Large language models in the patent lifecycle: A systematic review

Abstract

Patents constitute the world's largest repository of codified technical knowledge, yet their length, legal terminology, and linguistic ambiguity have limited the effectiveness of traditional text mining and keyword-based retrieval. Large Language Models (LLMs) offer new capabilities for processing this complexity, but the resulting literature is fragmented across disciplines, with no comprehensive synthesis of how these tools are being applied across the patent lifecycle. This study systematically reviews the literature on LLMs and Generative AI in patent analytics, following the PRISMA 2020 protocol, to map current applications, characterise dominant architectural trends, and identify priority research gaps. Using Scopus and Web of Science, we identified and screened a corpus of 129 peer-reviewed studies published from 2018 to early 2026, organised according to three patent lifecycle stages: creation and pre-filing (N = 36), examination and prosecution (N = 34), and post-grant management (N = 60). The review shows that research activity has grown sharply since 2024, accompanied by a shift from exploratory, single-task studies toward applied, multi-component systems. Across this literature, Retrieval-Augmented Generation (RAG) has become the dominant architectural response to hallucination risk, grounding generated content in verified external sources. The review further identifies persistent gaps in domain-specific evaluation metrics, multimodal reasoning, and governance frameworks for AI-generated patent content. For patent professionals and institutions, the findings indicate that LLMs are best positioned as assistive tools requiring human oversight rather than autonomous replacements, and that the central challenge going forward is not whether but how the patent system governs this transition.