AutoPatent: A Multi-Agent Framework for Automatic Patent Generation

TL;DR

Draft2Patent formalizes the task of converting inventor drafts into complete patents and introduces the D2P benchmark with 1,933 draft-patent pairs. The AutoPatent framework employs a multi-agent system (planner, six writers, and examiner) guided by PGTree and RRAG to generate full patents averaging $17K$ tokens, outperforming larger LLMs in objective metrics and human judgments. Key findings show that smaller models with AutoPatent can surpass bigger models in quality and consistency, and the system demonstrates strong generalization across baselines. This work highlights a practical pathway to automating patent drafting while acknowledging ethical and evaluative limitations and suggesting directions for automated patent evaluation.

Abstract

As the capabilities of Large Language Models (LLMs) continue to advance, the field of patent processing has garnered increased attention within the natural language processing community. However, the majority of research has been concentrated on classification tasks, such as patent categorization and examination, or on short text generation tasks like patent summarization and patent quizzes. In this paper, we introduce a novel and practical task known as Draft2Patent, along with its corresponding D2P benchmark, which challenges LLMs to generate full-length patents averaging 17K tokens based on initial drafts. Patents present a significant challenge to LLMs due to their specialized nature, standardized terminology, and extensive length. We propose a multi-agent framework called AutoPatent which leverages the LLM-based planner agent, writer agents, and examiner agent with PGTree and RRAG to generate lengthy, intricate, and high-quality complete patent documents. The experimental results demonstrate that our AutoPatent framework significantly enhances the ability to generate comprehensive patents across various LLMs. Furthermore, we have discovered that patents generated solely with the AutoPatent framework based on the Qwen2.5-7B model outperform those produced by larger and more powerful LLMs, such as GPT-4o, Qwen2.5-72B, and LLAMA3.1-70B, in both objective metrics and human evaluations. We will make the data and code available upon acceptance at \url{https://github.com/QiYao-Wang/AutoPatent}.

Figures (7)

• Figure 1: Draft2Patent Task. Automating patent drafting by simulating real-world scenarios.
• Figure 2: An overview of the AutoPatent framework, which includes eight agents and four steps for automatically generating a patent $\mathcal{P}$ from a draft $\mathcal{D}$. The fire represents the corresponding agents' parameters are fine-tuned, while the snowflake represents the agents' parameters are frozen. The magnifying glass represents using the description writer for retrieval, while the question mark indicates using the examiner agent for text review. The dashed arrows represent the model’s input stream, while the solid arrows represent the model’s output stream.
• Figure 3: PGTree structure. The input of the planning agent is the draft $\mathcal{D}$, while the output is a PGTree $\mathcal{W}$, represented as a two-layer multiway tree.
• Figure 4: Human evaluation results. For each comparison, the left number indicates the count of AutoPatent wins, the middle shows the ties, and the right represents the cases where AutoPatent loses.
• Figure 5: A case for repetition error of patent generated by SFT.