Automating Patent Drafting from Inventor Disclosures

Summary & Methodology Analysis

The Patent-MAF system employs a multi-agent orchestration strategy to handle the high-level reasoning required for patent drafting. The process begins with a manager agent that ingests raw disclosures and relevant figures to structure the output into specific artifacts. These artifacts are then routed to specialized drafters: one focused on generating claims and a vision-enabled module that generates specifications based on the visual input. A final polisher agent acts as a consistency layer, ensuring that the terminology remains unified across the entire document structure. This modular approach is designed to bridge the gap between early-stage notes and the rigorous legal requirements of patent filings.

Evaluation is grounded in several benchmarks, including the Dis2Pat dataset which targets more realistic patent drafting scenarios. The system also builds upon previous work like HUPD-DCG and EPD, which facilitate claim generation from full specifications, and Patent-CR, which treats claim revision as a post-editing task. In performance tests, Patent-MAF reached an overall content quality score of 85.4 for specification generation. This result surpasses both the open-source Qwen3-VL-72B model, which scored 81.8, and the closed-source GPT-5, which scored 85.1. For claim generation specifically, the system achieved a score of 86.7.

Despite these results, the system has clear limitations. The authors note that the dataset relies on pseudo-disclosures generated by an LLM, which may not fully replicate the ambiguity or incompleteness of actual inventor submissions. Furthermore, the framework is currently evaluated primarily on English-language patents. While the researchers used automatic and expert evaluations to validate the quality of their input data, the inability to capture the full diversity of real-world invention disclosures remains a challenge for future production readiness.