FlowPlan-G2P: Structured Paper-to-Patent

• The paper introduces FlowPlan-G2P as a three-stage framework that uses concept graph induction, section planning, and graph-conditioned generation to create patent descriptions.
• FlowPlan-G2P is designed to address the rhetorical and legal shift required to transform scientific prose into a structured, legally robust patent document.
• The framework demonstrates significant gains in metrics like Technical Content Fidelity and Legal-Professional Compliance over baseline methods.

FlowPlan-G2P is a structured generation framework for transforming scientific papers into patent descriptions. It is formulated for paper-to-patent generation in the specific sense of producing a patent-style “Description” section that is technically faithful, structurally organized, and legally usable, rather than a summary or stylistic paraphrase. The framework replaces direct paper-to-text rewriting with a three-stage pipeline—Concept Graph Induction, Paragraph and Section Planning, and Graph-Conditioned Generation—and is presented as an explicit model of the hidden intermediate reasoning performed by expert patent drafters when they identify invention components, organize them into canonical legal sections, and draft section-specific prose (Pan et al., 5 Jan 2026).

1. Problem setting and rhetorical-legal shift

The task addressed by FlowPlan-G2P is difficult because a scientific paper and a patent specification serve fundamentally different purposes. A scientific paper emphasizes novel findings, evidence, and theory, whereas a patent description must function as a legal-technical disclosure document. The framework therefore treats paper-to-patent generation as a rhetorical and legal shift rather than as a conventional style-transfer problem (Pan et al., 5 Jan 2026).

Two constraints are central. The first is rhetorical disparity: scientific writing argues that a discovery is valid and interesting, while patent writing explains the invention in a way that supports legal protection and implementation. The second is the enablement requirement. Patent law requires enough technical disclosure for a Person Having Ordinary Skill in the Art (PHOSITA) to reproduce the invention without undue experimentation. In this setting, drafting often requires logical expansion rather than compression. That point is decisive for methodology, because summarization-oriented instincts are misaligned with the patent-description objective.

The paper also foregrounds the structural rigidity of patent descriptions. Distinct sections such as Field, Background, Summary, and Detailed Description have fixed rhetorical roles, and the document must preserve a global problem–solution–implementation–effects logic across long spans of text. On this view, weaknesses in black-box text-to-text systems are not incidental. The critique is that direct LLM drafting and outline-guided chunk generation struggle with paragraph-level information flow, legal conformity, long-range logical consistency, technical faithfulness, hallucination suppression, and explicit support for enablement.

This framing implies that the central obstacle is not merely lexical mismatch between research prose and patent prose. A plausible implication is that the main bottleneck lies in the absence of an explicit intermediate representation for invention structure and section-level discourse organization.

2. Architecture and intermediate representations

FlowPlan-G2P is organized as a three-stage pipeline that mirrors how patent professionals work: first identify key invention elements and dependencies, then organize them into legally conventional sections, and then draft each section using patent-appropriate phrasing. The framework externalizes this workflow through two explicit intermediate artifacts: a refined directed concept graph $G^*$ and a section-level plan $P$ (Pan et al., 5 Jan 2026).

Stage	Output	Function
Concept Graph Induction	Refined directed concept graph $G^*$	Captures invention semantics
Paragraph and Section Planning	Plan $P=(S,T)$	Allocates and orders content
Graph-Conditioned Generation	Section-wise patent paragraphs	Realizes planned content in patent style

The paper describes this as a structured transformation process rather than direct rewriting. In that formulation, the graph serves as an interpretable representation of invention semantics, while the plan specifies content allocation and narrative order. The three claimed contributions are correspondingly organized around a structured transformation paradigm, a concrete framework integrating graph induction, section planning, and graph-conditioned generation, and an expert-centric workflow model that separates technical reasoning, structural planning, and legal drafting.

The significance of this decomposition lies in controllability. Instead of asking a single model invocation to infer invention logic, section boundaries, and legal rhetoric simultaneously, the system distributes those burdens across modular stages. This suggests that patent-description generation is being treated as a document planning problem with explicit discourse constraints rather than as unconstrained long-form generation.

3. Concept Graph Induction

The first stage converts a scientific paper $D$ into a patent-oriented structured representation whose output is a refined directed concept graph $G^*$. The reasoning process is decomposed into canonical patent-oriented categories: $$S = \{Field, TechProblem, PriorArt, Novelty, Solution, Implementation, Effects, Embodiment, Figure\}. \tag{1}$$ For each category $R_i$, the model generates structured reasoning conditioned on the document and on previously generated reasoning steps: $R_i = LLM(expert\_prompt;(D, R_{1:i-1})). \tag{2}$ This staged dependency is intended to preserve causal continuity across categories (Pan et al., 5 Jan 2026).

The prompting strategy is explicitly expert-oriented and aligned with standard patent phrasing. Examples include “The present invention relates to ...” and “However, conventional technology has the following drawbacks ...”. Prompts also include role-specific constraints, patent boilerplate, and context windows referencing previously generated sections. The method therefore does not attempt flat extraction of entities and relations from the source document; it asks the model to reinterpret the paper from a patent drafter’s perspective.

The extracted reasoning steps are converted into a directed concept graph

$G = (V, E),$

where nodes represent patent elements and the edge set is

$P$0

The graph may therefore contain nodes such as algorithms, modules, problems, embodiments, and effects, linked by relations such as solution-to-problem, implementation-to-solution, and effect-to-claim support.

To improve recall, the system generates three candidate graphs, $P$1. Graph 1 uses rule-based construction with predefined edge templates such as TechProblem $P$2 solves $P$3 Solution and Solution $P$4 implements $P$5 Implementation. Graphs 2 and 3 use LLM-based relation inference to capture implicit dependencies not encoded by those templates. The final graph $P$6 is then formed through merging and refinement procedures that enforce node coverage, relation consistency, and edge validity. The described mechanism includes pairwise semantic similarity, rule-based filtering, majority voting for conflicting edge types, and union semantics for nodes. It also prunes isolated nodes, removes redundant relations, removes invalid cycles, and verifies that required node types—Field, TechProblem, and Solution—are present, injecting placeholder nodes when necessary.

A notable control mechanism appears in the graph-merging prompt: it instructs the model not to invent new nodes or edges, to preserve node labels exactly, to prefer edge types that appear more than once, and to return JSON with "nodes" and "edges". The paper is equally explicit about what is not formalized. Although equations (1)–(3) are given, there is no end-to-end training loss or learned graph induction objective. The graph stage is procedural and prompt-driven rather than a supervised graph-learning regime. A common misconception is therefore to treat FlowPlan-G2P as a graph neural network or graph optimization system; in the paper, it is instead an LLM-guided symbolic construction process.

4. Paragraph and Section Planning

Once $P$7 is available, the second stage reorganizes it into a patent-style document plan

$P$8

where $P$9 is a set of section-specific subgraphs and $G^*$0 is their global order. This representation encodes both content allocation and narrative ordering. The planner uses the canonical patent sections Field, Background, Summary, Detailed Description, and Effects, and the paper gives the explicit example that Solution and Implementation nodes are aggregated into Detailed Description (Pan et al., 5 Jan 2026).

The system generates $G^*$1 candidate plans. Each candidate clusters graph nodes into section-specific groups using an LLM guided by hierarchical embeddings aligned with patent section types. The exact embedding model and clustering algorithm are not specified, but the acceptance mechanism is formalized through two criteria. For section $G^*$2, intra-section connectivity is

$G^*$3

This measures how tightly connected the assigned content is within the section. Semantic consistency is defined as

$G^*$4

where $G^*$5 is the entropy of node-type distribution in section $G^*$6. The acceptance thresholds are $G^*$7 and $G^*$8. If no candidate satisfies both, the system selects the plan with the highest combined score as fallback.

The planning stage also uses domain heuristics that directly encode patent conventions. Plans are pruned if they place embodiments before technical problems or if they separate figures from associated implementations. The final global narrative order is constrained to follow Problem -> Solution -> Implementation -> Effects. This is one of the clearest operationalizations of the framework’s expert-workflow premise.

The importance of this stage is architectural. It converts graph semantics into a verified outline before prose generation begins. This suggests that long-range rhetorical control is delegated to an explicit planner rather than left to emergent behavior during decoding.

5. Graph-conditioned realization and compliance model

The final stage generates the patent description section by section, conditioned on the verified subgraphs. For each section subgraph $G^*$9, the prompt is constructed as

$P=(S,T)$0

Generation is therefore controlled by a section-specific instruction and a linearized graph representation rather than by the full scientific paper (Pan et al., 5 Jan 2026).

The paper states that the model generates one paragraph per section, although the Detailed Description prompt explicitly asks for multiple embodiments and figure references within that section. Decoding uses a low temperature, with temperature = 0.2 and top-k = 10. Section-specific prompting is highly prescriptive: Field must be concise and technically specific; Background must frame prior art and its drawbacks; Detailed Description must include multiple embodiments and direct figure references; Effects must articulate technical advantages.

The appendix specifies required boilerplate for each section, including:

• Field: “The present invention relates to ...”
• Background: “According to the prior art ...” and “However, such technology has the following problems ...”
• Summary: “An object of the present invention is to provide ...”
• Detailed Description: “According to one embodiment of the present invention ...”, “In another embodiment ...”, and figure references such as “Figure 1 illustrates ...”
• Effects: “Therefore, according to the present invention, the effect is ...”

The prompts also explicitly require repetitive elaboration across embodiments, inclusion of alternative embodiments, and consistent figure references. Few-shot examples are retrieved from professional patent corpora to reduce stylistic drift, although the retrieval mechanism, selection strategy, and few-shot context size are not reported.

After generation, the framework performs post-generation validation through two checks: semantic fidelity, assessed by an LLM-based entailment metric comparing the paragraph to the source subgraph, and token-level coverage of key subgraph concepts. If discrepancies exceed a threshold, the paragraph is regenerated. The paper does not define a formal legal rule checker, and it is careful about the meaning of “legally compliant.” In this work, that phrase refers to stronger adherence to patent-description conventions and statutory disclosure norms, especially enablement or sufficiency of disclosure, rather than formal legal validity in court. The compliance model is thus operationalized through planning constraints, prompt templates, and evaluation with Pat-DEVAL rather than through formal legal verification.

6. Evaluation, empirical findings, and limitations

The experiments use Pap2Pat-EvalGold, a refinement of Pap2Pat that keeps only paper–patent pairs with cosine similarity $P=(S,T)$1 using Sentence-BERT and Author Overlap Ratio $P=(S,T)$2, yielding 146 high-quality pairs. The main baseline comparison uses Zero-Shot Prompting, Few-Shot Prompting, Pap2Pat, and FlowPlan-G2P, all with Claude-4.5 as the backbone model to isolate methodology. The evaluation framework is Pat-DEVAL, based on Chain-of-Legal-Thought (CoLT), with four dimensions: Technical Content Fidelity (TCF), Data Precision (DP), Structural Coverage (SC), and Legal-Professional Compliance (LPC). The paper also reports ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore, and includes a human evaluation by four patent experts or scholar on 75 generated descriptions, with Kendall’s $P=(S,T)$3 between Pat-DEVAL and human judgments ranging from 0.67 to 0.76 (Pan et al., 5 Jan 2026).

Model	TCF	DP	SC	LPC
Zero-Shot Prompting	1.8	1.5	1.9	1.6
Few-Shot Prompting	2.4	2.1	2.5	2.2
Pap2Pat	3.5	3.2	3.4	3.1
FlowPlan-G2P	4.6	4.5	4.7	4.8

These results are the paper’s central empirical claim. Relative to Pap2Pat, FlowPlan-G2P improves by $P=(S,T)$4 TCF, $P=(S,T)$5 DP, $P=(S,T)$6 SC, and $P=(S,T)$7 LPC. The paper also highlights a “Metric Paradox”: Zero-Shot achieves the highest BERTScore and ROUGE-L in one comparison, yet human experts rate it poorly on legal compliance, while FlowPlan-G2P receives lower BS and ROUGE-L than Zero-Shot but reaches Human-LPC $P=(S,T)$8. This is used to argue that surface lexical or embedding similarity is a poor proxy for patent quality.

Robustness is reported across three LLM backbones: Llama-4-scout, Deepseek-v3.1, and Claude-4.5. The paper emphasizes that methodology can dominate raw model scale, noting that Llama-4-scout + FlowPlan-G2P reaches 4.3, exceeding Vanilla Claude-4.5 at 2.3. At the same time, several limitations are explicit. The framework mainly targets descriptive sections and does not ensure real-time legal correspondence between descriptions and claims. It relies on general principles and static legal knowledge rather than jurisdiction-specific precedent. Important implementation details remain unspecified, including the exact clustering algorithm, the exact hierarchical embedding model, the retrieval method for few-shot patent examples, thresholds for entailment-based regeneration, and computational cost. The paper also does not provide a classical ablation study, detailed error-analysis tables, or full case studies.

A further source of confusion is nomenclature. The unrelated robotics paper “FlowPlan: Zero-Shot Task Planning with LLM Flow Engineering for Robotic Instruction Following” explicitly states that it does not define a method called “FlowPlan-G2P” (Lin et al., 4 Mar 2025). In the present context, FlowPlan-G2P denotes the paper-to-patent framework introduced in (Pan et al., 5 Jan 2026), not the zero-shot robotic planning system.

The broader implication is that specialized-domain generation benefits from explicit intermediate structure and domain-specific evaluation. In FlowPlan-G2P, that principle takes the concrete form of a prompt-driven symbolic graph, a gated section planner, and a compliance-oriented realization stage, all designed to transform scientific discourse into patent-description discourse under enablement and section-schema constraints.