Researchers have created a new dataset called Dis2Pat to help improve the performance of AI systems in generating patent applications from informal inventor disclosures. The dataset requires AI models to generate complete and legally coherent patent applications directly from these early-stage invention materials. To evaluate the current state of AI in this task, the researchers also propose a baseline model called Patent-MAF, which is a multi-agent framework for locally deployable patent drafting. Benchmark results show that current large language models (LLMs) struggle with patent drafting and that Patent-MAF outperforms other open-source models. --- Why it matters: This research matters to AI engineers because it highlights the limitations of current LLMs in generating complete and legally coherent patent applications, a crucial task in real-world patenting workflows. The proposed baseline model provides a strong foundation for future research in this area, which could lead to more accurate and efficient patent drafting. Source: https://arxiv.org/abs/2608.21249