AI/ ai · patents · rag · arxiv

Researchers Build an AI System That Drafts Full Patents

A logic-tree-guided AI framework drafts full patents, but its paper omits the actual numbers behind claimed gains over rival LLMs.

A new AI framework called LogicTree-RAG says it can draft entire patent applications instead of stitching together disconnected sections.

Researchers describe LogicTree-RAG in a paper posted to arXiv on September 28, 2026 (arXiv:2609.30943, https://arxiv.org/abs/2609.30943). The system builds a hierarchical logic tree that maps out a patent's technical elements before writing begins, generating each node through what the authors call evidence-guided recursive generation pulled from retrieved source material. A hybrid traversal step then converts that tree into the standard sections of a patent filing, including claims, background, and detailed description, without a human-drafted outline. The team says the approach beats unspecified LLM-based baselines on content quality, language conformity, and token efficiency, and produces longer, more structured documents in the process.

Patent drafting is a slog: examiners expect legally precise, technically exhaustive documents that stay consistent across dozens of pages, and most AI drafting tools handle only isolated sections or still need a lawyer to sketch the outline first. LogicTree-RAG's pitch is full-document automation with no expert-authored scaffolding, which, if it holds up, would move patent drafting closer to a genuinely automatable workflow rather than an AI-assisted one.

One catch: the abstract touts wins over strong LLM-based baselines without publishing a single benchmark number, so until the full paper or code lands, the improvement claims are unverified.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →