A new test for inventive step

Dr. R. Free (GB), CMS Cameron McKenna Nabarro Olswang, LLP

This article presents my idea of a new test for inventive step which is partially automated and uses “vector search” technology. I first presented this idea in a lecture in October 2023 as part of the Basel IP lectures in Switzerland.


In August 2023 the launch of AI-PreSearch (12) Introducing AI-PreSearch: A Revolutionary AI-Driven Search Tool to support our Patent Examiners | LinkedIn was announced as part of the EPO PreSearch framework and shortly afterwards the USPTO gave a webinar demonstrating a similar AI patent search tool. I understand both these AI search tools to use a type of “vector search” where text from a patent document is represented as a vector in a multi-dimensional space. To convert the patent document text into a vector in the multi-dimensional space a neural network is used such as a large language model with technology similar to chat GPT, GEMINI, Mistral Large, and others. The vector can be thought of as coordinates of a point in the multi-dimensional space. Because of the characteristics of the neural network, points in the space representing semantically similar content are nearby one another. The multi-dimensional space is pre-populated with vectors representing prior art documents; there may be hundreds of millions of vectors, one per prior art document. Each vector is referred to as an embedding since it “embeds” a patent document into the multi-dimensional space, also referred to as the embedding space. The vector search process starts from the position in the multi-dimensional space representing the patent application, and looks around in the space for neighbours. The “looking around” can be computing a cosine similarity metric between the vector representing the patent application and different ones of the vectors representing the prior art. Neighbouring vectors which are found represent potential prior art citations and can be offered to an Examiner as part of the PreSearch framework.

The new test

Suppose a new patent application is received and converted into an embedding vector using a large language model. The idea for a new test for inventive step is “the new application is inventive if the embedding space around the embedding vector of the new application, within a radius of x, is empty and there is a technical effect”. The diagram illustrates this idea although it shows only two dimensions and in reality there are many hundreds of thousands of dimensions. Because the grey space around the new application is empty there is an argument that the new application is inventive. This test assumes that the embedding space is smooth and continuous; which may not be the case since presumably the embedding space contains excluded matter holes.

How can the value of x be set?

Values of x could be found from historical data about granted patents and the state of the art. The historical values could then be used to determine a value for x to use now. Perhaps the value of x would be found to be different in different technical fields.

Can the test be used now?

EPO Examiners use the problem and solution approach when deciding whether or not to object on inventive step. The problem and solution approach is not mandatory and so there could be room for a vector space radius test as well. EPO technical board of appeal case T188/09 explains that an alternative to the problem and solution approach is allowed where one of the situations of T465/92 applies. Those situations include the case where there is no close prior art because the invention breaks new ground; which is exactly what the vector radius test is doing. Therefore there could be a possibility to use the new test as a two stage test for inventive step as follows:

What about combinations of documents?

The suggested test does not deal with combinations of documents and yet the Guidelines Part G, Chapter VII sets out how pieces of prior art may be combined in the context of the problem and solution approach.

It is possible to apply mathematical operations in the multi-dimensional vector space. Scientists have shown that mathematical operations applied to embeddings of words yield semantically meaningful outcomes in some cases. According to the following blog, Word Embeddings: Intuition behind the vector representation of the words | by Oleg Borisov | Towards Data Science “it has been found that the embedding vectors can magically understand that there are some relationships between the words like “a King to a Queen is as a Man to a Woman”, as well as that if we are talking with respect to a Country-Capital hyperplane, then Spain is related to Madrid similarly as Italy is related to Rome”. Therefore there could be a possibility that mathematical operations applied to embeddings of patent items yield something meaningful rather than something ridiculous. Starting from the vector representing the closest prior art document, does addition of a vector representing another prior art document lead towards the embedding of the patent application?

Fundamental technical problems

Representing patent items in vector space gives a new way to understand patent data and perhaps could lead us to find fundamental technical problems like those proposed in my epi Information 4|18 article “Technical problems in AI Inventions in the Light of the Guidelines for Examination in the EPO” October 2018. Fundamental technical problems are ones which cannot be divided and examples were “how to save space”, “how to save time”. Some fundamental technical problems seemed to arise from the laws of physics such as “how to save space” whereas some seemed to be human made such as “how to improve security” or “how to implement frontier AI safety”. If those fundamental technical problems exist perhaps we can find them as clusters of vectors in the AI-PreSearch vector database.


Using AI tools to help with assessment of inventive step by applying a radius in vector space may be a useful guide for patent stakeholders when used together with the problem and solution approach.