Benchmarking Patent Drafting from Inventor-Style Disclosures
The paper introduces Dis2Pat, a benchmark for turning informal inventor disclosures into full patent applications.
The authors argue that prior patent-drafting work starts too late in the process, using structured or already legalistic inputs. Their dataset is meant to reflect earlier invention materials written in de-legalized, inventor-style language. They also propose Patent-MAF, a locally deployable multi-agent drafting baseline. In their benchmark, Patent-MAF beats evaluated open-source models and stays competitive with large closed-source models. ArXiv · AI/CL/LG's note
The authors argue that prior patent-drafting work starts too late in the process, using structured or already legalistic inputs. Their dataset is meant to reflect earlier invention materials written in de-legalized, inventor-style language. They also propose Patent-MAF, a locally deployable multi-agent drafting baseline. In their benchmark, Patent-MAF beats evaluated open-source models and stays competitive with large closed-source models. ArXiv · AI/CL/LG's note
score 4