron 3 8B and Qwen3-Embedding 8B) and 0.342 for untrained PaECTER.
On perf200, Polhem 1.0 scores 0.228 nDCG@10. The next best model, Octen-Embedding 8B, scores 0.189, and untrained PaECTER 0.173.
What this means in practice
Polhem 1.0 is now live in Stilta, end to end. That matters because retrieval sets the ceiling for everything that follows: Stilta’s agents read the patent, break down its claims and evaluate each candidate reference, but they can only evaluate what retrieval brings back.
For invalidity, freedom-to-operate and landscape work, higher recall means fewer missed references, so our customers can rely on the most complete prior art picture available.
Why the name Polhem
Christopher Polhem (1661–1751) was a Swedish inventor and engineer, often called the father of Swedish mechanics. A model built to find earlier inventions seemed a fitting place for his name.
References
[1] Ayaou, I., Cavallucci, D., Chibane, H. (2025). “DAPFAM: A Domain-Aware Family-level Dataset to benchmark cross domain patent retrieval.” arXiv:2506.22141. Dataset: huggingface.co/datasets/datalyes/DAPFAM_patent, licensed CC BY-NC-SA 4.0.
[2] Zuo, Y., Gerdes, K., de la Clergerie, É., Sagot, B. (2025). “Patent Representation Learning via Self-supervision.” arXiv:2511.10657. perf200 is the prior art search set in huggingface.co/datasets/ZoeYou/downstream, licensed CC BY-NC 4.0.
[3] Ghosh, M., Erhardt, S., Rose, M. E., Buunk, E., Harhoff, D. (2024). “PaECTER: Patent-level Representation Learning using Citation-informed Transformers.” arXiv:2402.19411.
Stilta is an agentic AI system purpose-built for patent invalidity analysis - both prior art search and claim-level analysis. Built by ex-McKinsey QuantumBlack AI engineers.
To see how Polhem 1.0 performs on your patents, reach out at
team@stilta.com or book a demo below.
