Patentformer: A demonstration of AI-assisted automated patent drafting

TL;DR

Patentformer introduces an AI-assisted platform for automated patent specification drafting, addressing the challenge of producing legally coherent and technically accurate text from claims and drawing descriptions. It formalizes the input-output transformation with enriched inputs $\mathcal{T'} = (\mathcal{C'}, \mathcal{B'}, \mathcal{N'})$ to generate $\mathcal{S'}$ using a fine-tuned patent-oriented LLM, following preprocessing, mapping, and post-processing steps. The work provides a novel dataset, Patent-2015-2024-G06F, and demonstrates through a user study that the system achieves high legal and linguistic quality, while outlining limitations in image understanding and a path toward multimodal models. The results suggest practical potential for accelerating patent drafting and reducing dependence on expert attorneys, with future enhancements focused on LVLM integration and automated figure generation.

Abstract

Patent drafting presents significant challenges due to its reliance on the extensive experience and specialized expertise of patent attorneys, who must possess both legal acumen and technical understanding of an invention to craft patent applications in a formal legal writing style. This paper presents a demonstration of Patentformer, an AI-powered automated patent drafting platform designed to support patent attorneys by rapidly producing high-quality patent applications adhering to legal writing standards.

Figures (4)

• Figure 1: An example of a patent drawing (left), claim, specification, and a brief description of the drawing (right).
• Figure 2: Workflow of the Patentformer system. The user provides the claims $\mathcal{C}$, and the images $\mathcal{I}$ with descriptions $\mathcal{B}$ to the Patentformer, then the Patentformer processes the inputs automatically to an enhanced text version, e.g., $\mathcal{C'}$, $\mathcal{N'}$, and $\mathcal{S'}$, as described in wang2024patentformer. Furthermore, a user interface is provided to the user to map the correct relationship between claims with components and descriptions. Thereafter, the system automatically prepares the linked mapping model input $\mathcal{T'}$ and sends it to the fine-tuned large language model. Finally, the model outputs the generated specifications $\mathcal{S'}$ and is automatically cleaned to a clean version $\mathcal{S}$ before presenting to the user.
• Figure 3: User interface and output of the Patentformer system from input to generated specification.
• Figure 4: Human evaluation score across four qualities: linguistic, legal, figure descriptive, and technical quality. The user agrees with the legal (the most important aspect of the patent) and the linguistic quality of the drafting from the Patentformer and expects the potential from the multimodal version of the Patentformer to improve the figure descriptive and technical qualities.