Pith. sign in

REVIEW 5 major objections 5 minor 41 references

Knowledge Graph Fusion with Large Language Models for Accurate, Explainable Manufacturing Process Planning

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Small 3B LLM plus knowledge graph matches GPT-4o on CNC planning

desk verdict A sensible RAG-over-KG engineering demo undercut by a self-referential benchmark; the headline numbers are uninterpretable until the evaluation is redone with independent ground truth. read the letter →

arxiv 2506.13026 v1 pith:JYWCKZRI submitted 2025-06-16 cs.AI cs.CL

classification cs.AIcs.CL
keywords knowledgegraphretrieval-augmentedgenerationlargelanguagemodelsCNCmachiningprocessplanningnumerichallucinationentityrelationextractionon-premisedeployment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ARKNESS claims that automatically turning machining documents into a knowledge graph of context-carrying triples, then feeding the retrieved subgraph into a small language model, fixes the two failures that keep LLMs off the shop floor: numeric hallucination and missing provenance. On its 155-question benchmark covering tool sizing, feed-speed choices, and tolerance diagnostics, the paper reports that a 3B-parameter Llama augmented by ARKNESS matches GPT-4o accuracy while showing gains of up to 25 percentage points in multiple-choice accuracy, 22.4 points in F1, and 8.1 times ROUGE-L on open-ended responses. The practical point is that if the claim holds, a factory can run a private, on-premise assistant that answers process-planning questions with exact values and citeable evidence instead of sending proprietary data to a cloud model or trusting a hallucinating generalist. The paper further claims that the supporting knowledge graph can be built without manual annotation, removing the curation bottleneck that has limited earlier manufacturing knowledge graphs.

What carries the argument

The load-bearing object is the augmented triple, written as $\langle s, r, (v, c)\rangle$, in which every relationship carries its original verbatim context string $c$; this preserves the provenance that plain subject-predicate-object triples lose. Retrieval is a two-stage mechanism: cosine-similarity ranking of embedded triples gives a top-$K$ candidate pool, and a beam search with depth $d_{\max}$ and beam width $b$ expands outward from those candidates to adjacent triples sharing a node, so quantitative questions can pull in supporting facts that a flat top-$K$ search would miss. The final prompt concatenates system instructions, the retrieved triple-and-context set, and the query, conditioning whatever LLM is plugged into the pipeline. The mechanism is model-agnostic: the paper's experiments vary the LLM from 3B to 8B open-weight models up to large cloud models and hold the graph and retriever fixed.

What would settle it

Give the ARKNESS pipeline a held-out set of CNC parameter questions whose answers are verified by independent machinists and whose source text is excluded from the knowledge graph corpus; if the augmented 3B model then fails to match GPT-4o accuracy or cites triples that do not support its numeric answer, the central claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a retrieval-augmented generation pipeline built on an automatically constructed, multi-relational machining knowledge graph can make a small open-weight LLM numerically reliable and explainable enough for CNC process planning. The knowledge graph is built zero-shot: a large model reads each paragraph of heterogeneous machining documents, G-code annotations, and vendor datasheets, and emits augmented triples of the form (entity, relation, object, context), where the context is the verbatim sentence the relation came from. At query time the question is embedded, the top-K triples are selected by cosine similarity, and a beam search expands to neighboring triples so the model receives the minimal evidence-linked subgraph rather than just scattered facts. The paper reports that this grounding lifts open-source models across the board, with the largest gains on quantitative machining questions, and that a lightweight 3B Llama answers multiple-choice and open-ended questions at GPT-4o-level accuracy while reducing numeric hallucination by 22 percentage points.

Load-bearing premise

The load-bearing premise is that the 155-question benchmark's ground-truth answers are correct and representative, because the same model family that built the knowledge graph also wrote the test questions and no independent expert validation or public release of the test set is provided.

Editorial extensions

If this is right

  • A private, on-premise 3B-parameter assistant could answer tool-sizing and feed-speed questions with numeric exactness and traceable evidence, removing the data-sovereignty barrier to shop-floor LLM use.
  • Numeric hallucination, not reasoning ability, appears to be the main bottleneck for technical LLM answers; supplying verbatim context from a knowledge graph can suppress it more than model scale alone does.
  • Deeper graph traversal helps quantitative questions, while shallow retrieval suffices for content questions, so retrieval depth can be tuned to question type.
  • Knowledge graph construction for a new manufacturing domain can start from raw documents without manual labeling or ontology engineering.
  • Smaller models benefit disproportionately from the graph, suggesting cost-efficient deployment where the graph, not the model, carries the domain expertise.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same document-to-augmented-triple pipeline should transfer to other regulated technical domains, such as maintenance manuals, medical device procedures, or chemical batch records, where numeric precision and provenance matter more than conversational fluency; the paper does not test this.
  • A cleaner test of the retrieval mechanism would drop the graph's source sentences from the LLM's parametric knowledge by using deliberately unfamiliar material; the paper's drill-size examples come from standard charts the models may have seen in training.
  • Applying ARKNESS to open-ended questions with exact-match numerical scoring instead of ROUGE-L would test whether the reported 8.1 times ROUGE-L gain reflects true numeric accuracy rather than lexical overlap.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces ARKNESS, a retrieval-augmented generation framework for CNC process planning. GPT-4o is prompted to extract entity-relation triples, each with a contextual description, from technical machining documents; the triples are stored in a PostgreSQL database. At query time, a semantic embedding model scores triples by cosine similarity, a top-K pool is selected, beam search expands the neighborhood up to depth d_max, and the retrieved context is concatenated into the prompt of a chosen LLM. The authors evaluate multiple-choice and open-ended questions across seven models, reporting that KG augmentation improves accuracy, F1, and ROUGE scores, and claiming that a 3B-parameter Llama model augmented by ARKNESS matches GPT-4o. The central quantitative claims rest on a benchmark that was generated by GPT-4o from the same documents used to build the KG, which raises circularity concerns, and the headline numbers are not fully consistent with the reported tables.

Significance. If the claimed effects are real, the framework would be a useful contribution: it is model-agnostic, runs on-premise, provides provenance through triple-level context, and the qualitative examples in Tables 3-6 show dramatic corrections of numeric hallucinations. The experiments cover multiple model families and parameter scales, and the retrieval procedure is clearly described. However, the evaluation is not currently convincing: the benchmark is self-generated, the abstract overstates results relative to Table 2, and there is no non-KG retrieval baseline. The contribution is therefore plausible but unproven as reported.

major comments (5)
  1. [Section 4.3 and Section 4.1] GPT-4o is used both to construct the KG triples from the source documents and to generate all benchmark questions from the same documents, with no independent human expert validation or release of the test set. The abstract's characterization of the benchmark as 'industry-curated' is unsupported by the text. This makes the evaluation self-referential: retrieval can succeed by returning the exact source passage from which a question was written, and the reported gains may reflect distribution alignment with GPT-4o's own semantic preferences rather than genuine process-planning competence. This issue is load-bearing for every headline number in the abstract.
  2. [Section 4.3 (question counts)] The stated totals are internally inconsistent. Section 4.3 lists 65 content-specific multiple-choice + 45 machining-specific multiple-choice + 104 content-specific open-ended + 45 machining-specific open-ended, which sums to 259 questions, while the abstract and introduction state that the benchmark contains 155 questions. The denominators for the accuracies in Table 2 and the averages in Figure 6 are therefore ambiguous, and the reader cannot verify the reported percentages.
  3. [Abstract and Section 5, Table 2] The headline claim that 'a lightweight 3B-parameter Llama-3 augmented by ARKNESS matches GPT-4o accuracy while achieving a +25 percentage point gain in multiple-choice accuracy' is not supported by Table 2. The +0.250 accuracy gain on machining-specific questions is achieved by Gemini 2.0 Flash (0.267 to 0.517), not by Llama 3.2 3B, whose gain is +0.216 (0.367 to 0.583). Moreover, Llama 3.2 3B with KG augmentation (0.583) remains below GPT-4o with KG augmentation (0.733) on machining-specific questions; the 'matches GPT-4o' statement holds only for content-specific questions. The +22.4 pp F1 gain does apply to Llama 3.2 3B on machining-specific questions, so the abstract sentence combines results from different models misleadingly.
  4. [Sections 3.3 and 5 (missing retrieval baseline)] The experiments compare a no-context baseline with KG-augmented prompting, but never compare against ordinary chunk-based passage retrieval over the same documents. Since ARKNESS retrieval effectively returns triples with their original verbatim context, the observed gains could come from passage-level context injection rather than from the graph structure, beam-search expansion, or relational organization. A top-K passage-retrieval baseline is needed to isolate the specific contribution of the knowledge graph.
  5. [Section 5, Tables 3-6 (ground truth provenance)] The 'Grounded Answer' values shown in Tables 3-6 appear to be verbatim strings drawn from the source documents, and the open-ended reference answers are said in Section 4.3 to have been generated by GPT-4o from those same documents. The paper does not state how the ground-truth numeric values were verified by domain experts, nor whether the reference answers were independently checked. Without such verification, the large ROUGE improvements may partly measure how faithfully the model copies the retrieved context rather than how well it solves realistic process-planning problems.
minor comments (5)
  1. [Section 5, Table 2] Gemini 2.0 Flash-Lite shows a 0.200 decrease in content-specific F1 after KG augmentation; the text acknowledges this exception but the conclusion that KG augmentation 'consistently enhances performance across most models' is worded too strongly.
  2. [Section 3.3, Eq. (6)] The formula uses 'C_d\V' but the visited set was introduced as 'E_v'; the notation should be aligned to avoid ambiguity about which set is subtracted.
  3. [Section 5, Figures 4 and 6] No error bars or statistical significance tests are reported despite the statement that results are averaged over 10 runs, so the reader cannot assess whether the observed differences are stable across runs.
  4. [Throughout] There are several typos and inconsistent acronym usages, including 'transversal' for 'traversal', 'ROGUE' for 'ROUGE', and 'enviornments' in the conclusion; these should be corrected.
  5. [Section 4.2 and Table 2] The hyperparameter choice for Table 2 (top-K = 10, depth = 0) is described as 'minimal viable integration,' but the selection procedure for these values is not described; because Figures 4 and 5 show sensitivity to these parameters, the reported results should be accompanied by a clear hyperparameter policy or an ablation.

Circularity Check

2 steps flagged · score 6.0 of 10

Central headline gains rest on a GPT-4o-generated benchmark built from the same documents as the KG; open-ended ROUGE-L is measured against text inserted into the prompt.

  1. self definitional [Section 4.1 (Knowledge Graph) and Section 4.3 (Question Category)]
    "Each document was processed using the automated graph-construction pipeline detailed in Section 3.2 using GPT-4o. ... GPT-4o was implemented to generate each question from the documents chosen with a total of 65 content specific multiple choice, 45 machining specific multiple choice, 104 content specific open ended, and 45 machining specific open ended."

    The same model, GPT-4o, and the same source documents define both the knowledge graph and the test questions. Every benchmark question is therefore generated from a passage that GPT-4o also converted into triples in the KG, so the correct answer is, by construction, present in the retrieval index. A retriever that returns the passage from which the question was written will trivially contain the grounded answer. The claimed gains (+25 pp, +22.4 pp, 8.1x) are measured against a benchmark that reflects GPT-4o's own document-derived distribution, not an independent industry-curated test. This makes the central evaluation self-referential.

  2. other [Section 3.2 (Automated Graph Construction), Section 3.4 (Large Language Model Generation), and Section 5 (Open-ended results, Figure 6)]
    "RELATIONSHIP_DESCRIPTION: A concise description of the relationship directly sourced from the input text. ... Then the final prompt P given to the LLM is constructed by concatenating the system instructions, query, and retrieved knowledge graph information: P(q,C) = I_sys ⊕ C ⊕ q"

    The open-ended ROUGE-L evaluation compares generated answers against grounded answers that come from the same document text stored as relationship descriptions in the KG. The prompt explicitly feeds that retrieved text into the LLM and instructs the model to use it ('focus on the most important information'). A model can therefore score high ROUGE-L by copying or lightly rephrasing the retrieved context, because the reference string is already part of the input. The reported 8.1x ROUGE-L improvement is thus partly a copying artifact rather than evidence of independent synthesis, and it is a second, separate self-referential loop beyond the GPT-4o-generated benchmark.

full rationale

The paper's core contribution is an evaluated system, not a formal derivation, so circularity must be assessed through the evaluation loop. The load-bearing problem is that the benchmark is not independent of the system's knowledge source: GPT-4o built the KG from the selected documents (Section 4.1), and GPT-4o also wrote every test question from those same documents (Section 4.3). Consequently, the correct answers are guaranteed to be present in the KG's triples and relationship descriptions, and the reported accuracy, F1, and ROUGE-L gains reflect retrieval of the exact passages used to write the questions. The open-ended metric is further inflated because the reference text is inserted into the prompt as retrieved context, making high ROUGE-L achievable by extraction. The abstract's 'industry-curated' characterization is contradicted by the paper's own description of GPT-4o-generated questions, and the count inconsistency (65+45+104+45 = 259 questions versus the claimed 155) is unreconciled. No chunk-based RAG baseline is reported, so the specific benefit of graph structure over plain passage retrieval is untested. There is no load-bearing self-citation chain or imported uniqueness theorem; the circularity is in the benchmark design, not in the retrieval equations themselves. Score 6 reflects that one or more headline 'predictions' reduce by construction to the GPT-4o-generated benchmark and to reference text placed in the prompt, while the embedding and beam-search machinery are not themselves circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

ARKNESS introduces no new physical entities, particles, forces, or conserved quantities. It is a software pipeline built from known components. The main ledger items are the manually chosen retrieval hyperparameters and the unverified assumptions that GPT-4o extraction and GPT-4o-generated questions faithfully represent the machining documents.

free parameters (4)
  • top-K retrieval count = K = 10 for the main results in Table 2
    Retrieval depth and K are swept in Figure 4 and a specific value is chosen for the reported benchmark. Accuracy varies with K, so the headline gains depend on this hyperparameter.
  • graph traversal depth d_max = 0 for Table 2; deeper values analyzed in Figure 4
    The beam-search depth controls how much graph neighborhood is injected into the prompt. The paper reports performance stabilization at different depths, indicating this is a tuned design choice.
  • beam width b = not specified in the text
    Equation 6 defines beam width b as a chosen parameter, but no value is reported, which affects both retrieval quality and reproducibility.
  • semantic embedding model f = not named
    Equation 3 depends on the chosen embedding model for both query and triple embeddings. Without naming the model, the retrieval results cannot be replicated.
assumptions (4)
  • domain assumption The selected source documents contain the authoritative machining knowledge needed to answer the benchmark questions.
    Table 1 defines the KG corpus. If key tables or standards are missing, retrieval gaps would be misattributed to LLM or framework failure.
  • domain assumption GPT-4o's zero-shot triple extraction and GPT-4o-generated benchmark questions are faithful to the source documents.
    Sections 3.2 and 4.3 rely on GPT-4o for both KG construction and question generation, but no manual validation rate or error analysis is reported.
  • domain assumption Cosine similarity over embedded triple text is a sufficient relevance signal for quantitative machining facts.
    Equation 3 defines retrieval solely by embedding similarity. The paper does not test whether the most similar triples are the factually correct triples.
  • domain assumption The prompt-plus-context construction in Equations 9 and 10 reliably transfers retrieved facts into the final answer.
    The paper does not measure cases where the LLM ignores or misuses retrieved context, though it notes context overload effects for one model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge Graph Fusion with Large Language Models for Accurate, Explainable Manufacturing Process Planning." pith.science (2026). https://pith.science/paper/JYWCKZRI

@misc{pith2026250613026,
  author       = {Pith},
  title        = {Pith review of: Knowledge Graph Fusion with Large Language Models for Accurate, Explainable Manufacturing Process Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JYWCKZRI}},
  note         = {Machine review of arXiv:2506.13026}
}
read the original abstract

Precision process planning in Computer Numerical Control (CNC) machining demands rapid, context-aware decisions on tool selection, feed-speed pairs, and multi-axis routing, placing immense cognitive and procedural burdens on engineers from design specification through final part inspection. Conventional rule-based computer-aided process planning and knowledge-engineering shells freeze domain know-how into static tables, which become limited when dealing with unseen topologies, novel material states, shifting cost-quality-sustainability weightings, or shop-floor constraints such as tool unavailability and energy caps. Large language models (LLMs) promise flexible, instruction-driven reasoning for tasks but they routinely hallucinate numeric values and provide no provenance. We present Augmented Retrieval Knowledge Network Enhanced Search & Synthesis (ARKNESS), the end-to-end framework that fuses zero-shot Knowledge Graph (KG) construction with retrieval-augmented generation to deliver verifiable, numerically exact answers for CNC process planning. ARKNESS (1) automatically distills heterogeneous machining documents, G-code annotations, and vendor datasheets into augmented triple, multi-relational graphs without manual labeling, and (2) couples any on-prem LLM with a retriever that injects the minimal, evidence-linked subgraph needed to answer a query. Benchmarked on 155 industry-curated questions spanning tool sizing and feed-speed optimization, a lightweight 3B-parameter Llama-3 augmented by ARKNESS matches GPT-4o accuracy while achieving a +25 percentage point gain in multiple-choice accuracy, +22.4 pp in F1, and 8.1x ROUGE-L on open-ended responses.

Figures

Figures reproduced from arXiv: 2506.13026 by the authors.

Figure 1
Figure 1. Illustration contrasting a knowledge graph (top) and a large language model (bottom) handling the same [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the framework with incorporating knowledge graph with large language model, illustrating [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Overview of the knowledge graph construction and graph transversal for user queries. The left side depicts [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Graph of large language model model accuracy as a function of graph traversal depth and Top K triple retrieval [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Radar plots of average large language model accuracy performance across knowledge graph completeness [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Average performance over 10 runs for evaluated large language models, comparing baseline outputs with [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 34 canonical work pages

  1. [1]

    Wiessner, P

    M. Wiessner, P. Blaser, S. Böhl, J. Mayr, W. Knapp, K. Wegener, Thermal test piece for 5-axis machine tools, Precision Engineering 52 (2018) 407–417

  2. [2]

    Hoang, H

    D. Hoang, H. Errahmouni, H. Chen, S. Rachuri, N. Mannan, R. ElKharboutly, M. Imani, R. Chen, F. Imani, Hierarchical representation and interpretable learning for accelerated quality monitoring in machining process, CIRP Journal of Manufacturing Science and Technology 50 (2024) 198–212

  3. [3]

    Z. Chen, D. Hoang, R. Chen, F. Imani, Distributed hyperdimensional computing for real-time data aggregation and interpretable quality monitoring in manufacturing, in: ASME International Mechanical Engineering Congress and Exposition, volume 88605, American Society of Mechanical Engineers, 2024, p. V002T03A092

  4. [4]

    Z. Li, Y . Dai, Z. Sun, C. L. Guan, T. Lai, H. Xu, X. Zhou, A sub-micron precision machining and measurement method of long travel metal guideways, Journal of Manufacturing Processes 133 (2025) 947–956

  5. [5]

    forinsightsconsultancy.com/reports/computer-numerical-control-cnc-machine-market/ ,

    ForInsights Consultancy, Computer numerical control (cnc) machine market forecast 2030, https://www. forinsightsconsultancy.com/reports/computer-numerical-control-cnc-machine-market/ ,

  6. [6]

    Hoang, N

    D. Hoang, N. Mannan, R. ElKharboutly, R. Chen, F. Imani, Edge cognitive data fusion: From in-situ sensing to quality characterization in hybrid manufacturing process, in: International Manufacturing Science and Engineering Conference, volume 87240, American Society of Mechanical Engineers, 2023, p. V002T06A029

  7. [7]

    Hoang, H

    D. Hoang, H. Chen, M. Imani, R. Chen, F. Imani, Brief paper: Multi-task brain-inspired learning for interlinking machining dynamics with parts geometrical deviations, in: International Manufacturing Science and Engineering Conference, volume 88117, American Society of Mechanical Engineers, 2024, p. V002T05A012

  8. [8]

    Spanaki, D

    K. Spanaki, D. Dennehy, T. Papadopoulos, R. Dubey, Data-driven digital transformation in operations and supply chain management, 2025

Show all 41 references
  1. [9]

    URL: https:// assets.new.siemens.com/siemens/assets/api/uuid:3d606495-dbe0-43e4-80b1-d04e27ada920/ dics-b10153-00-7600truecostofdowntime2022-144.pdf

    Siemens AG, The True Cost of Downtime 2022, Technical Report, Siemens, 2022. URL: https:// assets.new.siemens.com/siemens/assets/api/uuid:3d606495-dbe0-43e4-80b1-d04e27ada920/ dics-b10153-00-7600truecostofdowntime2022-144.pdf

  2. [10]

    N. C. Nwasuka, U. Nwaiwu, Computer-based production planning, scheduling and control: a review, Journal of Engineering Research 12 (2024) 275–280

  3. [11]

    M. Raza, Z. Jahangir, M. B. Riaz, M. J. Saeed, M. A. Sattar, Industrial applications of large language models, Scientific Reports 15 (2025) 13755

  4. [12]

    K. Šket, D. Potoˇcnik, M. Ficko, S. Klanˇcnik, Enhancing g-code programming in cnc machining using chatgpt: A comparative study of gpt-3.5 and gpt-4.0, Available at SSRN 4940034 (2024)

  5. [13]

    J. Jeon, Y . Sim, H. Lee, C. Han, D. Yun, E. Kim, S. L. Nagendra, M. B. Jun, Y . Kim, S. W. Lee, et al., Chatcnc: Conversational machine monitoring via large language model and real-time data retrieval augmented generation, Journal of Manufacturing Systems 79 (2025) 504–514

  6. [14]

    Rosati, F

    R. Rosati, F. Antonini, N. Muralikrishna, F. Tonetto, A. Mancini, Improving industrial question answering chatbots with domain-specific llms fine-tuning, in: 2024 20th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA), IEEE, 2024, pp. 1–7

  7. [15]

    Kanimozhi, Y

    S. Kanimozhi, Y . Sriker, et al., Explorative deployment of fine-tuned large language model for on-site computerized numeric control machine operator assistance, in: 2024 IEEE Silchar Subsection Conference (SILCON 2024), IEEE, 2024, pp. 1–6

  8. [16]

    Y . Xiao, S. Zheng, J. Shi, X. Du, J. Hong, Knowledge graph-based manufacturing process planning: A state-of- the-art review, Journal of Manufacturing Systems 70 (2023) 417–435

  9. [17]

    P. Wen, Y . Ma, R. Wang, Systematic knowledge modeling and extraction methods for manufacturing process planning based on knowledge graph, Advanced Engineering Informatics 58 (2023) 102172

  10. [18]

    Hoang, D

    D. Hoang, D. Gorsich, M. Castanier, F. Imani, Vector-symbolic knowledge graphs for enhanced memorization and reasoning in digital manufacturing, Available at SSRN 5097516 (2025)

  11. [19]

    L. Wang, H. Cheng, R. Wang, X. Huang, Machining scheme selection of features based on process knowledge graph and improved cosine similarity matching, Machines 13 (2025) 188

  12. [20]

    J. Guo, J. Wu, J. Bian, Q. He, Knowledge graph-based machining process route generation method, in: International Conference on Human-Computer Interaction, Springer, 2023, pp. 35–48

  13. [21]

    C. Cai, Z. Jiang, H. Wu, J. Wang, J. Liu, L. Song, Research on knowledge graph-driven equipment fault diagnosis method for intelligent manufacturing, The International Journal of Advanced Manufacturing Technology 130 (2024) 4649–4662

  14. [22]

    Y . Li, H. Zhao, H. Jiang, Y . Pan, Z. Liu, Z. Wu, P. Shu, J. Tian, T. Yang, S. Xu, et al., Large language models for manufacturing, arXiv preprint arXiv:2410.21418 (2024)

  15. [23]

    P. Wang, J. Karigiannis, R. X. Gao, Ontology-integrated tuning of large language model for intelligent maintenance, CIRP annals 73 (2024) 361–364

  16. [24]

    C. Qiu, B. Li, H. Liu, S. He, C. Hao, A novel method for machine tool structure condition monitoring based on knowledge graph, The International Journal of Advanced Manufacturing Technology 120 (2022) 563–582

  17. [25]

    C. Auer, M. Lysak, A. Nassar, M. Dolfi, N. Livathinos, P. Vagenas, C. B. Ramis, M. Omenetti, F. Lindlbauer, K. Dinkla, et al., Docling technical report, arXiv preprint arXiv:2408.09869 (2024)

  18. [26]

    Hurst, A

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford, et al., Gpt-4o system card, arXiv preprint arXiv:2410.21276 (2024)

  19. [27]

    Soori, F

    M. Soori, F. K. G. Jough, R. Dastres, B. Arezoo, A review in capabilities and challenges of 5-axis cnc milling machine tool operations, Preprint (2024). 19 KG-Fused LLMs for Explainable Process Planning

  20. [28]

    Y . Ye, T. Hu, C. Zhang, W. Luo, Design and development of a cnc machining process knowledge base using cloud technology, The International Journal of Advanced Manufacturing Technology 94 (2018) 3413–3425

  21. [29]

    F. N. Guo, Exploring the application of industrial robots in cnc machining, Journal of Global Humanities and Social Sciences 4 (2023) 231–235. URL: http://ojs.bonfuturepress.com/index.php/GHSS/article/ view/1499. doi:doi:10.61360/BoniGHSS232014990503

  22. [30]

    URL: https://academy.titansofcnc.com/files/Fundamentals_of_CNC_Machining.pdf, desk Copy

    Autodesk, Inc., Fundamentals of CNC Machining: A Practical Guide for Beginners, United States, 2014. URL: https://academy.titansofcnc.com/files/Fundamentals_of_CNC_Machining.pdf, desk Copy. Document Number: 060711

  23. [31]

    Nugrahanto, H

    I. Nugrahanto, H. Gunawan, H.-y. Chen, Innovative approaches to sustainable cnc machining: A machine learning perspective on energy optimization (2023)

  24. [32]

    Ihsan, Y

    M. Ihsan, Y . Sumantri, Y . Irawan, Integration of taguchi and promethee for cnc milling machining parameter optimization on aa6061, International Journal of Mechanical Engineering Technologies and Applications 5 (2024) 96–107

  25. [33]

    Zhang, J

    S. Zhang, J. Bai, Research on cnc programming and machining process based on cad/cam technology, Applied Mathematics and Nonlinear Sciences 9 (2024) 1–18. URL: https://doi.org/10.2478/amns-2024-0516 . doi:doi:10.2478/amns-2024-0516

  26. [34]

    H. Wu, X. Wang, X. Deng, H. Shen, X. Yao, Review on design research in cnc machine tools based on energy consumption, Sustainability 16 (2024) 847

  27. [35]

    Soori, F

    M. Soori, F. K. G. Jough, R. Dastres, B. Arezoo, Robotical automation in cnc machine tools: a review, acta mechanica et automatica 18 (2024)

  28. [36]

    Version 3.2, released 25 Sept 2024

    Meta AI, Llama 3.2 model card, https://github.com/meta-llama/llama-models/tree/main/models/ llama3_2, 2024. Version 3.2, released 25 Sept 2024

  29. [37]

    Grattafiori, A

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al., The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024)

  30. [38]

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, et al., Qwen2. 5 technical report, arXiv preprint arXiv:2412.15115 (2024)

  31. [39]

    URL: https://deepmind.google/technologies/gemini/ flash/

    Google DeepMind, Gemini 2.0 flash, 2025. URL: https://deepmind.google/technologies/gemini/ flash/

  32. [40]

    URL: https://deepmind.google/technologies/gemini/ flash-lite/

    Google DeepMind, Gemini flash lite, 2025. URL: https://deepmind.google/technologies/gemini/ flash-lite/. 20

  33. [2023]

    18 KG-Fused LLMs for Explainable Process Planning

    Accessed: 2025-05-12. 18 KG-Fused LLMs for Explainable Process Planning

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.