LogicTree-RAG: Logic Tree-guided Retrieval-Augmented Generation for Long-form Patent Drafting

Conference: NeurIPS2026 (task archive; this note is based on arXiv v1)
arXiv: 2609.30943v1
Area: Information Retrieval & RAG
Keywords: logic tree, retrieval-augmented generation, long-form generation, patent drafting, evidence attribution

TL;DR

LogicTree-RAG organizes a research paper into a retrievable, recursively expandable logic tree with evidence associations, then uses section-specific hybrid traversal to generate a long patent-description draft, achieving 20.22k output tokens, 43.79 coverage, and 66.34 source-combined factuality on the Pap2Pat test set; these results do not certify patentability or legal validity.

Background & Motivation

Turning a paper into a patent is not simply replacing academic wording with legal wording. A paper typically presents its contribution through problems, methods, and experiments, whereas a patent description must organize the technical problem, system components, operating procedures, and implementation details into a standalone disclosure. Even a model that accepts the entire paper may produce only a few thousand tokens of summary. Expanding separate chunks can instead repeat the same module, overdevelop one branch, and omit another essential branch. Sufficient input context and complete technical output are different capabilities.

Prior patent-generation methods often address localized tasks such as titles, abstracts, or claims. For document-scale drafting, COPGEN/Pap2Pat uses chunked outlines, while AutoPatent requires draft information. The authors argue that these inputs are often extracted from reference patents in the corresponding task setups, which may not reflect a setting with only an invention disclosure. This work aims to construct its own organization directly from a research paper instead of relying on a detailed reference-patent outline at inference time. The paper is only a proxy for an invention report; this does not establish that a published paper still meets filing-timing or novelty requirements.

The actual control problem is therefore determining which technical content has already been covered, which details belong to each branch, and which evidence should guide the next expansion. Core idea: use a logic tree with source-evidence associations as generation state, perform branch-specific retrieval, expansion, and refinement on that tree, then transform the same structure into a patent description with section-appropriate breadth and depth.

Method

Overall Architecture

The input is a research paper used as a proxy for an invention disclosure, and the output is description content within a long patent draft. The stages are “Semantic Chunking & Tree Initialization,” “Node-Aware Evidence Search,” “Evidence-Guided Expansion & Refinement,” and “Section-Specific Hybrid Traversal.” The system first constructs a source-document knowledge base and a shallow technical structure, processes nodes through a FIFO queue, and finally uses different traversal orders for Background, Summary, and Detailed Description.

Each node represents a technical element, such as a principle, step, or subsystem, and stores generated content together with corresponding source spans. This intermediate representation is not a formal proof tree: parent–child edges express technical organization rather than verified entailment. Associating evidence spans or their IDs with a node also does not mean that every sentence in the node is supported by those spans.

The diagram represents inference-time data flow and state updates only, not training supervision. Search results determine whether expansion proceeds; new children return to the queue, refinement changes the current node, and section generation begins only after a stopping condition is reached.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Research paper"] --> B["Semantic Chunking &<br/>Tree Initialization"]
    B --> C["Node-Aware<br/>Evidence Search"]
    C -->|Novel evidence exists| D["Evidence-Guided<br/>Expansion & Refinement"]
    D -->|New children enter FIFO| C
    C -->|Already covered: skip and process next node| C
    D -->|Queue empty or budget reached| E["Section-Specific<br/>Hybrid Traversal"]
    C -->|Queue empty or budget reached| E
    E --> F["Background, Summary,<br/>Detailed Description"]

Algorithm 1 explicitly lists Background, Summary, and Detailed Description as output sections. Although the abstract and human evaluation discuss a “complete patent” and Claim Structure, the supplied algorithm contains no independent claims-generation step. A Claims module cannot be inferred or invented.

Key Designs

1. Semantic Chunking & Tree Initialization: preserve technical context before building an expandable scaffold

Fixed-length chunks can separate a term definition from subsequent steps or combine adjacent paragraphs on different topics. The system first uses PyMuPDF to extract text blocks, fonts, coordinates, and paragraph layouts, recovering section and paragraph boundaries. It then computes cosine similarity between adjacent paragraph embeddings and merges paragraphs when similarity meets the threshold and accumulated length stays below the limit. Retrieval units thus aim to contain coherent technical meaning rather than arbitrary strings of a fixed length.

The mechanism is carried by the paper's Eq. (2). Similarity determines whether adjacent paragraphs are merged, not whether generated content is correct:

\[ s_{i,i+1}=\operatorname{sim}(\mathcal{E}_{p}(x_i),\mathcal{E}_{p}(x_{i+1})),\qquad s_{i,i+1}\geq\tau_s. \]

The default semantic threshold is 0.8, and the chunk limit is 512 tokens. Chunk vectors are stored in a Milvus document knowledge base; subsequent queries must also retrieve the corresponding source text. Layout reconstruction locates boundaries, while semantic merging determines which adjacent paragraphs should remain together. Layout information itself is not semantic evidence.

The LLM next generates an initial technical description as the root and induces the first layer of core technical elements from it. The root summarizes the invention, internal nodes organize higher-level concepts, and leaves contain details awaiting expansion. First-layer nodes enter the subsequent queue rather than becoming a fixed patent-section outline. The main text and Algorithm 2 describe initialization using the chunk set, but the actual initialization prompt in Appendix E explicitly uses only introduction chunks. This scope difference matters: the knowledge base covers the paper, while the root-description prompt is narrower.

Nodes have generated text and evidence sets. In interpreting that association, node text should be distinguished from the spans/evidence IDs used to trace its sources. The paper imposes a nonempty evidence-set constraint:

\[ \forall v\in V,\quad \mathcal{E}(v)\neq\emptyset. \]

This constraint requires an attributable source, not verified support for every detail. The root is associated with the chunk set, and new children receive the current retrieval context. No automatic proof establishes that every technical statement follows from those sources, particularly when later internal-knowledge completion adds details not confirmed by the paper.

2. Node-Aware Evidence Search: select evidence that belongs more specifically to the current branch

Consider sibling nodes for “camera tracking” and “coordinate calibration.” Retrieval based only on similarity to the current node can repeatedly return a general system overview for both. The system first queries the knowledge base with the current node text, retrieving up to 100 candidates. It then reranks them by subtracting their maximum similarity to another sibling from their similarity to the current node. Evidence must be relevant and relatively better suited to the current branch, reducing competition for the same generic explanation.

The crucial operation subtracts the most similar sibling, not average similarity to all nodes. The paper's Eq. (6) is:

\[ S_{dr}(d)=\operatorname{sim}(d,c(v_t))-\max_{v_e\in V_f^{child}\setminus\{v_t\}}\operatorname{sim}(d,c(v_e)). \]

The current node's parent is \(v_f\), and \(V_f^{child}\) is that parent's child set. The comparison therefore concerns siblings of the current node, not its own children. However, the main-text Remark describes local filtering as overlap with the current node's children, which conflicts with the equation. This note follows the equation and its variable definitions while retaining the wording discrepancy.

The main text describes score-threshold filtering. The actual configuration in Appendix F.3 instead keeps the reranked top-5 and adaptively sets the threshold to the score of the 5th candidate. This controls evidence quantity by rank and should not be described as a fixed-threshold experiment. Retrieval and reranking are followed by a global novelty check: if other branches already cover the evidence, the current node is skipped. Local sibling discrimination assigns responsibilities under one parent; the global check limits cross-branch duplication. They are distinct mechanisms.

The main text summarizes the global scope as outside the parent branch. Algorithm 2 instead excludes the parent's direct child set in its set expression, without fully specifying exclusion of every descendant in that branch. Both descriptions convey cross-branch coverage control, but the coverage decision and exact exclusion boundary are insufficiently detailed. They do not establish a strict semantic deduplication proof.

3. Evidence-Guided Expansion & Refinement: add technical hierarchy and resolve within-node gaps separately

Tree construction uses breadth-first FIFO scheduling, initialized with existing leaves. Each iteration removes the front node, performs node-aware search and the global coverage check, and skips the node if no new evidence remains. Otherwise, expansion precedes a check for refinement. This scheduling avoids exhausting the budget on one deep branch before other central technical components have been developed.

The expansion prompt uses the retrieved evidence to elaborate the current contribution into more detailed technical paragraphs, which are segmented into semantic units and instantiated as new children. These children retain the current evidence association and are appended to the queue. Their own content becomes the query in a later iteration. Expansion changes tree size and hierarchy rather than merely lengthening the original node; it progressively guides retrieval into finer technical regions.

Refinement addresses a different gap: a node can belong to the correct branch while saying only “perform calibration,” without definitions, inputs, outputs, parameters, or computation steps. A detection prompt first identifies under-specified exact phrases from the node. Source availability then determines the completion route. For domain terms appearing in the source with available supporting evidence, evidence-grounded retrieval-augmented generation updates the node. When relevant source evidence is unavailable, the model uses internal knowledge to provide generic, implementation-oriented descriptions. This is completion, not a new source-backed proof.

The paper reports that the internal-knowledge route is triggered in only 8.6% of steps, and removing it has relatively small effects on coverage and factuality. The main gains therefore align more closely with improved organization of retrieved evidence than extensive expansion from model knowledge. Infrequent use does not eliminate risk: when the source does not specify a parameter or implementation, generic completion can turn plausible background knowledge into a definitive disclosure that does not belong to the invention.

The main text stops when the queue is empty or accumulated content reaches a budget. Implementation settings additionally cap the tree at 64 nodes and depth 3, with the root at depth 0. A budget controls growth but does not guarantee coverage of every necessary technical element. Equal node counts also do not mean equally useful evidence across branches. Missing source information should remain visible rather than equating budget-limited completion with technical completeness.

4. Section-Specific Hybrid Traversal: establish breadth before elaborating implementations in depth

The completed tree is not serialized in one universal order. Background and Summary use BFS, prioritizing higher-level concepts and multiple technical branches so that a single implementation detail does not dominate the invention's overall picture. Detailed Description first uses BFS and LLM-based semantic clustering to identify coherent logical subtrees, then applies DFS within each subtree to elaborate the technical chain.

This distinction follows section function. A summary explains the central components and their cooperation; a detailed description explains how an individual component operates. BFS breadth is not mechanically equal word allocation, and DFS depth does not license invented implementation details. Section text should still be constrained by the constructed nodes and their source evidence. The pure-DFS ablation produces longer output with lower coverage, showing that length is not a substitute for section balance.

The controllable variables are tree budget, node granularity, and traversal strategy, not formally verified legal correctness. Parent–child structure can expose conceptual dependencies for professional inspection and revision, but no algorithmic step verifies claim scope, novelty, inventive step, or disclosure sufficiency under a particular jurisdiction.

A Worked Example

The following explanation uses the optical patient-localization material in Appendix F.5. It is not a new experiment reported by the paper and does not provide clinical operating advice.

The introduction describes tracking a single infrared reflective marker with CCD cameras and supplementing existing image-guidance systems with independent position verification. Initialization could form a system overview and branches for “optical tracking,” “coordinate calibration,” and “workflow.” These first-layer names are illustrative, not a fixed tree published by the authors.

When processing “coordinate calibration,” retrieval should focus on marker coordinates and correspondences between camera and room coordinate systems rather than repeat the tracking overview. The sibling-similarity penalty helps distinguish this branch from “optical tracking.” If other branches have not covered the evidence, source-described calibration procedures are expanded into children and added to FIFO.

If a node says only “solve relative pose using SVD,” refinement should retrieve coordinate inputs and transformation steps from the source instead of inventing unreported accuracy or parameters from general knowledge. After tree construction, BFS summarizes component cooperation, while DFS within the calibration subtree explains the steps in Detailed Description. Evidence associations identify sources for inspection; they do not establish clinical safety or patent legal validity.

Loss & Training

LogicTree-RAG is an inference-time orchestration framework with no new training loss or task-specific fine-tuning. The default backbone is Qwen3-80B, with temperature 0.3, top-p 1.0, and a maximum of 8,192 generated tokens per call. Multiple calls accumulate into the long output. The COPGEN SFT variant uses separate training settings; its training gains cannot be attributed to LogicTree-RAG.

Key Experimental Results

Main Results

Pap2Pat contains 1813 paper–granted-patent pairs. The main results use 500 test pairs, and COPGEN SFT uses 1,000 training pairs. LCFO contains 252 long documents and evaluates non-patent document expansion. The cache provides figure descriptions but not a complete set of reliably transcribable numerical results, so LCFO scores are not reconstructed.

Coverage uses the generated document as premise and reference-patent sentences as hypotheses, averaging maximum NLI entailment probabilities after BM25 retrieval to assess reference-content coverage. \(\mathcal{F}_{Pat}\) reverses the direction to assess support for generated sentences from the reference patent. The premise for \(\mathcal{F}_{Src}\) is actually the reference patent plus the source paper, not the source paper alone. ROUGE-L is a textual-overlap proxy; Style combines n-gram profiles and StyloMetrix; Repetition measures repeated n-grams in sliding windows; Coherence uses DiscoScore. None certifies legal or clinical truth.

The table selects rows from main-text Table 1. COPGEN† retains the original symbol. Appendix F.2 describes empty-outline and long-outline settings, but Table 1 does not clearly explain the symbol mapping, so † is not assigned an inferred input condition.

Method	Output tokens	Coverage ↑	FPat ↑	FSrc ↑	ROUGE-L ↑	Style ↑	Repetition ↓	Coherence ↑	Time s ↓
Qwen3-80B	8.91k	39.09	37.50	44.88	34.08	35.06	6.05	95.71	81
GPT-5 DT	2.95k	40.91	46.70	62.04	20.56	41.52	7.25	97.34	146
COPGEN	7.69k	39.21	57.63	62.39	32.56	59.54	15.47	97.11	265
COPGEN†	8.96k	32.85	52.44	58.16	28.96	43.38	18.94	97.21	216
COPGEN w/ SFT	25.85k	40.77	50.12	59.69	38.72	62.21	21.58	97.50	221
LongWriter	10.81k	37.21	36.65	45.21	21.12	34.26	6.67	95.62	197
LogicTree-RAG	20.22k	43.79	58.92	66.34	38.91	65.68	4.14	97.45	311
Contamination Control	18.96k	42.95	57.79	66.57	40.28	64.28	4.60	97.23	298

The main result improves coverage, factuality, style, and repetition control relative to these generative baselines. However, COPGEN w/ SFT has higher Coherence, 97.50 versus 97.45, and longer output. The reference-patent and source-paper heuristic rows are not generation baselines, so “best among generative baselines” cannot be extended to every row of the full table.

Contamination Control uses a temporally filtered post-2024 subset, not a difficulty-matched paired comparison with the full test set. Its FSrc of 66.57 and ROUGE-L of 40.28 exceed the main result's 66.34 and 38.91, respectively, so it is inaccurate to describe every metric as slightly lower. Temporal filtering reduces some contamination concerns but does not prove absence of pretraining overlap for every backbone.

Ablation Study

The following table uses the original values and metric directions from main-text Table 2.

Config	Output tokens	Coverage ↑	FPat ↑	Style ↑	Repetition ↓
w/o Logic Tree	9.10k	39.85	56.92	60.47	16.86
Logic Tree w/o Retrieval	8.21k	40.56	51.16	59.05	9.38
w/o Semantic Chunking	19.42k	42.09	53.74	62.27	10.93
w/o Refinement	18.58k	39.92	56.85	61.87	4.09
w/o Evidence-Grounded refinement	19.60k	40.16	56.98	62.64	4.67
w/o Internal Knowledge refinement	18.95k	43.58	58.85	65.62	5.21
w/o Node-aware Reranking	20.85k	42.60	57.13	64.79	8.32
w/ DFS Traversal	21.13k	41.15	58.68	65.12	5.34
LogicTree-RAG	20.22k	43.79	58.92	65.68	4.14

Removing the tree changes coverage from 43.79 to 39.85, while removing retrieval changes FPat from 58.92 to 51.16, showing different roles for organization and source support. Without node-aware reranking, output becomes longer but repetition rises from 4.14 to 8.32. Pure DFS also produces longer output with lower coverage. Output length alone therefore does not explain the gains.

The paper's claim that the full method is best on every metric is too strong: w/o Refinement has repetition of 4.09, lower than the full method's 4.14. Refinement improves coverage, factuality, and style, but the table does not show dominance on every metric.

Key Findings

Appendix G.3 checks whether generated content at retrieval-triggered nodes is supported by retrieved evidence. Human auditing covers only 20 cases. “Fully supported” describes support relative to the supplied sources, not world truth, clinical effectiveness, or legal correctness.

Evaluation	Method	Fully % ↑	Partial %	Unsupported % ↓
LLM-as-judge	COPGEN	70.4	20.2	9.4
LLM-as-judge	LogicTree-RAG	85.6	12.0	2.4
Human Audit	COPGEN	60	30	10
Human Audit	LogicTree-RAG	75	25	0

• Compute has explicit bounds: Appendix G.4 reports an average of 20 LLM calls, 311s, and 7.60k total input tokens per patent. Call proportions are 5% initialization, 45% tree construction, and 50% traversal. The 64-node and depth-3 limits are configuration caps, not realized counts for every document. The 20 calls do not establish a separate call for each node.
• Efficiency is not fully matched cost: Token efficiency is output tokens divided by total input tokens. It does not substitute for API prices or quality at identical cost. Shared decoding settings do not imply identical backbones, training data, call counts, output lengths, or reference-outline information.
• Human preference evidence is limited: Appendix G.5 reports blinded evaluation by 3 patent practitioners on 10 samples. Win rates against COPGEN for Overall, Legality, Claim Structure, Antecedent Basis, and Fidelity are 70.5, 65.7, 85.2, 82.1, and 78.5. How decimal rates are aggregated from this sample size and win/loss/tie judgments is insufficiently explained; stable legal conclusions do not follow.
• Structural control can transfer, but numerical boundaries remain: LCFO changes only the output schema. Section analysis reports up to 17.4% improvement in Detailed Description coverage, but the cache lacks complete figure values and an explicit paired row. This percentage is not converted into absolute percentage points or used to reconstruct a chart.

Highlights & Insights

• Retrieval becomes evidence allocation with branch responsibilities: Conventional RAG asks whether evidence is relevant to the query; this method also asks whether it belongs more strongly to a neighboring branch. Relative relevance could reduce repeated use of the same overview in module-level technical reports.
• Expansion and refinement address different failures: Expansion adds technical hierarchy, while refinement resolves gaps within existing nodes. Separating an omitted subprocess from an inadequately explained subprocess helps localize incomplete generation.
• Traversal order becomes a writing-control variable: The same nodes can support a global summary or a detailed implementation narrative. Separating organization from final prose helps reviewers determine whether an omission arose during tree construction or textual realization.

Limitations & Future Work

• Not a formal verification system: Nonempty evidence sets and higher support rates do not guarantee sentence-level entailment; partially supported and unsupported content remains. Detail-level source verification and explicit internal-knowledge labels could prevent conjecture from being presented as source fact.
• A gap in complete-patent scope: Algorithm 1 shows only three description-section categories, while human evaluation includes claim structure. An unpublished Claims pipeline cannot be inferred from its evaluation scores; generation and verification protocols need clarification.
• Contradictory human-evaluation status: The main text and Appendix G.5 describe conducted expert evaluation, whereas Appendix H treats practitioner assessment as future validation. This note preserves the inconsistency rather than reconciling it into an unambiguously completed study.
• Reproducibility gaps in sources and definitions: All-chunk versus introduction-only initialization, sibling-versus-child wording, and the global coverage exclusion scope require clarification. Retrieval similarity is not strict novelty detection, and missing figure values should not be invented.
• Professional review remains necessary: Consistent technical disclosure does not establish novelty, inventive step, patentability, legal validity, or filing readiness. This is an early-stage drafting assistant, not a replacement for patent attorneys. The medical example also provides no evidence of clinical reliability.

Related Work & Insights

• vs COPGEN/Pap2Pat: COPGEN selects sources and generates text around outline chunks. This work induces a logic tree and controls its growth with sibling discrimination and cross-branch coverage checks. Reduced dependence on detailed reference outlines comes at the cost of uncertainty in tree construction and coverage decisions.
• vs LongWriter/LongWriter-Zero: These methods primarily enhance long-output capability, whereas this work emphasizes source evidence and technical hierarchy. The ablations suggest evaluating length and completeness separately rather than judging writing quality by tokens alone.
• vs ReAct/Plan-then-Execute: Appendix G.2 replaces orchestration with generic agents while retaining the same core tools, obtaining coverage of 40.28 and 42.02 versus 43.79 for tree orchestration. This supports explicit structural state in this task, not a universal claim that generic agents are worse at all long-form generation tasks.

Rating

• Novelty: 4/5. Combines branch-specific retrieval, evidence-associated trees, and section traversal into a clear inference-time framework.
• Experimental Thoroughness: 3/5. Main experiments, ablations, costs, and small human audits are informative, but evaluation scope and human aggregation need clarification.
• Writing Quality: 3/5. The central mechanism is understandable, with inconsistencies in initialization scope, node-relation wording, and human-evaluation status.
• Value: 4/5. Offers useful organization principles for source-constrained technical drafts without guaranteeing legal or clinical applicability.